An Introduction To Text-To-Speech Synthesis
1 like41 views
The document is a textbook on text-to-speech synthesis. It provides an overview of the key topics and challenges in developing text-to-speech systems. These include analyzing the differences between written and spoken language, understanding human communication processes, organizing the input text into linguistic units like words and sentences, decoding and interpreting the text, predicting prosody from the text, converting text to speech sounds, and techniques for synthesizing speech. The textbook contains 18 chapters that delve deeper into each of these areas from both theoretical and engineering perspectives.
![1 INTRODUCTION
This is a book about getting computers to read out loud. It is therefore about three things: the
process of reading, the process of speaking, and the issues involved in getting computers (as
opposed to humans) to do this. This field of study is known both as speech synthesis, that is the
“synthetic” (computer) generation of speech, and text-to-speech or TTS; the process of converting
written text into speech. As such it compliments other language technologies such as speech
recognition, which aims to convert speech into text, machine translation which converts writing
or speech in one language into writing or speech in another.
I am assuming that most readers have heard some synthetic speech in their life. We experi-
ence this in a number of situations; some telephone information systems have automated speech
response, speech synthesis is often used as an aid to the disabled, and Professor Stephen Hawking
has probably contributed more than anyone else to the direct exposure of (one particular type of)
synthetic speech. The idea of artificially generated speech has of course been around for a long
time - hardly any science fiction film is complete without a talking computer of some sort. In fact
science fiction has had an interesting effect on the field and our impressions of it. Sometimes (less
technically aware) people believe that perfect speech synthesis exists because they “heard it on
Star Trek”1. Often makers of science fiction films fake the synthesis by using an actor, although
usually some processing is added to the voice to make it sound “computerised”. Some actually
use real speech synthesis systems, but interestingly these are usually not state of the art systems,
as these sound too natural, and may mislead the viewer2 One of the films that actually makes a
genuine attempt to predict how synthetic voices will sound is the computer HAL in the film 2001:
A Space Odyssey [265]. The fact that this computer spoke with a calm and near humanlike voice
gave rise to the sense of genuine intelligence in the machine. While many parts of this film were
wide of the mark (especially, the ability of HAL to understand, rather than just recognise human
speech), the makers of the film just about got it right in predicting how good computer voices
would be in the year in question.
1 Younger readers please substitute the in-vogue science-fiction series of the day
2 In much the same way, when someone types the wrong password on a computer, the screen starts flashing and saying
“access denied”. Some even go so far as to have a siren sounding. Those of us who use computers know this never
happens, but in a sense we go along with the exaggeration as it adds to the drama.
1](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-25-320.jpg)
![2 Chapter 1. Introduction
Speech synthesis has progressed remarkably in recent years, and it is no longer the case
that state-of-the-art systems sound overtly mechanical and robotic. That said, it is normally fairly
easy to tell that it is a computer talking rather than a human, and so substantial progress is still to
be made. When assessing a computer’s ability to speak, one fluctuates between two judgments.
On the one hand, it is tempting to paraphrase Dr Johnson’s famous remark [61] “Sir, a talking
computer is like a dog’s walking on his hind legs. It is not done well; but you are surprised to
find it done at all.” And indeed, even as an experienced text-to-speech researcher who has listened
to more synthetic speech than could be healthy in one life, I find that sometimes I am genuinely
surprised and thrilled in a naive way that here we have a talking computer: “like wow! it talks!”.
On the other hand it is also possible to have the impression that computers are quite dreadful at
the job of speaking; they make frequent mistakes, drone on, and just sound plain wrong in many
cases. These impressions are all part of the mysteries and complexities of speech.
1.1 WHAT ARE TEXT-TO-SPEECH SYSTEMS FOR?
Text-to-speech systems have an enormous range of applications. Their first real use was in reading
systems for the blind, where a system would read some text from a book and convert it into speech.
These early systems of course sounded very mechanical, but their adoption by blind people was
hardly surprising as the other options of reading braille or having a real person do the reading
were often not possible. Today, quite sophisticated systems exist that facilitate human computer
interaction for the blind, in which the TTS can help the user navigate around a windows system.
The mainstream adoption of TTS has been severely limited by its quality. Apart from users
who have little choice (as in the case with blind people), people’s reaction to old style TTS is
not particularly positive. While people may be somewhat impressed and quite happy to listen to
a few sentences, in general the novelty of this soon wears off. In recent years, the considerable
advances in quality have changed the situation such that TTS systems are more common in a
number of applications. Probably the main use of TTS today is in call-centre automation, where a
user calls to pay an electricity bill or book some travel and conducts the entire transaction through
an automatic dialogue system Beyond this, TTS systems have been used for reading news stories,
weather reports, travel directions and a wide variety of other applications.
While this book concentrates on the practical, engineering aspects of text-to-speech, it is
worth commenting that research in this field has contributed an enormous amount to our general
understanding of language. Often this has been in the form of “negative” evidence, meaning that
when a theory thought to be true was implemented in a TTS system it was shown to be false; in
fact as we shall see, many linguistic theories have fallen when rigorously tested in speech systems.
More positively, TTS systems have made a good testing grounds for many models and theories,
and TTS systems are certainly interesting in their own terms, without reference to application or
use.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-26-320.jpg)
![4 Chapter 1. Introduction
1.3 THE ENGINEERING APPROACH
In this book, we take what is known as an engineering approach to the text-to-speech problem.
The term “engineering” is often used to mean that systems are simply bolted together, with no
underlying theory or methodology. Engineering is of course much more than this, and it should be
clear that great feats of engineering, such as the Brooklyn bridge were not simply the result of some
engineers waking up one morning and banging some bolts together. So by “engineering”, we mean
that we are tackling this problem in the best traditions of other engineering; these include, working
with the materials available and building a practical system that doesn’t for instance take days to
produce a single sentence. Furthermore, we don’t use the term engineering to mean that this field
is only relevant or accessible to those with (traditional) engineering backgrounds or education. As
we explain below, TTS is a field relevant to people from many different backgrounds.
One point of note is that we can contrast the engineering approach with the scientific ap-
proach. Our task is to build the best possible text-to-speech system, and in doing so, we will use
any model, mathematics, data, theory or tool that serves our purpose. Our main job is to build an
artefact and we will use any means possible to do so. All artefact creation can be called engineer-
ing, but good engineering involves more: often we wish to make good use of our resources (we
don’t want to use a hammer to crack a nut); we also in general want to base our system on solid
principles. This is for several reasons. First, using solid (say mathematical) principles assures us
we are on well tested ground; we can trust these principles and don’t have to experimentally verify
every step we do. Second, we are of course not building the last ever text-to-speech system; our
system is one step in a continual development; by basing our system on solid principles we hope
to help others to improve and build on our work. Finally, using solid principles has the advantage
of helping us diagnose the system, for instance to help us find why some components do perhaps
better than expected, and allow the principles of which these components are based to be used for
other problems.
Speech synthesis has also been approached from a more scientific aspect. Researchers who
pursue this approach are not interested in building systems for their own sake, but rather as models
which will shine light on human speech and language abilities. As such, the goals are different,
and for example, it is important in this approach to use techniques which are at least plausible
possibilities for how humans would handle this task. A good example of the difference is in the
concatenative waveform techniques which we will use predominantly; recording large numbers of
audio waveforms, chopping them up and gluing them back together can produce very high quality
speech. It is of course absurd to think that this is how humans do it. We bring this point up
because speech synthesis is often used (or was certainly used in the past) as a testing ground for
many theories of speech and language. As a leading proponent of the scientific viewpoint states,
so long as the two approaches are not confused, no harm should arise (Huckvale [225]).](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-28-320.jpg)
![Section 1.4. Overview of the book 5
1.4 OVERVIEW OF THE BOOK
I must confess to generally hating sections entitled “how to read this book” and so on. I feel that
if I bought it, I should be able to read it any way I damn well please! Nevertheless, I feel some
guidelines may be useful.
This book is what one might call an extended text book. A normal text book has the job of
explaining a field or subject to outsiders and this book certainly has that goal. I qualify this by
using the term “extended” for two reasons. Firstly the book contains some original work, and is
not simply a summary, collection or retelling of existing ideas. Secondly, the book aims to take
the reader right up to the current state of the art. In reality this can never be fully achieved, but
the book is genuinely intended to be an “all you ever need to know”. More modestly, it can be
thought of as “all that I know and can explain”. In other words: this is it: I certainly couldn’t write
a second book which dealt with more advanced topics.
Despite these original sections, the book is certainly not a monograph. This point is worth
reinforcing: because of my personal involvement in many aspects of TTS research over the last 15
years, and specifically because of my involvement in the development of many well-known TTS
systems, including Chatr [53], Festival [55], and rVoice, many friends and colleagues have asked
me whether this is a book about those systems or the techniques behind those systems. Let me
clearly state that this is not the case; Text-to-Speech Synthesis is not a system book that describes
one particular system; rather I aim for a general account that describes current techniques without
reference to any particular single system or theory.
1.4.1 Viewpoints within the book
That said, this book aims to provide a single, coherent picture of text-to-speech, rather than simply
a list of available techniques. And while not being a book centred on any one system, it is certainly
heavily influenced by the general philosophy that I have been using (and evolving) over the past
years, and I think it is proper at this stage to say something about what this philosophy is and
how it may differ from other views. In the broadest sense, I adopt what is probably the current
mainstream view in TTS, that this is an engineering problem, that should be approached with the
aim of producing the best possible system, rather than with the aim of investigating any particular
linguistic or other theory.
Within the engineering view, I again have taken a more specialist view in posing the text-to-
speech problem as one where we have a single integrated text analysis component followed by a
single integrated speech synthesis component. I have called this the common form model (this
and other models are explained in Chapter 3.) While the common form model differs significantly
from the usual “pipelined” models, most work that has been carried out in one framework can be
used in the other without too much difficulty.
In addition to this, there are many parts which can be considered as original (at least to
my knowledge) and in this sense, the book diverges from being a pure text book at these points.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-29-320.jpg)
![Section 2.2. Human Communication 15
with kicking or buckets.
From this basic model, we can now explore some of the fundamental features of language. In
doing so, we follow Hockett [214], who stated the principle defining properties of verbal language.
Among these are:
Arbitrary
Phonemes do not have inherent meaning but words do; a word is in effect a form/meaning pair in
which the correspondence between the meaning and the form (ie. the sequence of phonemes) is
arbitrary. Because of this, there is no way to guess the meaning of a word if it is not learned, so as
a reader, you won’t know the meaning of the word PANDICULATION unless you are informed of its
meaning. From this, it follows that words of similar form can have complete different meanings,
so BIT and PIT mean quite different things despite only having a slight sound difference, whereas
DOG and WOLF are similar in meaning, but share no sounds.
Duality
The duality principle states that verbal language is not one system but two; the first system uses a
small inventory of forms, phonemes, which sequentially combine to form a much larger inventory
of words. The phonemes don’t carry meaning but the words do, and this association is arbitrary.
The second system combines the large number of words into an effectively limitless number of
sentences.
Productive
The productive property means that there is an effectively unlimited number of things that one
can say in a verbal language. It is not the case that we have a fixed number of messages and in
communication simply choose one from this selection; rather many of the sentences we produce
are unique, that is, the particular combination of words we use have never been used before. Of
course many of the sentences we say are conventionalised (“How do you do?” etc), but most are
unique. It might seem questionable if this is really the case as most people don’t feel they are being
particularly “original” when the speak. However, a simple examination of the combinatorics of
combining even a modest vocabulary of 10,000 words shows that a vast number of possibilities
exist (certainly larger that the number of atoms in the universe).
Discrete
Both the basic types of linguistic unit, the phoneme and the word, are discrete. Phonemes are dis-
tinct from one another and form a finite and fixed set. The set of words is less fixed (we can create
new words), but nevertheless all words are distinct from one another and are countable. What
is remarkable about the discreteness property of language is that the primary signal of language,
speech, is not discrete and hence there is a mismatch between form and signal. Speech signals
are continuous functions of pressure over time (see Chapter 10) and so speaking is a process of](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-39-320.jpg)
![Section 2.2. Human Communication 17
Semantics is sometimes given a broad definition as the study of “meaning” in the general
sense, but is sometimes given a much narrower definition, and is concerned only with certain as-
pects of meaning, such as analytic truth, reference, and logic. Finally pragmatics concerns a num-
ber of related areas including the aspects of meaning not covered by semantics, the interaction of
language, a speaker’s knowledge and the environment and the use of language in conversations and
discourses. Sometimes pragmatics is described as being “parallel” to semantics/syntax/phonology
etc as it is the study of how language is used.
It is important to realise that these terms are used both to refer to the linguistic level and the
area of study. In traditional linguistics these terms are often taken as absolute levels that exist in
language, such that one sometimes sees “turf wars” between phoneticians and phonologists about
who best should describe a particular phenomenon. From our engineering perspective, we will find
it more convenient to adopt these terms, but not worry too much about the division between each;
rather it is best to view these levels as salient points in a continuum stretching from pragmatics to
acoustics.
2.2.3 Affective Prosody
We now turn to a different component of human communication. Broadly speaking, we can state
that prosody has two main communicative purposes, termed affective and augmentative. The
affective use of prosody can be used to convey a variety of meanings: perhaps the most straight-
forward case is when is is used to indicate primary emotions such as anger or pain. Prosody is
particularly good at this; when we hurt ourselves and want to tell someone we usually let out a
yell rather than say “look, I’ve cut off my fingers” in a calm measured tone. The use of prosody
to convey primary emotion is also largely universal; when listening to someone who speaks a dif-
ferent language we find they use more or less the same form to indicate the same meanings. In
fact this extends to animals too; we are often in little doubt as to what an animal snarl or growl is
trying to convey. Beyond primary emotions we can consider a whole range of meanings that are
expressible via prosody. These include secondary emotions such as angst or distrust, speech acts
such as question, statement and imperative; modes of expression such as sarcasm or hesitancy and
so on.
Many of these phenomena are language specific (in contrast to primary emotion), and have a
conventional or arbitrary character such that a speaker of that language has to learn the appropriate
prosodic pattern in the same way that they learn the form/meaning correspondences for words in
verbal language. In fact we can see these effects as lying on a scale with universal systems lying
at one end and conventional arbitrary systems lying at the other. In many languages, a rising pitch
pattern at the end of a sentence indicates a question; but this is not universally so. Ladd [269]
shows that the pattern is reversed in Hungarian such that certain questions are indicate by a falling
pitch.
One of the central debates in the field of prosodic linguistics is just how far this scale extends.
Crudely speaking, the argument concerns whether prosody is mainly a universal, non-arbitrary](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-41-320.jpg)
![20 Chapter 2. Communication and Language
in how conversations happen. In fact there are a number of guiding principles involved, which we
shall now explore.
The first concerns the understood conventions of communication. In human conversation,
participants don’t just say everything on their mind; to do so would swamp the conversation.
Rather, they only say what they need to in order to get their point across. We can state this
as a case of effectiveness and efficiency; we wish to effectively communicate our message so
that it is understood, but also to do so efficiently so that we don’t waste energy. In fact it is
clear that conversation works by adopting a set of understood conventions that govern or guide
the conversation. The most famous exposition of these principles was by Grice, who defined
four main principles, of conversation which are now commonly known as Grice’s maxims [184].
These principles state for instance that it is expected that one participant should not be overly long
or overly short in what is said; one should say just the right amount to make the thought clear.
Grice’s maxims are not meant as prescriptive guides as to how conversation should progress,
or even statements about how they do progress, but rather a set of tacit conventions, which we
normally adhere to, and which we notice when they are broken.
The second factor concerns the common ground that the participants share. This is a general
term which covers the background knowledge, situation and mutual understand that the conver-
sational participants share. When in the same room, they can point to objects or refer to objects,
actions and so on, in a different way to how such things can be referred to if they are more distant.
If one says “have a seat” and there is a chair in the room, this means something straightforward.
The situation can significantly alter the intended meaning of an utterance. Take for instance the
sentence “my car has broken down”. Now this can be taken as a simple statement of fact, but if
someone says this upon entering a mechanic’s workshop, the implication is that the person’s car
has broken down and he’d like the mechanic to fix it. If by contrast, this is the first thing someone
says when arriving for dinner, the intention might be to apologise for being late because an un-
forseen travel delay occurred. To take another example, if two experts in the same research field
are talking, they can talk assuming that the other has a detailed knowledge of the subject matter.
This is different to when an expert talks to a lay person.
Finally, the knowledge of the channel helps form the message. Participants don’t generate
messages and then decided whether to speak them or write them down; they know what medium
is going to be used before they start. This allows the participants to best exploit the strengths
of each form of communication; in spoken language extensive use may be made of prosody; in
written language the author may take the liberty of explaining something in a complicated fashion
knowing that the reader can go back and read this several times.
2.3.2 Generation
In our model, the process of converting a meaning into a message is called message generation
(elsewhere the term language generation is often used, but this is too potentially confusing for
our purposes). To demonstrate how generation works, let us use the above mentioned communi-](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-44-320.jpg)
![Section 2.4. Discussion 23
as too hight.
2.3.5 Understanding
Understanding is the process of creating a thought from the message. Understanding can be
thought of as the reverse of generation and also bears similarity to decoding in the sense that we are
trying to recover a thought from a representation (the message), which also includes information
generated from other sources such as the background, situation and so on.
In understanding, we also have an ambiguity problem, but this is quite different from the
ambiguity problem in decoding. For example, it is perfectly possible when hearing a speech
signal to decode it into the correct message, but then fail to understand it because of ambiguity.
For instance, if someone says “did you see the game last night”, it is quite possible to correctly
identify all the words, but fail to determine which game is being referred to.
2.4 DISCUSSION
This chapter has introduced a model of communication and language that will serve as the basis for
the next chapter, which looks at text-to-speech conversion in detail, and the subsequent chapters,
which explore the various problems that occur and the techniques that have been developed to
solve them. We should stress that what we have presented here is very much a model; that is a
useful engineering framework in which to formulate, test and design text-to-speech systems.
In building this model, we have drawn from a number of fields, including semiotics, informa-
tion theory, psychology, linguistics and language evolution. These fields have overlapping spheres
of subject matter, but unfortunately are often studied in isolation such that none really gives as
complete and detailed a model of communication as we might wish for. Each of these separate
disciplines is a vast field in itself, so we have only touched on each area in the briefest possible
way.
Considering semiotics first, we should look at the works of Ferdinand de Saussure, who
is widely acknowledged to have invented the field of semiotics and is considered by many as
the first modern linguist. Saussure was the first to emphasise the significance of the arbitrary
nature of form and meaning in communication and proposed that the study of the relationship
between the two should be the central topic in linguistics [391], [392]. The philosopher Charles
Sanders Pierce also pioneered early work in semiotics and semantics, and gave us the framework
of iconic and symbolic semiotics which is used here [350]. The second main area we draw from
is communication theory or information theory, and from this we have the principles of encoding,
decoding, channels, messages and signals. The pioneering work in this field was produced by
Claude Shannon [400], who was originally interested in the problem of accurate transmission of
information along telephone wires (he worked for the telephone company AT&T). It soon became
clear that the implications of this work were much broader (Weaver [401]) and it has been widely](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-47-320.jpg)
![24 Chapter 2. Communication and Language
adopted as an approach to the analysis of signals arising from linguistic sources, such that this
information theoretic approach is now dominant in speech recognition and many other fields.
Saussure did not in general deal much with signals, and Shannon and Weaver did not deal much
with meaning, but their work clearly intersects regarding the issue of form and message (the term
form is used in linguistics and semiotics and message in information theory, but they both refer to
the same thing.)
The field of linguistics is often described as the science of language, but in the main this
field focuses very much on just the meaning/form verbal part of language, as opposed to the
entire process which includes participants, conversations, signals and so on. Linguistics generally
approaches language without reference to use; the idea is that a language is an entity and it can
be studied independently of any purpose to which it is put3. This approach is part tradition and
part pragmatic; as linguistics can trace its lineage back to the philosophical musings of ancient
Greek philosophers. In addition there is the practical justification that by isolating effects it makes
them easier to study. In modern times, the field of linguistics has been dominated by Noam
Chomsky[89], [90], [88], who in addition to many other innovations has championed the idea that
the primary function of language is not communication at all, but rather forms the basis of the
human mind; it evolved for this reason and its communicative purpose was a useful by product.
Despite this mismatch between linguistics and our more practical needs, this field does provide the
basics of a significant chunk of the overall picture of human communication, including the models
of phonetics, phonology, words, syntax and so on that are central to our problem. Many linguistics
text books tend to be heavily skewed towards one of the many incompatible theories within the
discipline, so care needs to be taken when reading any book on the subject. A good introductory
book which explains the “hard issues”, which is accessible to those with a scientific background
and which doesn’t follow any particular theory is Lyons [290].
The study of human communication in terms of participants, conversations, background
knowledge and so on is often performed outside the field of mainstream linguistics by psycholo-
gists and philosophers. Good books on human conversation and language use include Clark [95]
and Clark and Clark [96], Searle [395], Grice [184] and Power [361].
2.5 SUMMARY
Communication
• A semiotic system has three levels of representation; meaning, form, and signal.
• in affective communication, the meaning, form and signal are continuous and simply related.
• in iconic communication, the form resembles the meaning in some way. Iconic systems can
be either discrete or continuous
3 In fact in surveying more than twenty introductory text books in linguistics found in the Cambridge University book
shop, I did not find even one that mentioned that language is used for communication!](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-48-320.jpg)
![Section 3.1. Speech and Writing 27
evolved using sound as its medium is not clear, perhaps it was down to an ability to make distinct
patterns more easily with the mouth, than say by waving hands. It is important to realise that this
choice of sound may well have been a complete accident of evolution, today we communicate in
many different ways with vision either by reading or by interpreting sign language. For most of
human history, speech was the only (sophisticated) means of communication, but a second key
invention, perhaps 10,000 years ago, was that of writing.
To begin with, we need to look at the nature of speech and writing in more depth. We will
do this from a number of angles, looking first at the physical differences between these signals,
before moving on to look at how these differences influence communication.
3.1.1 Physical nature
The speech signal is a continuous, acoustic waveform. It is created by the operation of the vocal
organs in response to motor control commands from the brain. We only have a partial understand-
ing of how this process works, but it is clear that the coordinated movement of the vocal organs to
produce speech is extremely complicated. One of the main reasons why the relationship between
the speech signal and the message it encodes is that speech is continuous and the form is discrete.
In addition to this, speech is highly variable with respect to the discrete message as many other
factors, including prosody, speaker physiology, a speaker’s regional accent and so on all affect the
signal. A single message can generate many potential signals. Finally, speech is not permanent;
unless a recording device is present, once something is spoken, it is lost for ever and if no listener
is present or a listener does not hear the speech then the message that the speech encodes will
never be decoded.
Writing is fundamentally different from speech in a number of ways. Firstly writing is
primarily seen as visual: traditionally, writing always took the form of visible marks on a page
or other material, and today we are quite used to seeing writing on computer and other screens.
This is not however a complete defining property; we consider braille a form of writing and its key
benefit is of course that one does not need to be able to see to read it. It has been argued [199] that
the main defining feature of writing is that it is permanent in a way that speech is not. Whether
the writing is created with a pen on ink or impressions on a page, the key point is that once written
it does not disappear. We can of course destroy what we have done, but this is not the point; the
idea is that with writing we can record information, either so that we can access it at a later date,
or send it over some distance (as a letter for instance).
What then does writing record? Writing can be used to record natural language, but in saying
this we should make it clear that writing attempts to record the linguistic form of a message, rather
than the actual signal. That is, if we write down what someone says, we do not attempt to record
all the information in the speech, such that a reader could exactly mimic the original; rather we
decode the speech into a linguistic message, and then encode that. Of course quite often the
message we wish to write has never been spoken, and so when writing we generate a linguistic
form from our thoughts, and then encode this as writing without ever considering speech.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-51-320.jpg)
![Section 3.1. Speech and Writing 33
quipu system, in which numbers could be stored and arithmetic performed. This ability for writing
systems to encode non-linguistic information is therefore an ancient invention, and it continues
today in printed mathematics, balance sheets, computer programs and so on.[64], [19], [6]
Consider for a moment numbers and mathematics. Without getting to far into the debate
about what numbers and mathematics actually are, it is worth discussing a few basic points about
these areas. Firstly, in most primitive cultures, there is only a very limited number system present
in the language; one might find a word for one, two and three, then a word for a few and a word
for lots.
More sophisticated number communication systems (as opposed to numbers themselves) are
a cultural invention, such that in Western society, we think of numbers quite differently, and have
developed a complex system for talking and describing them. Most scientifically and mathemati-
cally minded people believe that numbers and mathematics exist beyond our experience and that
advancing mathematics is a process of discovery rather than invention. In other words, if we met
an alien civilisation, we would find that they had calculus too, and it would work in exactly the
same way as ours. What is a cultural invention is our system of numerical and mathematical no-
tation. This is to some extent arbitrary; while we can all easily understand the mathematics of
classical Greece, this is usually only after their notational systems have been translated into the
modern system: Pythagoras certainly never wrote x2 +y2 = z2. Today, we have a single commonly
used mathematical framework, such that everyone knows that one number raised above another
(e.g. x2) indicates the power. The important point for our discussion is to realise that when we
write x2 +y2 this is not encoding natural language; it is encoding mathematics. Now, it so happens
that we can translate x2 +y2 into a language such as English and we would read this as something
like “x to the power of two plus y to the power of two”. It is vital to realise that x2 + y2 is not
an abbreviated, or shorthand form of the English sentence that describes it. Furthermore, we can
argue that while mathematics can be described as a (non-natural) language, in that it has symbols,
a grammar, a semantics and so on; it is primarily expressed as a meaning and a written form only;
the spoken form of mathematical expressions is derivative. We can see this from the ease at which
we can write mathematics and the relative difficulties we can get in when trying to speak more
complex expressions; we can also see this from the simple fact that every mathematically literate
person in the western word knows what x2 + y2 means, can write this, but when speaking has to
translate into words in their own language first (e.g. EKS SQUARED PLUS Y SQUARED in English
IX QUADRAT PLUS YPSILON QUADRAT in German).
3.1.7 Semiotic systems
A semiotic system is a means of relating meaning to form. Natural language is a semiotic system,
and probably has by far the most complicated form to meaning relationship, and as we have just
seen mathematics is another. Most semiotic systems are much simpler than natural language as
they are cultural inventions rather than biological processes (recall that while mathematics itself
may be a property of the universe, the mechanism we use for its form and encoding are cultural](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-57-320.jpg)
![44 Chapter 3. The Text-to-Speech Problem
ware where extra licences (and hence cost) is required when the size of the deployment increases.
The main technical drawback of canned speech is that it can only say a fixed number of
things. This can be a severe drawback in even simple applications where for example a telephone
number is to be read in an answer machine application. A common solution is to attempt to splice
together recordings of individual words or phrases so as to create new utterances. The result of
such operations varies greatly from acceptable (but clearly spliced) to comically awful. That said,
even the resultant poor naturalness of this is often chosen over TTS in commercial situations.
Faced with the choice between fully natural but inflexible canned speech, and somewhat un-
natural but fully flexible TTS, some researchers have proposed limited domain synthesis systems
which aim to combine the benefits of both. There are as many different approaches to this. Some
systems attempt to mix canned speech and TTS. Black and Lenzo [52] proposed a system for clev-
erly joining words and carrier phrases, for use in applications such as a talking clock. The phrase
splicing approach of Donovan et al [139] used recorded carrier phrases and cleverly spliced with
unit selection synthesis from recordings of the same speaker. A somewhat different approach is
to use what is basically the same system as for normal unit selection synthesis but to load the
database of recordings with words and phrases from the required domain [440], [436], [394], [9].
3.5 KEY PROBLEMS IN TEXT-TO-SPEECH
In this section we identify the main challenges in high quality text-to-speech. First we describe
four areas which we believe are the key problems that have been recognised as the main goals of
TTS system building (but note, the problems are not always described in this way). Then we go
on to describe two additional problems, which are starting to attract more attention and which we
believe will be central to TTS research in the future.
3.5.1 Text classification with respect to semiotic systems
As we explained, it is a mistake to think that text is simply or always an encoding of natural
language. Rather, we should see text as a common physical signal that can be used to encode
many different semiotic systems, of which natural language is just one (rather special) case.
There are two main ways to deal with this problem. The first is the text-normalisation
approach, which sees the text as the input to the synthesiser and tries to rewrite any “non-standard”
text as proper “linguistic” text. The second is to classify each section of text according to one of
the known semiotic classes. From there, a parser specific to each classes is used to analyse that
section of text and uncover the underlying form. For natural language the text analysis job is
now done; but for the other systems an additional stage is needed, where the underlying form is
translated into words.
Let us consider the semiotic class approach. Assume for a moment that we can divide an
input sentence into a sequence of text tokens, such that the input sentence](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-68-320.jpg)
![46 Chapter 3. The Text-to-Speech Problem
The important point to realise is that there is often no single correct answer to the translation issue,
and so systems need to be sensitive to the user expectations.
A further difficulty with translation is that for some content there is no commonly agreed
rendering in words. While times, dates and currencies normally present no problem, classes such
as email addresses, computer programs and other technical information has no agreed reading. For
example, taking the email address:
(7) pat40@cam.ac.uk
We can easily parse this into
user = pat40
top level domain = uk
secondary domain = ac
local domain = cam
In effect as all email programs do just this. The problem is how should we read this? Should
pat40 be read P A T FOUR ZERO, P A T FORTY, or PAT FORTY?
In summary then, there are three tasks involved here.
1. Identify which semiotic class a token belongs to. This is often the most difficult task.
2. Analyse each token to find its underlying form. Apart from the case of natural language, this
is often quite easy as the semiotic systems are cultural inventions and have been designed to
be easy to understand.
3. Translate into natural language. It is often quite easy to generate at least one legitimate
translation but sometimes difficult to generate the particular version that a user expects.
3.5.2 Decoding natural language text
While most semiotic systems have a fairly simple and direct signal-to-form mapping, this in gen-
eral is not so for natural language. Once the text has been properly classified in terms of its
semiotic class, the general level of ambiguity with respect to form is greatly reduced; it is only in
the area of natural language where significant ambiguity still remains.
Ambiguity occurs in many different guises. The most basic is homograph ambiguity, where
two words share the same form. In cases such as
(8) lets now project the image
and
(9) the project was running behind schedule
we see that two separate words PROJECT-VERB and PROJECT-NOUN both share the same letter
sequence, but have different meanings and different pronunciations. In addition to homograph
ambiguity, syntactic ambiguity is also reasonably common. In the example (from [363])](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-70-320.jpg)
![48 Chapter 3. The Text-to-Speech Problem
What then makes one voice sound more natural than another? Again, this is a question that
we are not in a position to answer fully, but we can provide a sketch of the factors involved. Firstly,
any system that produces obviously non-human artefacts in the speech will readily be judged as
unnatural. It is all too easy to generate speech with pops, clicks, buzzes and an endless variety
of other mechanical sounds. Even speech that is absent from any of these “error” type of sounds
can readily sound unnatural. While speech exhibits considerable variability in the realisation of a
particular phoneme, this is controlled by a number of factors including phonemic context, position
in sentence, supra-segmental influence and so on. Getting this variation just right is a key task.
Beyond this, we can consider a number of issues which are related to individuality. In normal
speech, any given instance of speech is obviously speech from someone; there is no such think
as speaker-neutral speech. So in a sense, we are completely conditioned to the fact that speech
comes from people, all of whom sound slightly different. So when listening to a synthesiser, one
naturally thinks that it is from someone, and to sound natural, this someone must have a believable
sounding voice. This aspect of naturalness has virtually nothing to do with the message being
spoken, it is entirely a property of the system that encoded the speech. Hence we must make the
synthesiser sound like someone, either a copy of a real person’s voice, or the creation of a new
voice that could pass for a real person unproblematically.
3.5.4 Intelligibility: encoding the message in signal
The final central task is intelligibility. For our purposes, we define intelligibility as the ability of
a listener to decode the message from the speech. As such, this does not include any measure of
comprehension directly, but the idea is that if the listener can decode the message properly they
should be able to understand it with the same ability as they could the same message spoken by a
human. Because of that we can sometimes measure intelligibility with comprehension tests; this
is usually done with content that may be hard to decode, but is often not difficult to understand
(e.g. a simple sentence address or telephone number).
Of all the TTS problems, intelligibility is probably the easiest to solve. In fact, it is possible
to argue that this problem was solved some time ago, as was possible to generate fairly intelligible
sounding speech in TTS systems from the late 1970s onwards. In fact, it can be shown in tests
that the intelligibility of a modern TTS system is often only marginally better than much older
systems such as MITalk [10], once any text errors have been dealt with2. From this it is wrong
to conclude that little progress has been made on this issue. Crudely speaking, the story is that
while the intelligibility of older systems was reasonably good, research focused on improving
naturalness which was uniformly rated as being very poor with these systems. Hence the idea in
the research community (though rarely stated as such) was to improve naturalness without making
intelligibility any worse.
2 Today’s systems are massively more accurate with respect to text analysis, and so any text processing errors in
systems such as MITalk have to be eliminated if a comparison of synthesis encoding with respect to intelligibility is to
be performed.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-72-320.jpg)
![Section 4.4. Processing Documents 69
For less casual conventional writing, the situation can be more difficult. In email messages,
one finds that some writers do not use an upper cases characters at the start of a sentence, or that
new-lines are used as sentence boundaries and no ending punctuation is found. We also come
across the problem in email that the form of the headers is intended for email programs to parse
and process and not for reading aloud. So the notion of “sentence” is unclear here. To get around
this, we simply have to decide in what form we will read this information and so what form the
sentences will take in these genres. One approach is to have each line as a sentence, another is to
somehow process the header into a spoken list or other structure.
4.4 PROCESSING DOCUMENTS
4.4.1 Markup Languages
It is useful to be able to control the behavior of the TTS system at run time. While global parame-
ters can be set in a number of ways, often it is useful to have very precise instructions which affect
specified words. This is most effectively done by interleaving instructions to the TTS system with
the actual text to be spoken. In older TTS systems, it was common for each to have its own type
of markup. So for example in the Bell Labs system [411],, there are special sequences of escape
characters that help the user specify what behaviour the system is to follow.
The disadvantage of this is that these special characters would mean nothing to a different
TTS system, so a document marked up for one TTS system would not work with another. To
get around this, Taylor and Isard [435], [444] proposed the speech synthesis markup language
(SSML) which aimed to provide a general set of markup tags which could in principle be used
by any TTS system. This was extended by a number of further projects including the spoken text
markup language by Sproat et al [411], and the java speech markup language developed by
Andrew Hunt at Sun Microsystems. These served as the basis of the synthesis part of the W3C
standard VoiceXML (also called SSML, but not to be confused with the original).
All these schemes are based on the idea of generalised markup, first developed in the type-
setting industry. There, the idea of markup had been around for a long time, where a typesetter
would mark some text to be printed in 12 point, or in italics or so on. The problem with such
markup is that instructions is that they are physical, meaning that they are very specific to the type
of printing being used. If someone wished to print the document in a bigger font for instance, the
insistence on using 12 point (which may have seemed large in a 10pt document) means that the
intended effect is not achieved. Generalised markup attempts to get round this by splitting the type
setting into a two stage process. The author, or editor, adds logical markup to the text, for example
by indicating that something is a heading, or requires emphasis. In the second stage, this is then
rendered by a process much more specific to the actual printing mechanism being employed. As
the document itself only contains logical instructions, it is much more portable, and can be printed
on a number of devices and in a number of styles.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-93-320.jpg)
![72 Chapter 4. Text Segmentation and Organisation
the raw text for each sentence, and then passing this data structure through a series of modules
until a speech waveform is generated. We will term this the utterance structure.
In many early systems, the utterance structure was no more than a string that was passed
between modules. In doing this systems could adopt either an overwrite paradigm or a addition
paradigm. For instance in an overwrite system if the input text was
The man lived in Oak St.
The text normalisation module might produce output like:
the man lived in oak street
And the subsequent pronunciation module might then produce output like:
dh ax m a n l i v ax d ax n ou k s t r ii t
The point is, in each case, the module takes its input and only outputs the particular results from
that module. In an addition paradigm, each module adds its output to the output from the previous
modules. So in the above example, the output from the text normalisation module might produce
something like:
The|the man|man lived|lived in|in Oak|oak St|street .|null
And the subsequent pronunciation module might then produce output like:
The|the|/dh ax/ man|man|/m ae n/ lived|lived|/l ih v ax d/
in|in|/ax n/ Oak|oak|/ow k/ St|street|/s t r iy t/ .|null
The advantage of addition systems is that it is often hard to know in advance of development
what each module requires as input, and if we delete information we run the risk of supplying a
module with impoverished input. For example, we might find that the phrasing module (which
is called after the text normalisation module) requires punctuation, which can clearly not be used
if the information has been deleted. The trouble with string based addition systems is that they
can become unwieldy rather quickly. Even in the example above, with a short sentence and the
operation of only two modules, we see that the string is rather difficult to read. Furthermore, each
module will have to parse this string, which becomes increasingly more complicated as the string
becomes richer.
It is common therefore for modern systems to adopt an approach which follows the addition
paradigm, but does so in a way that allows easier and clearer access to the information. Two of the
most widely used formalisms are the heterogeneous relation graph or HRG formalism [441],](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-96-320.jpg)
![Section 4.5. Text-to-Speech Architectures 73
used in the Festival system, and the delta system [201], used in a number of commercial and
academic systems.
Both these are based on the idea of building data structures based on linguistic items2 An
item can be any single linguistic unit, including a word, phone, syllable, pitch accent or other.
The item is a feature structure which is a data structure common in computational linguistics
which defines a collection of properties as an unordered list of key-value pairs also known as
attribute-value pairs. Attributes must be atomic and drawn from a finite set of defined values.
Attributes must also be unique within the feature structure, so that when given an attribute, at most
one value will be found. From a computer science perspective, feature structures are very similar
to associative arrays or maps, and can also be though of as lookup tables or dictionaries. In
mathematics, they can be thought of as finite partial functions; that is, functions which have a
finite domain, and for which not all domain values are defined. An example feature structure is
given below:
word :
NAME abuse1
POS noun
TEXT abuse
PRON /@buws/
(4.1)
In the above example, all the values are atomic. It is also possible to have feature structures
themselves as values, which is useful for grouping particular types of information together:
phone :
NAME p
STRESS 1
DISTINCTIVE FEATURES
VOICED false
MANNER stop
PLACE bilabial
(4.2)
(4.3)
These items represent a single linguistic entity. As there are usually several of these in each
sentence we need an additional data structure to store these. In the HRG formalism, relations are
used to group items of a similar type. In the standard HRG formalism, we have three types of
relation: list, tree and ladder. Many types of linguistic data are best stored in lists, and this is how
word, syllable and phone information is normally stored. Other types of linguistic data are most
2 For clarity, we will adopt HRG terminology throughout our explanation.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-97-320.jpg)
![74 Chapter 4. Text Segmentation and Organisation
(c)
(b)
(a)
Figure 4.1 The three types of relation: a) list relation, b) tree relation and c) ladder relation
S
VP
NP
(a)
(b)
this is
parent
daughter1
daughter2
parent
daughter2
an example
daughter1
this is
next
previous
next next
previous previous
an example
daughter2
Figure 4.2 a) list relation containing word items. b) syntax relation containing syntax items
.
suited to trees, and this is how syntax, prosodic phrasing and syllable structure can be stored. Less
common is the use of ladders, which are used to represent connections between linguistic data
that is not organised hierarchically, as in a tree. The ladder structure can be used to implement
the data structures of autosegmental phonology [179], which is used for instance to represent the
relationship between syllables and intonation accents. The three types of structure are shown in
Figure 4.1.
One way of configuring the data structures is to have a separate relation for the output of each](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-98-320.jpg)
![76 Chapter 4. Text Segmentation and Organisation
s
word
phoneme
structure 1
syllable
structure 2
syllable
syllable
nucleus
rhyme
coda
onset
ch r
@
coda
k
u
r
t
onset rhyme
nucleus
syl
structure
syl
Figure 4.4 Thing
from any other item by navigating through the relations.
An alternative to the HRG formulation is the Delta formulation [201]. This mainly differs
in that the connection between the relations (called streams in Delta) is determined by a chart.
An example of a delta structure is shown in Figure 4.4. Here, each stream/relation is a list, but
navigation is performed by going to the edge of the item and ascending or descending through the
chart. The delta system is somewhat simpler and often easier to use than the HRG formalism, but
is somewhat less powerful in that it cannot represent general tree information.
4.6 DISCUSSION
4.6.1 Further Reading
The ontological status of linguistic entities was a prime concern of linguistics in the middle part of
the 20th century, but surprisingly has not been the subject of more recent work. Good references
that discuss the nature of words, sentences, phonemes and so on include Lyons [290] and Bloom-
field [57]. Tokenisation in general does not receive much attention in the literature as it is always
assumed to be a relatively minor (and simple) task.
The original SGML book is Goldfarb [178], which describes in detail the justification for
markup languages based on logical rather than physical markup. XML is by is inherently tied in
with the development of the web, and so good references for it can be found on the W3C website
(www.w3c.org). The original SSML and justification for its use in TTS can be found in Taylor
and Isard [444]. As the current version of SSML was developed as a web standard and references
to this can again be found on the W3C website.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-100-320.jpg)
![Section 5.2. Text Classification Algorithms 85
A → d|T
which reads A converts to d if a trigger token T is present in the sentence. The main difference
is that T does not have to be near to A. Often this approach is called a bag of features approach
as it uses evidence based on other tokens without considering their order in the sentence. Token
collocations are the most common type to consider to use as we can observe the tokens directly,
but it is also possible to use words as triggers also. These are obviously not directly observable
so we have to perform some classification before we can determine the words, and of course even
once found, our algorithm may not have chosen the word correctly.
If we consider some sentences with an ambiguous token , we can see quite clearly how
collocation information can help (from Yarowsky [507]).
... it monitors the lead levels in drinking... LEAD-METAL
... median blood lead concentration was ... LEAD-METAL
... found layers of lead telluride inside ... LEAD-METAL
... conference on lead poisoning in ... LEAD-METAL
... strontium and lead isotope zonation ... LEAD-METAL
... maintained their lead Thursday over ... LEAD-AHEAD
... to Boston and lead singer for Purple ... LEAD-AHEAD
... Bush a 17-point lead in Texas, only 3 ... LEAD-AHEAD
... his double-digit lead nationwide. LEAD-AHEAD
... the fairly short lead time allowed on ... LEAD-AHEAD
One simple rule based systems works as follows. Firstly, we require a lexicon which gives
the orthography for each word. With this, we can easily determine check whether two words in the
lexicon have the same orthography, and from this compile a list of tokens which have two or more
word forms. Next we write by hand a set of rules which fire (i.e. the activate) on the presence of
the trigger token, for example:
concentration LEAD-METAL
levels LEAD-METAL
singer LEAD-AHEAD
dog LEAD-AHEAD
At run-time, we then move through the sentence left-to-right examining each token in turn. When
we find one that is ambiguous (that is, it appears in our list), we look for trigger tokens at other
positions in the sentence which will form a collocation with the current token.
In longer sentences we may however find several collocation matches, and so exiting the list
on finding the first may result in an incorrect decision. One extension then is to search the entire
list and choose the word that gives the most matches. An alternative is to consciously order the list
in some way, with the idea that the choices at the top of the list are more definite indicators than
those else where. For example, if we find bass followed immediately by guitar, we can be](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-109-320.jpg)
![86 Chapter 5. Text Decoding: Finding the words from the text
almost certain that this is BASS 1 (rhyming with “base”) and not the type of fish. Another way of
using the rules is to run the most specific rules first, so if we have a rule which has several trigger
tokens, or a very rare trigger token (e.g. strontium) these fire before more general rules. The
idea here is that very specific rules are much more likely to be appropriate for that word only.
There are fundamental difficulties with regard to this and similar hand written rule ap-
proaches. Firstly, it can be extremely laborious to write all these rules by hand. Secondly, we
face situations where we have an ambiguous token but no trigger tokens that are present in our
collocation list. Both these problems amount to the fact that the mapping from the feature space
to the label is only partially defined. Finally, we face situations where say 2 rules match for each
word - which do we then pick? For these reasons, we now consider machine learning and statistical
techniques.
5.2.5 Decision lists
One of the weaknesses of the simple collocation list is that it doesn’t tell given any indication of
how strong a collocation might be. While we can use ordering (as in the case of bass guitar),
finding the correct ordering can be difficult and can still cause problems when contradictory rules
fire. An alternative is then to use statistics to give us an indication of how strong or how likely a
particular collocation is. One such method is to use decision lists, introduced by Rivest [373], and
used in TTS by Yarowsky [506], [507], Gale et al [169] [125] and Sproat [411].
The basic idea of decision lists is to determine the strength of the link between a token and
its trigger, and this is done by measuring the frequency of occurrence. In the following table (taken
from [507]), we see the raw counts which have been found for the token bass:
Trigger BASS-FISH BASS-MUSIC
fish 0 142
guitar 136 0
violin 49 0
river 0 48
percussion 41 0
salmon 0 38
from this it is clear that the trigger words are very strong disambiguation indicators, but also
that some triggers occur more frequently that others. From these counts, we can calculate the
conditional probability that of seeing a word given the collocation, e.g.
P(BASS-FISH|guitar,bass)
P(BASS-MUSIC|guitar,bass)
We then compute the log-likelihood ratios for each collocation:
abs
log
P(BASS-FISH|collocationi)
P(BASS-MUSIC|collocationi)](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-110-320.jpg)
![94 Chapter 5. Text Decoding: Finding the words from the text
emails pat40@cam.com
urls http://mi.eng.cam.ac.uk/ pat40
computer programs for (i = 0; i n; ++i)
addresses 4318 East 43rd Street
real estate CGH, sng whit fem
So it is important to have a good knowledge of the application area in order to ensure a good and
accurate coverage of the type of semiotic systems that will be encountered.
It should be clear that a simple classifier based on token observations will only be sufficient
in the simplest of cases. If we take the case of cardinal numbers, we see that it would be absurd
to list the possible tokens that a cardinal class can give rise to; the list would literally be infinite.
To handle this, we make use of specialist sub-classifiers which use regular expression generators
to identify likely tokens from a given class. One way of doing this is to run regular expression
matches on tokens and use the result of this as a feature in the classifier itself. So we might have
(34) Match [0-9]+ ?
as a rule, which will return true if the token is composed entirely of an arbitrary number of digits.
More specific regular expressions can also be used which handle numbers of a fixed length (e.g.
years in the modern era being 4 digits long,) or for real numbers which contain a decimal point.
In addition to these open class rules, we also see that semiotic classes form a hierarchy. So the
money class is composed of the number class and usually a currency symbol (e.g. $). A further
issue is that often several tokens form a coherent semiotic group, such that if we have
(35) 8341 Belville Dr
we have three tokens that together form a single address. Hence we need an ability to group the
tokens after or during classification.
The process of semiotic classification can therefore be summarised as:
1. Define the set S = {s1,...sN} of semiotic classes
2. For each decide if this is a closed or open class.
3. Closed classes can be enumerated, open classes require regular expression matches.
4. For each token:
(a) run the open class expressions and record the answer for each as a feature
(b) Using the open class features and the normal features, run a classifier (e.g. the naive
Bayes classifier), record the most probable class si for each token.
5. Group consecutive tokens of the same class together
6. Perform class-specific decoding and verbalisation
To take a simple example, suppose we have the sentence:
(36) You should meet him at 3:30pm at 1994 Melville Drive
and we have the set of semiotic classes as just defined. For each open class we would have a
regular expression as follows:](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-118-320.jpg)
![Section 5.3. Non-Natural Language Text 95
natural language [A-Za-z]+
cardinal numbers [0-9]+
telephone numbers 0131 525 6800, 334-8895
years 1[0-9][0-9][0-9] | 2[0-9][0-9][0-9]
time hours : minutes part-of-day
hours [0-9] | 1[0-9] | 2[0-4]
minutes [0-6][0-9]
part-of-day am | pm
addresses cardinal number proper-name street-type
proper-name [A-Z][a-z]+
street-type Dr| St| Rd| | Drive| Street | Road
For each token we run each regular expression and record its answer as a feature:
text natural language cardinal year hours minutes part-of-day proper-name
You 1 0 0 0 0 0 1
should 1 0 0 0 0 0 0
meet 1 0 0 0 0 0 0
him 1 0 0 0 0 0 0
at 1 0 0 0 0 0 0
3 0 1 0 1 0 0 0
: 0 0 0 0 0 0 0
30 0 1 0 0 1 0 0
pm 1 0 0 0 0 1 0
at 1 0 0 0 0 0 0
1994 0 1 1 0 0 0 0
Melville 1 0 0 0 0 0 1
Dr 1 0 0 0 0 0 1
The results of these matches can now be used as features, along with general features such as “is
this the first word in the sentence” (which is useful for telling a normal natural language word
from a proper name). All the features are then given to the classifier which makes a final decision
on semiotic class.
With regard to the issue of grouping, we can use a combination of the classifier and regular
expression matches to ensure that a sequence of tokens such as 3, :, 30, pm are seen as a single
time. A more sophisticated approach is available if we make use of the HMM tagging algorithm
as one of the features of an HMM is its ability to include multiple tokens in a single class, and
also, via the Viterbi search algorithm, ensure that the most probably global sequence of classes is
found, which helps resolves problems where two classes claim overlapping tokens.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-119-320.jpg)
![Section 5.4. Natural Language Text 101
5.4.2 Homograph disambiguation
The main problem in natural language text decoding is dealing with homographs, whether they be
accidental, part-of-speech or abbreviation homographs. Many systems choose to handle each of
these separately, and in particular abbreviation homographs are often with at the same time or in
the same module as semiotic classification. Typically this is done in systems which don’t have an
explicit semiotic classification stage. An alternative approach is to handle these in a single module.
Regardless, the general approach is to use one of the token disambiguation techniques described
in Section 5.2.1. Here we discuss some of the specifics of homograph disambiguation.
Turning to POS homographs first, we see that the classic way to resolve these is to use a POS-
tagger of the type described in Section 5.2.2. One problem with this approach is that unfortunately
POS-taggers only have a certain accuracy (96%-98% are commonly quoted figures [299]), and
often it is in the hard cases of resolving noun/verb pairs where the errors are made. Recall that the
POS tagger has two components, one which models sequences of tags and one which generates
tokens given a tag. The tag sequence n-gram is a very powerful constraint and nearly always gives
the correct answer in cases such as:
(47) The/det project/noun was/aux behind/prep schedule/noun
(48) The/det problem/noun is/aux how/wrb to/inf project/verb the/noun missile/noun
accurately/adv
but in sentences such as
(49) The new images project well on to the wall
POS-taggers often struggle as the n-gram sequence of tags alone is not sufficient to disambiguate in
this case. Part of the problem with POS taggers is that they are often formulated as general purpose
NLP tools, and as such are not focused on the goals of how they might be used in a particular
application. One simple improvement then is to adapt the tag set so that the discriminative power
in the algorithm is geared towards the distinctions we really care about.
An alternative approach is to use one of the word sense disambiguation algorithms as these
can be more easily focused for each homograph case. A further advantage in using the WSD
approach is that we can use this single approach for all type of homograph, not just the POS
homographs that the tagger deals with. To give an example, consider the abbreviation homograph
Dr which can encode DOCTOR or DRIVE. Both of these are often collocated with a proper noun,
but differ in their position relative to that proper noun. A simple feature that distinguishes many
examples is then whether the Dr occurs before or after the proper noun:
(50) Dr + proper-noun → DOCTOR
(51) proper-noun + Dr → DOCTOR
for example
(52) Dr Williams → DOCTOR WILLIAMS](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-125-320.jpg)
![Section 5.5. Natural Language Parsing 103
UNKNOWN-43), and the token and label are added to the run-time lexicon. The pronunciation is
determined by the grapheme-to-phoneme convertor, which we will describe in Section 8.4. These
set of operations can be found in the right half of Figure 5.3.
5.5 NATURAL LANGUAGE PARSING
Our concern so far has been the task of identifying the word sequence from the token sequence,
and to a large extent this is all the text decoding that we are required to do. Once we have un-
ambiguously identified the word sequence, we can use the lexicon and grapheme-to-phoneme
conversion to generate a sound representation which can then be input to the synthesiser itself.
It can be argued however that the word sequence alone is not enough, and that in addition it is
advantageous to conduct further processing to determine the underlying structure of the sentence.
Semantic analysis is generally thought too difficult and/or unnecessary, but a number of systems
perform a syntactic analysis which finds the correct or best syntactic tree for the input sentence.
There are several justifications for wishing to perform this additional level of analysis.
1. Proper parsing can resolve part-of-speech ambiguity problems that finite state (ie HMM
parsers) fail on, such as Example 49 above.
2. The syntax can help determine sentential prominence patterns (explained in Section 6.8).
3. The syntax gives a basic grouping of words, and these groupings strongly influence the
prosodic phrase structure of the utterance (explained in Section 6.7).
5.5.1 Context Free Grammars
In Chomsky’s early work [88], he proposed that the syntax of natural languages like English could
be described at a basic level by a context free grammar (CFG). A CFG defines a number of rules
of the form
(54)
S → NP VP
NP → DET NP
VP → V NP
VP → PP
PP → P NP
The rules operate generatively meaning that they are applied by starting with a basic top node (S
standing for “sentence”) and then applying rules in turn until a string of terminal nodes (ie words)
are generated. Chomsky’s idea was that this finite set of these rules could generate the near infinite
set of sentences in natural language. In addition, the operation of the rules defined the structure
structure of a sentence. The operation of context-free rules is easily represented by a syntactic
tree:](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-127-320.jpg)
![106 Chapter 5. Text Decoding: Finding the words from the text
The major weakness with this type of parsing is that it only considers syntactic factors, semantics
is completely ignored. That is why a parser often returns parses that are “correct” but seem silly.
Some more advanced models of syntax do incorporate semantic information [78], [177], [312],
but in general most syntax models, especially in engineering, only make use of syntactic rules.
A sensible alternative then is instead of trying to find the parse which the author intended, or
which makes most sense, we can try and find the most likely parse. That is, we build a probabilistic
model of the grammar and use the parser to find the most likely parse. There are a number of ways
this can be done, but the most common is to assign probabilities to the rules themselves so that
we say a non-terminal expands to each of its productions with a certain probability. When the
parsing is complete, each set of rules has an associated probability and the highest can be chosen
as the right answer. A number of specialist databases called tree-banks are available for training
PCFGs. These have trees and nodes labelled by hand, and from these a grammar can be trained and
probabilities assigned to each rule. For unlabelled data (that is sentences which we have no trees
for) the inside-outside algorithm can be used. This is the PCFG equivalent of the Baum-Welch
algorithm used to train hidden Markov models, and is explained in Section 15.1.8.
We can therefore identity three common approaches to parsing:
1. Define a CFG by hand and use a deterministic search algorithm (i.e. the parser) to find the
parse or parses which best fit the input
2. Define a CFG by hand and use a probabilistic search algorithm to find the parse or parses
which is most likely given the input
3. Learn the CFG from data and use this with a probabilistic search algorithm to find the parse
or parses which is most likely given the input.
5.6 DISCUSSION
Figure 5.3 shows the flowchart for the text decoding system described in this chapter.
At this stage, we will discuss a number of alternate formulations to the one presented here.
In the (distant) past, text forms such as Dr were often see as abberations as far as text and lan-
guage processing was concerned to the extent that these were often called text anomalies [276] .
The “real job” of text/language processing was to perform syntactic and morphological process-
ing. For instance in the MITalk book 6 pages are devoted to text anomalies and 48 to syntactic,
morphological and phonological processing. As can be seen from this chapter, in modern systems,
the opposite balance is seen. Another view is that this problem is one of text normalisation in
which the a form like Dr should be converted into its true text form doctor. In such systems
there is often no formal distinction between the concept of words and tokens. A more recent ap-
proach is to define such text forms as non-standard words (NSWs) [413]. This approach takes
the notion of these text forms much more seriously, and for example notes that the text forms for
a word are quite productive and can’t simple be listed. The NSW approach uses decision tress](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-130-320.jpg)
![Section 5.6. Discussion 109
lists of words, their pronunciations and their text forms. This means that morphological analysis
for instance, is much less necessary than before and that data driven methods are quite practical
in terms of training (we have the data) and run time (we can perform searches or run algorithms
with large numbers of parameters/rules). Thirdly, and perhaps most profoundly, pure linguistic
analysis such as syntactic parsing has fallen out of favour for the simple reason that it has failed
“to deliver”. Few doubt that a full, accurate and true description of a sentences linguistic structure
would be very useful, but what form exactly should this take? Modern linguistics has more theo-
ries of syntax than even specialists can keep up with, and disagreements run to the very nature of
what syntax is and what it represents (and there are now plenty who see it as an epiphenomenon).
From a more practical perspective, despite the use of tree-banks, there simply is no such thing
as a theory-independent model of syntax which we can use for engineering purposes. All use a
grammar and all such grammars make strong claims about the nature of language (such as whether
it is context free). Next, and maybe because of these problems, we find that even the best parsers
are really quite inaccurate and relying on their output can cause quite noticeable errors in other
parts of the system. Finally, we find that even if we had a good model of syntax and an accurate
parser, we still may not find syntactic analysis helpful. As we shall see in the next chapter, the
main uses for syntactic structure are in determining prominence, intonation and phrasing; but even
if we knew the syntax it is far from certain how to derive these other sorts of information. The
relationship between syntactic and prosodic structure is a great void in linguistics.
Many regard these problems with syntax and other types of “deep” linguistic analysis as ones
of expediency; with a more work in these fields we will gradually improve the theories, grammars
and parsers to the extent that accurate syntactic structures can be assigned to arbitrary text. An
alternative viewpoint is that these difficulties are genuinely profound, and given the enormous
work that has been put into the area of theoretical and applied syntax with so little solid progress,
that perhaps the nature of the problem itself is incorrectly formulated and that a new approach to
the organisational properties of language is required.
5.6.1 Further reading
In commercial TTS, text analysis is seen as vitally important as any error can cause an enour-
mous negative impression with a listener. Despite this, the problem has only received sporadic
attention amongst academic researchers. Notable exceptions include various works by Richard
Sproat [410], [411], [412], [413], who tackles nearly all the problems in text analysis (and many
in prosody, explained in the next chapter). The Bell Labs system that Sproat worked on is partic-
ularly well known for its very high accuracy with regard to text processing. In a separate line of
work Sproat et al investigated a number of machine learning approaches to semiotic classification
and verbalisation. David Yarowsky [506], [507], introduced an approach which showed the use of
the WSD approach to homograph resolution. A more traditional view is given in the MITalk book
[10].
Outside the field of TTS we find that a lot of the work in the field of natural language](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-133-320.jpg)
![114 Chapter 6. Prosody Prediction from Text
(60) ((IN TERMS) ((OF THE NECESSARY) (POLITICAL EXPEDIENCY))) (( THE RESULT) (( WAS
CLEARLY) (A GOOD ONE)) (FOR SHAW))
A question that has interested linguists for some time concerns the relationship between syn-
tactic phrases (ie. the syntax tree) and prosodic phrasing. While there is clearly some relationship
(easily seen in the above examples), finding the general nature of this is extremely difficult. Factors
which complicate the relationship include:
1. Prosodic phrasing seems “flatter” in that while syntactic phrasing is inherently recursive and
can exhibit a large degree of nesting, if levels in prosodic phrasing exist there are only a few,
and there is no sense in which the phrases within an utterance seem to embed so strongly.
2. Prosodic phrasing is to some extent governed by purely phonological, phonetic or acoustic
factors which can override the syntax. In the classic example, Chomsky [89] commented on
the fact that in
(61) THIS IS THE DOG THAT WORRIED THE CAT THAT KILLED THE RAT THAT ATE THE
MALT THAT LAY IN THE HOUSE THAT JACK BUILT
nearly every speaker says this with a flat prosodic phrasing like:
(62) THIS IS THE DOG | THAT WORRIED THE CAT | THAT KILLED THE RAT | THAT ATE
THE MALT | THAT LAY IN THE HOUSE | THAT JACK BUILT
whereas the syntactic phrasing is deeply recursive and embedded
(63) (THIS IS (THE DOG ( THAT WORRIED THE CAT (THAT KILLED THE RAT (THAT ATE
THE MALT (THAT LAY IN THE HOUSE (THAT JACK BUILT)))))))
The speech patterns of the utterance seem to override the syntactic patterns.
3. Phrasing is particularly prone to speaker choice, and while “rules” exist for where phrase
breaks may or may not occur, quite often it appears optional whether a phrase break will be
found in a particular utterance. Hence all of the following are possible:
(64) JOHN AND SARAH WERE RUNNING VERY QUICKLY
(65) JOHN AND SARAH | WERE RUNNING VERY QUICKLY
(66) JOHN AND SARAH WERE RUNNING | VERY QUICKLY
So as we can see the relationship between syntactic and prosodic structure is complicated and
a general theory which links syntax and prosody is yet to be developed. Some have even argued
that the difficulty really lies with our models of syntax, and that if we developed syntactic models
which took more account of the ways sentences were spoken some of these problems might be
resolved [1].
6.2.2 Models of Phrasing
While there is no single accepted theoretical model of phrasing, all models can be encompassed
within the idea that a sentence has a prosodic phrase structure. The main point of disagreement](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-138-320.jpg)
![Section 6.2. Phrasing 115
revolves around the details of this phrase structure, and in particular how nested or deep it should
be.
Some have proposed that phrasing should be represented by a tree and that this should act
as a parallel tree to the syntax tree. As such the prosodic tree can be generated by a CFG, and so
can have recursion and arbitrary depth. Perhaps the most elegant model of this type is the model
of metrical phonology from Liberman [284], [286]. The central idea is that any sentence, or
part of sentence has a metrical tree which specifies its phrase and stress patterns. The lowest unit
in the tree is the syllable, and so metrical phonology provides a unified theory of syllable, word
and phrasal patterns. Ignoring the syllable level for the moment we can see the general form of a
metrical trees example below:
W
W
was
S
clearly
S
W
W
a
S
S
good
W
one
S
W
for
S
shaw
The S and W labels will be explained later, but for now simply note that the metrical tree
groups words together. This model seems particularly good at explaining the patterns within com-
pound noun phrases, which we will re-examine in Section 6.3.1. A major criticism though is that
the theory implies unlimited depth in trees, which implies that a) there are many graded subtleties
in the actual production or perception of phrase breaks, or b) that this representation maps to
another surface representation that is flatter and has fewer levels. Many solutions have been pro-
posed, both as amendments to the original theory2 and as alternatives. The theory of compound
prosodic domains proposed by Ladd [268] allows a more limited recursion with arbitrary depth
but not full recursion. Another variant, called the strict layer hypothesis [398] is much more
constrained and states that every word has a fixed number of prosodic constituents above it. This
and many other theories can be seen as forming a model called the prosodic hierarchy. In this,
we have a tree as before, but with a fixed number of levels, each of which is its own type. A rich
version of this might include the levels utterance, intonational phrase, phonological phrase, word,
foot, syllable and segment. In this intonational phrase means a large phrase with a clear intona-
tional tune (see below), a phonological phrase is a smaller phrase that spans only a few words,
2 The field of metrical phonology still survives today, but unfortunately bears little resemblance to the elegance and
power of the original theory.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-139-320.jpg)
![116 Chapter 6. Prosody Prediction from Text
and a foot is a unit which starts with a stressed syllable and covers all the unstressed syllables until
the next stressed syllable.
It is reasonable to say that nearly all other models are of this general type, and differ only
in that they use differing numbers of levels or names for those levels. One of the most popular,
especially for engineering purposes has 2 levels of sentence internal phrase, giving 4 levels over
all:
1. Utterance
2. Intonational phrase/ Major phrase
3. Phonological phrase/ Minor phrase
4. Word
in this scheme, an utterance might look like
Major
Minor
the result
Minor
was clearly
Minor
a good one
Minor
for shaw
A slightly different view is to not concentrate on the constituent phrases, but rather the
breaks between the phrases. This view has been around for some time, but was formalised most
clearly in the ToBI model in the concept of break index [406]. The break index is a number which
exists between every pair of words, for example:
(67) IN 1 TERMS 3 OF 1 THE 1 NECESSARY 3 POLITICAL1 EXPEDIENCY 4 THE 1 RESULT 3
WAS 1 CLEARLY 3 A 1 GOOD 2 ONE 3 FOR 1 SHAW 6
The numbers give the strength of the break such that between two normal words in running
speech we have a 1, and for words which are deemed more separate we have 2 and so on, extending
up to 6. A simplified version of this is frequently used, where 3 is given to a minor phrase break
and 4 to a major phrase break. Break index systems and strict depth tress are isomorphic, and so
the distinction as to which is used is often more of expediency rather than deep rooted theory.
A major problem in resolving issues with prosodic phrasing is that there always seems to
be a tension between what we can describe as a top down approach, where we use the linguistic
structure of the sentence to dictate the phrasing, and a bottom up approach, where we listen to real
speech and use clues from pausing, timing patterns and pitch to show where the phrases should
be. No model yet proposed has come up with a satisfactory method of doing both of these and
reconciling the results.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-140-320.jpg)
![Section 6.3. Prominence 119
This is often explained by the notion that the main stress on BOYS “repels” the stress on -TEEN so
that stressed syllables do not directly occur as neighbours.
6.3.2 Discourse prominence patterns
When sentences are spoken in discourses or dialogues, additional patterns can occur. One com-
monly cited affect is the difference in prominence levels between given and new information. This
idea says that when a word is first used in a discourse, it will often be quite strong, but on second
and subsequent mentions it is actually weaker, so we might have:
(84) FIRST MAKE SURE THAT PAUL ANSWERS THE PHONE.
(85) THEN TELL PAUL TO BRING THE BOOK OVER TONIGHT.
in which the first mention of PAUL is prominent and the second is not. While this given/new
distinction has traditionally received considerable attention in the literature, many recent empirical
studies on real speech have shown that this affect is not so common as traditionally though [328],
[329], [23], [330]. First, it is far more common that the second sentence would have a pronoun
instead of the original name
(86) THEN TELL HIM TO BRING THE BOOK OVER TONIGHT.
and secondly, as Aylett [24] points out, even when words are repeated in full form in dialogues,
there is very little evidence of prominence patterns being governed by the given/new distinction.
What does occur, is that the degree of prominence is heavily influenced by the redundancy in the
dialogue, such that if the speaker feels the listener has a high probability of understanding then he
or she may well give a word less prominence. Crucially, the word receiving less prominence does
not have to be the same as the original word, hence
(87) THEN TELL THE IRISH GIT TO BRING THE BOOK OVER TONIGHT.
here as IRISH GIT is obviously related to PAUL it may receive less prominence. Prominence
patterns of these sort are clearly augmentative and will be explained further in Section 6.5.3.
We can identify a second type of discourse prominence effect which we shall term emphasis.
This occurs when a speaker wants to draw attention to a particular word:
(88) IT WAS REALLY HUGE
(89) I DIDN’T BELIEVE A WORD HE SAID
(90) I’VE HEARD IT ALL BEFORE
(91) THATS JUST TOO BAD
(92) ITS IMPORTANT TO DO THAT BEFORE YOU ENTER THE ROOM
(93) NOT THAT ONE
and so on. In general, emphatic prominence is used for augmentation purposes (see below in
Section 6.5.3) where the speaker wants to make particularly sure that the listener understands
what is being said. A particularly salient example is contrastive prominence of the type:](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-143-320.jpg)
![Section 6.3. Prominence 121
also neatly deals with the issue of prominence shift, which occurs in the following example for
many speakers:
W
S
city
W
hall
S
S
parking
W
lot
(the prominence has shifted from HALL to CITY). Just as with phrasing however, it seems difficult
to believe that when it comes to a whole sentence such as:
W
W
W
I
S
left
S
W
my
S
car
S
W
in
S
W
the
S
W
S
city
W
hall
S
S
parking
W
lot
]
that speakers can make and listeners discern all these different degrees of phrasing. This has lead to
researchers proposing a fixed number of types or levels of prominence, but as with phrasing, there
is virtually no agreement as to how many there should be. Probably the most common models for
practical engineering purposes are:
1. a metrical tree in which the values on the high level nodes are ignored
2. a system with three levels of prominence
(a) prominent/stressed
(b) un-prominent/unstressed
(c) reduced
3. as above, but with an extra level for emphatic prominence.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-145-320.jpg)
![124 Chapter 6. Prosody Prediction from Text
if you ask a child whether they ate all the cookies, they find it easy to say NO but less hard to say
this convincingly).
Affective prosody has a certain degree of language universality. In all languages, raising
one’s voice and shouting are seen as expressions of angry or hostile states, while speaking in a
softer voice is seen as conciliatory or calming. In fact the relationship is so direct that often we
say “she spoke in a calm voice” etc. Affective prosody is not completely universal however, and in
terms of primary emotion, there are significant differences across languages. Even within English
it is common to find that one group of speakers are regarded as brash and another meek, when
in fact all that is different are the “settings” and “thresholds” with which the basic emotions are
expressed.
In general, affective prosody has a fairly direct meaning-to-form relationship and hence the
more extreme the form the more extreme the emotion. This implies that, unlike verbal language,
the affective prosody system is not arbitrary with regard to meaning/form. A number of emotion
categorisation systems have been proposed. One common theory holds that there are six basic
emotions (anger, happiness, sadness, disgust, fear and surprise) from which follow a number of
secondary emotions such as puzzlement or intrigue [110], [113], [153]. Some however argue
that there is no fixed set of emotions and rather emotion is best expressed in a multi-dimensional
system. One such proposal advocates three independent axes of emotion called activation, evalu-
ation and power [393], [126].
In addition to emotion, affective prosody includes speaker attitude. The difference between
emotion and attitude is that only emotion can exist independent of communication. As an illustra-
tion; you can be happy on your own, but you need someone else to be sarcastic with. Apart from
this distinction, attitude functions similarly to emotion in that it is fairly direct, has a significant
degree of universality and does not have a arbitrary or dual nature.
One point of debate in the literature on both emotion and affective prosody concerns whether
there is such a state as “no emotion” or “no affective prosody”. The argument is that if we are able
to increase the level of intensity of anger, then we should be able to decrease this to the point
where there is no anger. Similarly with the other emotions. This issue is important as it seems
that the total affective content of an utterance can vary greatly and can be contrasted with the total
propositional content of the verbal part of the utterance. These two effects seem to trade-off such
that we can have sentences that are nearly purely affective “thats amazing” and others which are
purely propositional “the sum of the square of the opposite and adjacent sides...”. This later type
of sentence is what we have previously termed discourse neutral.
6.5.2 Suprasegmental
If a sentence is spoke with little or no affective content, ie. in a discourse neutral manner, we still
see characteristic patterns in the phrasing, rhythm, pitch, voice quality and timing. Typical effects
include phones at the ends of sentences or phrases being lengthened, syntactically salient words
(e.g. heads) having more emphasis, F0 levels being higher at the starts of sentences and so on.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-148-320.jpg)
![126 Chapter 6. Prosody Prediction from Text
munication (whereby they would use the least effort in generating their utterances) with a desire
to be effective, whereby they wish to ensure the message is correctly decoded and understood.
In an ideal communicative setting, where the participants know each other, can see each
other and are able to converse without background noise or other distractions, it is nearly always
the case that the verbal message can be communicated without need for any augmentative prosody.
Whether any prosody is used therefore is purely a matter of the affective nature of the utterance.
Successful communication is possible because the verbal channel is highly redundant, to the extent
that the speaker should always be able to decode and understand the message. In other conditions
we face a different situation, so for example in noisy conditions we might see that the speaker
speaks with greater vocal effort and more slowly. Likewise we may see the same effect if the
speakers don’t know each other so well, or in more polite situations in which the speaker wants to
be seen to be making a clear effort to communicate effectively.
While it may be fairly obvious that increased articulation effort will lead to a higher prob-
ability of a phoneme string being correctly recognised, it is with other aspects of augmentative
prosody that the situation becomes particularly interesting. For example, it is frequently asserted
that yes/no questions such as
(97) IS THIS THE WAY TO THE BANK?
having rising intonation in English. In fact empirical studies of data show that the most common
intonation for such sentences is exactly the same pattern used for statements. While speakers
can use rising intonation, they often don’t. This behaviour is best explained with reference to the
communicative model. In examples such as 97, it is usually obvious that a yes/no question is being
asked, there really is little need for a speaker to use the additional cues that a rising intonation
pattern would provide. However if the speaker believes she may be misunderstood, or if she is
trying to sound polite, the rising intonation pattern may appear. By comparison, it is common to
ask “yes/no questions” that are have the same syntax as statements, but in which the speaker does
now in fact use a rising intonation pattern to make sure these are interpreted correctly:
(98) THIS IS THE WAY TO THE BANK?
We see this effect in all aspects of prosody. The given/new distinction, where it is often claimed
that subsequent mentions of an entity are less prominent than first mentions, is rarely observed in
real speech, partly because it is quite clear from the verbal content which in fact is intended [328],
[24].
In classic example sentences such as
(99) BILL DOESN’T DRINK BECAUSE HE’S UNHAPPY
which are spoken without phrase breaks may potentially be ambiguous but in general the listener
will know enough of the dialogue context to correctly decode this. If however the speaker is less
sure of a correct decoding, phrase breaks may be inserted to aid communication, and likewise with
every other aspect of prosody.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-150-320.jpg)
![130 Chapter 6. Prosody Prediction from Text
6.7 PHRASING PREDICTION
First we will examine the issue of phrasing prediction which is how to generate a prosodic
phrase structure for a sentence from the text. For purposes of illustration we will adopt the model
described in Section 6.2.2 where we have a major phrase and a minor phrase and this is probably
the most commonly used in TTS.
6.7.1 Experimental formulation
To help in our exposition, we will formalise the problem in the manner suggested by Taylor and
Black [443]. In this, our input is a list of tokens:
T = t1,t2,t3,...,tn,
and between every token is a juncture, so that a full sentence is:
T = t1, j1,t2, j2,t3, j3,..., jn−1,wn,
The job of the phrasing algorithm is to assign values to every ji drawn from an inventory of
juncture types, which we will define to be { non-break, minor-break, major-break, sentence}.
About 1 in 5 junctures are of type non-break (that is the gap between normal words) and so we
have to be careful in assessing the accuracy of any algorithm - if we use one which always assigns
non-break we will have achieved 80% accuracy on a simple junctures correct measure. Because
of this, Taylor and Black proposed a number of measures, later extended by Busser et al [77].
These are calculated from the following basic quantities:
Breaks B number of breaks in the corpus
Insertions I a non-break has been incorrectly marked as a break
Deletion D a break has been incorrectly marked as a non-break
Substitution S a break of one type has been incorrectly marked as a break of another type
Number of junctures N total tokens in the corpus -1
These are then used to calculate the following scores:](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-154-320.jpg)
![132 Chapter 6. Prosody Prediction from Text
2. deterministic content, function (DCF): Place a phrase break every time a function word
follows a content word.
The first algorithm, DP, works very well in sentences of the type:
(101) The message is clear: if you want to go, you have to go now
For DP algorithm to work, we must have an accurate means of determining underlying punctua-
tion. This is so that sentences like
(102) But really!? You can go -- but it won’t do you any good.
have the punctuation sequence !? and --mapped to a single underlying form because we don’t
want to have an additional phrase break between the ! and the ?. The DCF algorithm works well
in sentences such as
(103) It was taking a long time to gather all the information
where we place a phrase break after every function word that follows a content word:
(104) it was taking | a long time | to gather |all the information
These can be combined into a single algorithm, deterministic content function punctuation
(DCFP), which places a break after every content word that precedes a function word and also at
punctuation. The algorithms were tested on a large corpus of read speech which had phrase breaks
labelled by hand. The results for these algorithms given in Taylor and Black [443] are:
Algorithm breaks correct junctures correct juncture insertions
Deterministic Punctuation 54.274 90.758 0.052
Deterministic Content Function Punctuation 84.40 70.288 31.728
These figures basically show that that the DP algorithm massively under predicts and the DCFP
algorithm massively over predicts. The DP algorithm result is worth examining further because
it shows that placing a phrase break at a punctuation mark is in fact a very safe thing to do; this
nearly always coincides with a real break. The only problem is that many other breaks are missed.
The DPCF algorithm over predicts far too much, which lead to the idea that if finer distinctions
could be made than just content word and function word, a more accurate algorithm would be
possible.
A number of more sophisticated deterministic systems have been proposed [26], [186], [10]
which make of rules for specific cases. For example, the verb balancing rule of Bachenko and
Fitzpatrick [26] works through a sentence left to right and compares the number of words in a
potential phrase formed with the verb and the syntactic constituents to the left, and the number of
words in the constituent to the right. The potential phrase with the shortest number of words is
chosen as the correct one.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-156-320.jpg)
![Section 6.7. Phrasing prediction 133
6.7.3 Classifier approaches
The general historical trend in phrase break prediction is the same as with most problems we ex-
amine: as more data has become available, researchers have moved away from rule based systems
towards trainable systems. This is partly because the new data showed how errorful rule based
systems were, and partly because it allowed algorithms to have their parameters and rules auto-
matically inferred from the data. We now turn to data-driven techniques, where interestingly we
see a similar pattern of algorithm type to that used in text classification. In this section we examine
local classifier algorithms, which share many features in common with WSD algorithms. In the
next section we examine HMM approaches, which are very similar to POS taggers.
Wang and Hirschberg [363] introduced the idea of using decision trees for phrase break
prediction. Decision trees allow a wide variety of heterogeneous features, examples of which are
given below:
• total seconds in the utterance
• total words in the utterance
• speaking rate
• time from start of sentence to current word
• time to end of sentence from current word
• is the word before the juncture accented?
• is the word after the juncture accented?
• Part-of-speech
• syntactic category of constituent immediately dominating the word before the juncture
• syntactic category of constituent immediately dominating the word after the juncture
These features (and others) were used as question in the tree, which was grown in the normal way.
Wang and Hirschberg trained and tested their system on 298 utterances from the ATIS corpus,
giving the following results (the results aren’t available in the formats we described above):
Boundaries correct 88.5%
Non-Boundaries correct 93.5%
It should be pointed out that this approach can’t strictly be used for TTS purposes as acoustic
features (e.g. time in seconds) measured from the corpus waveforms were used in addition to
features that would be available at run time. Following this initial work, a number of studies have
used decision trees [264] [418], and a wide variety of other machine learning algorithms have
been applied to the problem including memory based learning [77] [402], Bayesian classifiers
[516], support vector machines [87] and neural networks [157]. Similar results are reported in
most cases, and it seems that the most important factors in the success of a system are the features
used and the quality and quantity of data rather than the particular machine learning algorithm
used.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-157-320.jpg)
![134 Chapter 6. Prosody Prediction from Text
0
50
100
150
200
250
0 2 4 6 8 10 12 14 16 18 20
Figure 6.2 Histogram of phrase lengths.
6.7.4 HMM approaches
The HMM approach to phrase-break prediction was introduced by Taylor and Black [443] and has
been extended by others [37], [369]. In their basic system, Taylor and Black proposed that the
decision on whether a phrase break should or should not be placed was dependent on two main
factors:
1. The parts-of-speech surrounding the juncture
2. The distance between this proposed break and other breaks
The first of these factors is simply a restatement of what we have already examined; i.e. that
syntactic patterns have a strong influence on phrase break position. The second factor is used in the
classification approaches, but is given particular importance in the HMM model. The hypothesis
is that in addition to any high level syntactic, semantic or dialogue factors, phrasing operates in a
semi-autonomous fashion in that speakers simply like to place phrase breaks at regular intervals.
This can be explained to some extent by the fact that speakers need to pause to get breath, but it
seems that the phrase breaks occur much more frequently than is needed for breathing purposes
and so this alone can not explain the patterns. Regardless of why however, it is clearly the case that
speakers do exhibit strong tendencies to place phrase breaks at regular intervals, a fact noted by
many researchers in this area [26], [156]. Figure 6.2, which shows a histogram of phrase lengths
in words, clearly demonstrates this pattern.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-158-320.jpg)
![Section 6.8. Prominence Prediction 137
6.7.5 Hybrid approaches
More recently a number of approaches have been proposed which combine the advantages of
decision tree approaches (use of heterogeneous features, robustness to curse of dimensionality)
and the HMM approach (statistical, global optimal search of sequences). In addition, there has
been somewhat of a re-awakening of use of syntactic features due to the provision of more robust
parsers. Rather than attempt an explicit model of prosodic phrasing based on trying to map from
the syntax tree, most of these approaches use the syntax information as additional features in a
classifier [508], [209], [257].
While the pure HMM approach uses a discrete probability distribution over observations
(in our case POS tags), in practice, all we need is a means of finding the likelihood of seeing
the observation given the juncture type P(C|J). Many have therefore tried alternative ways of
generating this. One of the most popular is actually to use decision trees, as these can be used to
given probabilities as wall as simple decisions on classification. As it stands, a decision tree gives a
posterior probability, not a likelihood, but a likelihood value can be obtained by either ensuring that
the tree is trained on an equal amount of break and non-break data, or by normalizing the posterior
by the basic frequency of breaks and non-breaks. The advantage in this combined approach is that
we can use any features that require, but also perform the global Viterbi search using n-grams to
find the best overall sequence of breaks [403], [369], [423].
6.8 PROMINENCE PREDICTION
Prominence prediction algorithms follow the same basic approaches as phrase break prediction
where we have simple deterministic algorithms, sophisticated deterministic algorithms and data
driven algorithms.
6.8.1 Compound noun phrases
The simplest prominence prediction algorithm simply uses the concatenation of the lexical promi-
nence patterns of the words. So for:
(105) IN TERMS OF THE NECESSARY POLITICAL EXPEDIENCY TO ENSURE SURVIVAL — THE
RESULT WAS A CLEARLY A GOOD ONE FOR SHAW
we might generate a prominence pattern of:
(106) IN TERMS OF THE NECESSARY POLITICAL EXPEDIENCY TO ENSURE SURVIVAL THE
RESULT WAS A CLEARLY A GOOD ONE FOR SHAW
The main ways in which this rule fails (again ignoring affective and augmentative effects) are
in compound noun phrases and certain function words uses. The prominence patterns in com-
pound noun phrases in English are complex and can seem quite baffling. When we consider the
prominence patterns in street names such as](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-161-320.jpg)
![138 Chapter 6. Prosody Prediction from Text
(107) WINDSOR AVENUE
(108) BEAUFORT DRIVE
(109) PARK LANE
it seems quite obvious what the pattern is, until we consider examples with the word STREET
(110) WINDSOR STREET
(111) BEAUFORT STREET
(112) PARK STREET
in which the prominence pattern is reversed. Sproat [410] conducted a very thorough study into
compound noun phrasing patterns, and concluded that in general, there is no simple syntactic
pattern which governs prominence. Instead he proposed an extensive set of rules based partly on
synactic and partly on semantic patterns. The most basic rule says that compound noun phrases
with two words assign their prominence according to whether the two words are taken as a single
compound “word” or a phrase. In general in English, phrases receive their main prominence on
the last word in the phrase. English words, and particularly nouns, by contrast tend to have their
prominence on the first syllable. This can be seen in the difference between
(113) BLACKBIRD meaning the type of bird
(114) BLACKBIRD a bird which just happens to be black
This principle can then be generalised to cope with all two word cases, with the idea being
that only conventionalised cases are treated as compound words which receive their prominence
on the first word. In addition to this, there are a number of semantically driven rules , which range
from the general to the more specific:
(115) Furniture + Room → RIGHT, e.g. kitchen table
(116) Proper-name + street → LEFT, e.g. Park street
For more compound noun phrases involving three or more words, a CYK parser (see Section
5.5) is used to assign the internal structure to the noun phrase, after which prominence values can
be assigned. To resolve ambiguity, a set of heuristics is used to find the best parse. These are:
1. Prefer parses determined by semantic rather than syntactic means. (This partly works be-
cause semantic parses are much more specific)
2. Choose phrase based parses over compound word bases parses
3. Choose parses with right branching structure.
Additional rules are used to shift prominence in cases where the parses would produce two
prominent syllables side by side.We also find that many types of non-natural language entities
have characteristic patterns. For example, it is common place to find telephone numbers read
with a particular stress pattern (and indeed phrasing). Producing the correct stress patterns for
these is often fairly straight forward and the creation of these patterns can be integrated into the
verbalisation process.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-162-320.jpg)
![Section 6.8. Prominence Prediction 139
6.8.2 Function word prominence
Less work as been carried out on the second main problem of determining the prominence patterns
in function word sequences. The most we can say about this is that not all function words are equal
when it comes to propensity for reduction. At one extreme, we have determiners (THE and A)
which in their default form are always reduced, and at the other extreme closed class words such
as HOWEVER, WELL and SO can receive the most prominence in a sentence. Given a sequence of
content words, a reasonable approach is then to give each a score according to where it comes in
the strength list, and starting with the strongest, add prominence at regular intervals. A possible
hierarchy, from weakest to strongest, might be:
articles THE, A, AN
auxiliary verbs WAS, HAS, IS
prepositions TO, FROM, BY
particles IF, THEN, WELL
conjunctions THAT, AND, BUT
pronouns HE, HIM, SHE, ME, THEM
modals WOULD, COULD, SHOULD
Note that auxiliary verbs (e.g. HAVE) can function as main verbs also. We find a fairly clear
pattern that when used as an auxiliary the word is often reduced, when used as a main verb it is
often unreduced or prominent. With this table, we can then use a rule which states that prominent
syllables should not occur side by side, and so in a sequence of function words we first assign
prominence to the strongest word given in the table, ensure its neighbours and reduced and then
look for the next strongest word. This then can give the stress pattern for the sentence we saw in
Section 6.3.1:
(117) WHAT WOULD YOU HAVE DONE IF IT HADN’T BEEN FOR ME?
6.8.3 Data driven approaches
Prominence prediction by deterministic means is actually one of the most successful uses of non-
statistical methods in speech synthesis. This can be attributed to a number of factors, for example
the fact that the rules often don’t interact or the fact that many of the rules are base on semantic
features (such that even if we did use a data driven technique we would still have to come up
with the semantic taxonomy by hand). Sproat notes [410] that statistical approaches have had
only limited success as the issue (especially in compound noun phrases) is really one of breadth
and not modelling; regardless of how the prominence algorithm actually works, what it requires
is a broad and exhaustive list of examples of compound nouns. Few complex generalisations are
present (what machine learning algorithms are good at) and once presented with an example, the
rules are not difficult to write by hand.
That said, with the provision of larger, labelled corpora, the natural progression has still
been towards using machine learning techniques to predict prominence. We should probably not](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-163-320.jpg)
![140 Chapter 6. Prosody Prediction from Text
be surprised to find that these techniques follow the basic paradigm used for phrase break predic-
tion, where we obtain a corpus of speech, label the prosodic content of this by hand (in this case
prominence levels) and train an algorithm to predict these values from features available to the
system at run time. Approaches tried have included decision tress [207], [378], [147], memory
based learning [303], transformation based learning [200] and neural networks [72], [323], [389].
6.9 INTONATIONAL TUNE PREDICTION
Here we have explicitly separated prominence from the tune part of intonation, and so for our pur-
poses intonation prediction is specifically the prediction of intonational tune from text, rather than
the broader definition of this problem that encompass predicting tune, accentuation and sometimes
phrasing.
Of all the prosodic phenomena we have examined intonational tune is the most heavily re-
lated to augmentative and particularly affective content. In situations where these effects are ab-
sent, we can say to a first approximation that all utterances have in fact the same intonational tune;
the only differences occur as to where the pitch accents and boundary tones which make up this
tune are positioned. Hence we can almost argue that for discourse neutral synthesis, there simply
isn’t any intonational tune prediction to be done. In other words, the real task is to predict a suitable
F0 contour that expresses the prominence and phrasing patterns and encodes the suprasegmental,
rather than true prosodic patterns of the utterance.
While we can describe prominence and phrasing in quite abstract high level terms, this is
significantly harder with intonation as all theories to a greater or lesser extent make explicit refer-
ences to F0 patterns, levels and dynamics. Given this, and the fact that for the most part we are
generating discourse neutral suprasegmental intonation, we will leave the entire topic of intonation
until Chapter 9.
6.10 DISCUSSION
At first glance it may seem that predicting prosodic form from text is an impossible task. Two
significant barriers stand in our way; first, the text is greatly under specified for the type of in-
formation we require as textual encoding of prosody is more or less non-existent. Secondly, the
uncertainty surrounding what prosodic form should be makes it hard to label data, train algorithms
and basically even know if we are headed in the right direction.
6.10.1 Labelling schemes and labelling accuracy
Taking the second issue first we see that in every area of prosody there is considerable disagree-
ment as to how to represent prosodic form. The problem isn’t so much that researchers disagree](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-164-320.jpg)
![Section 6.10. Discussion 141
completely; there is wide spread agreement about how to label many utterances; in many cases it
is quite clear that a particular utterance has a strongly prominent syllable, has a clear phrase break
and has a distinctive intonational tune. There are many “black and white” cases where all theories
and all labellers agree. The problem is rather that there is a significant grey area, where we can’t
tell if this really sounds prominent, where we aren’t sure that there is a phrase break between two
words and where we can’t really decide on the intonational tune for an utterance.
There are of course grey areas in other aspects of linguistics. As we saw in Section 4.2, pin-
ning down the exact definition of a word can be tricky, and we will come across similar difficulties
in Section 7.3.2 when we consider the definition of the phoneme. But in general the agreement
among researchers is much much higher with regard to these phenomena. Furthermore, so long as
we are aware of the difficulties, the definitions of word and phoneme which we use in this book
are good enough, meaning that, as engineers, using these models does not seem to result in signif-
icant loss of quality in synthesis. Studies that have examined inter-labeller reliability of prosodic
schemes have shown the true extent of the problem; for instance in the original publication of the
ToBI system, Silverman et al reported that the agreement on phrase breaks was only 69% for four
labellers, and 86% for whether a syllable is prominent (bears a pitch accent in their paper). These
figures might not seem too bad, but are about an order of magnitude worse than the equivalent
results for verbal, phonetic transcription, and up to two or three orders of magnitude worse for the
transcription used for text analysis in Chapter 5. Furthermore, when we consider that the number
of choices to be made in prosodic labelling is often small (say choose one break index value from
5) we see just how difficult labellers find this task. These results for ToBI are not particularly un-
usual, consistently similar figures have been reported for many other schemes [426], [488], [487].
Furthermore it is misleading to blame poor labelling agreement on the labellers not being “expert”
enough - non-experts can readily label many of the types of data we require in TTS, and so in
comparing the figures for other types of labelling we are in fact comparing like with like.
Why does these problems arise?. Essentially we face two related issues;: which labelling
scheme, model or theory to use; and how to assign the labels arising from this to a corpus of
speech. Most researchers involved in both the scientific investigation and engineering implemen-
tation are aware that there are a huge number of differing and often incompatible theories and mod-
els of prosody. The temptation then is therefore to seek a “theory-neutral”, “common-standard” or
“engineering” system which we can use to label the data and so avoid having to nail our colours
to any one particular theoretical mast. We have to emphasise at this point that such an approach is
folly; there is simply no such thing as a theory neutral scheme. If we take the break index scheme
for instance, this explicitly states that there are a fixed number of types of breaks, that no recursion
or limited recursion exists, that phrasing is not explicitly linked to prominence patterns and so on.
Furthermore, there is no sense in which we can just “label what we hear” in the data; all
labelling schemes are based on a model, the fundamentals of that model are implied when we
label speech, and if our scheme is wrong, inaccurate or at fault in some other way, we run a high
risk of enforcing inappropriate labels on our data. If we then use these labels as ground truth, we
are running a severe risk of enforcing our training and run-time algorithms to make choices that](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-165-320.jpg)
![142 Chapter 6. Prosody Prediction from Text
are often meaningless.
6.10.2 Linguistic theories and prosody
Given these difficulties with labelling schemes and models, it is worth asking whether we need a
model of prosodic form at all; perhaps it would be possible to simply define a mapping from the
text or semantics to the acoustics. This viewpoint certainly has some merit, but we should be wary
of dismissing the idea of prosodic form entirely. A brief digression into the theoretical history of
prosody may help explain why.
It is probably fair to say that the field of linguistics has always had a troubled relationship
with prosody. In fact many accounts of language completely ignore prosody (e.g. prosody is
hardly ever mentioned in any work by Chomsky): whether this is because it is deemed to be quite
separate from verbal language or simply too tricky to handle is hard to say. If we however want a
full account of language and communication we have to include prosody.
Older linguistic accounts of prosody tended to treat it as quite a different system, and this to
a degree was unproblematic. But with the rise of more formal theories of language, new theories
of prosody were put forward that used the new techniques of formal language theory, feature
structures and so on. The problem facing these researchers was how closely should a formal
theory of prosody follow the formal theories of verbal language. Unfortunately (in my opinion)
too much emphasis was placed on an insistence that prosodic phenomena could be explained with
the same mechanisms for verbal language, and hence we saw grammars of intonation, intonational
phonology and schemes for pitch accent description which were meant to be the direct equivalent
of normal phonology. To be fair, external pressures pushed research in this direction; there was
a strong desire to move from the very impressionistic and informal models of the past towards a
more rigorous model and in addition there was the idea that if prosody could be explained with
the same mechanisms as verbal language this would somehow be more parsimonious and would
also ensure prosody was taken seriously as a linguistic area of study (instead of just being ignored
as mentioned above).
Ladd [269] points out that a major justification for the development of a phonology for
prosody was to counter attempts in linguistics to correlate say F0 values with syntax patterns.
While correlations may exist, building such a model ignores the fact that different speakers have
different “parameters” in their F0 encoding model, for example that high pitched voices will pro-
duce different patterns to low pitched ones. A better model is to posit a system of abstract prosodic
form, where all speakers of an language/accent would share the same representations for similar
semantic and functional purposes. The actual acoustic encodings of these might vary with speaker
physiology, but the prosodic form would be the same for all speakers. In formal terms, all we
are saying is that there are a number of levels, or interacting systems of prosody. In summary the
phonological school of prosody is nearly certainly right in advocating abstract description systems
for prosodic form, but is perhaps less right in advocating that these should be formulated with
the same mechanisms as verbal language. Prosody has an abstract structure, but it operates in a](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-166-320.jpg)
![144 Chapter 6. Prosody Prediction from Text
from the text) explain the presence of phrasing to a certain degree. The other features explain the
rest of the behaviour, but as we don’t have access to these, they effectively appear as noise in our
data and variance in our models.
The consequence of this, taken in conjunction wit the issue of very low level of transcribed
agreement, may well mean that many prediction algorithms are as accurate as theoretically possi-
ble; no amount of extra refinement in the techniques or extraction of alternative features will make
much difference. The reason why the algorithms can not improve is due to noise in the data, from
either or both or these sources.
6.10.4 Prosody in real dialogues
In Sections 6.8 and 6.7 we have seen how to predict prominence and phrasing patterns from text.
The techniques described though in general only operate on single sentences, which are usually
declarative in nature. Much of interest in prosody takes part at a different level, where many of
the main phenomena occur in conversational, discourse, emotive or spontaneous communication.
Hence most TTS systems only model a small range of the possible prosodic phenomena. We can
however imagine systems which go beyond current capabilities. If speech synthesis were used in
more conversational applications, the discourse effects of prosody it would become important to
model the discourse effects of prosody. In many current dialogue systems, the interaction feels
very artificial because the system is only capable of simple questions and declarative sentences.
The study of prosody in discourse in a rich topic of research in linguistics, but little has be done
in terms of well adopted practices in TTS systems. Hirschberg [208] gives a review of many
discourse and functional aspects of intonation and describes much of the work that has been done
in discourse prosody in TTS.
We need to reiterate some points about the augmentative function of prosody and how this
affects studies into prosodic phenomena. While it is widely accepted that prosody does undergo
more variation, or undergoes more speaker choice than other parts of language, it is our belief that
the degree to which this occurs is vastly underestimated in most studies. If we compare prosodic
form to verbal form, we see that while speakers can indeed vary the articulation of a word such as
HAT, they are very constrained in doing, and too much deviation from the canonical pronunciation
runs the risk of the word being decoded incorrectly.
A key reason for the difference in degree of variation between verbal and prosodic language
is to do with their functional use in communication. To a large extent, we can study the verbal part
of language without reference as to why a particular utterance is used in a discourse but the same
situation simply does not hold for prosody. The way in which prosody is used is primarily driven
by the discourse, and taking sentences out of their discourse and studying them without reference
to this, leads to problematic analyses. In verbal phonetics a common practice is to carefully elicit
sentences spoken under laboratory conditions. These conditions are artificial and can lead to dis-
crepancies between the phonetic/acoustic patterns observed between “canonical” speech and real,
spontaneous speech. In general though, a considerable amount can be learned from controlled](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-168-320.jpg)
![Section 6.10. Discussion 145
experiments in this way. The situation is radically different in prosody however, and eliciting indi-
vidual sentences in laboratory conditions is extremely risky in that these bear virtually no relation
to how prosody is used in real, spontaneous speech. Many studies unfortunately impose a strong
discourse bias, so that when we examine utterances elicited for the purposes of studying say the
given/new distinction, what we are doing is asking speakers to produce sentences in a very artifi-
cial discourse context. The effect of this is to grossly exaggerate the way the prosody is generated,
such that effects such as given/new or contrastive prominence can seem quite clear, distinct and or-
derly. In real, spontaneous speech by contrast, the discourse situation and the speaker and listeners
mental models of each other dictate how prosody is used. In these situations, the discourse context
and information content is often so obvious and redundant that prosody is simply not required for
further disambiguation. This helps explain while earlier laboratory based studies of prosody claim
clear patterns and effects, but when studies of real data are performed [328], [24], these effects are
often far less clear.
6.10.5 Conclusion
The take home message from this discussion is that the theoretical difficulties with prosodic mod-
els and the under specification of prosody in text mean that this is an inherently difficult problem
and should be approached with a degree of caution. While it is possible to continually drive down
the error rate with regard to semiotic classification, homograph disambiguation and so on, it is
unrealistic to assume the same can be done with prosody.
6.10.6 Summary
Prosodic form
• It is usual to consider three basic components of prosodic form: phrasing, prominence and
intonation.
• Phrasing is used to group words together in the utterance, and is influenced by top down
factors such as syntax, and bottom up factors which place phrase breaks at roughly equal
intervals.
• Prominence is used to give extra strength to certain words. It is partly determined by syntax,
where in particular it governs stress patterns in compound noun phrases and function word
sequences.
• Intonation is the use of pitch to convey prosodic information.
• Very little agreement exists on how to represent prosodic form, which makes the design of
practical algorithms more difficult than in other areas of TTS.
Prosodic function
• Affective prosody conveys emotion and attitude in communication.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-169-320.jpg)
![7
PHONETICS AND
PHONOLOGY
This chapter gives an outline of the related fields of phonetics and phonology. A good knowledge
of these subjects is essential in speech synthesis as they help bridge the gap between the discrete,
linguistic, word-based message and the continuous speech signal. More traditional synthesis tech-
niques relied heavily on phonetic and phonological knowledge, and often implemented theories
and modules directly from these fields. Even in the more modern heavily data driven synthesis
systems, we still find that phonetics and phonology have a vital role to play in determining how
best to implement representations and algorithms.
7.1 ARTICULATORY PHONETICS AND SPEECH PRODUCTION
The topic of speech production examines the processes by which humans convert linguistic mes-
sages into speech. The converse process, whereby humans determine the message from the speech
is called speech perception. Together these form the backbone of the field know as phonetics. In
general,
Regarding speech production, we have what we can describe as a complete but approximate
model of this process. That is, in general we know how people use their articulators to produce
the various sounds of speech. We emphasise however that our knowledge is very approximate; no
model as yet can predict with any degree of accuracy how a speech waveform from a particular
speaker would look like given some pronunciation input. The reason for this lack of precision
is that the specifics of the production process are in fact incredibly complex; while we can make
solid generalisations about how a [s] sound is pronounced for instance, it is another matter en-
tirely to explain how this varies from speaker to speaker, how [s] is affected by speech rate, how
surrounding sounds influence it and how this is affected by prosodic interaction. The fact that this
knowledge is incomplete should of course not be a surprise to readers of this book; if we fully
understood this process we could simply implement it in a computer and thereby solve the speech
synthesis problem. In fact, as we shall see, in the field of text-to-speech, our lack of knowledge has
led us to abandon (for the time being at least) the path of trying to mimic the human production
process directly. However, what knowledge we have can be put to good use, and hence a solid
147](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-171-320.jpg)
![148 Chapter 7. Phonetics and Phonology
grounding in what speech production and perception knowledge we do have is an essential part of
constructing a high quality synthesiser.
Probably the main difficulty in discovering the mechanisms of human speech production
is that of measurement and data collection. Of course this is a problem common to all science;
before telescopes only the simplest models of astronomy were possible; afterwards detailed and
accurate observations could be made which allowed the rapid development of theories. We see the
same in the scientific study of phonetics; in the early days there were no instruments available for
examining speech (phonetics even predates the invention of sound recording). With the invention
first of sound recording, then spectrographic analysis, x-ray films and other techniques, speech
scientists were able to examine the processes in more detail. Such instruments and more elemen-
tary techniques (such as looking directly into a speaker’s mouth or simply trying to determine the
position of one’s own tongue) have led to the level of understanding we have today and which we
shall briefly describe next.
Progress in speech perception has been significantly slower. The lack of progress can be
explained simply by the difficulty in extracting information about the process. While some tech-
niques exist for examining the cochlea and analysing brain patterns when listening, in general it
is very difficult to collect concrete data on the problem, and as a result scientists have little to go
on when developing theories. What ever progress has been made is typically conducted within
within the framework of experimental psychology (e.g. [58]]. This significant imbalance between
our knowledge of production and perception helps explain why phonetics is in general approached
from a speech production point of view; that is simply the area of the speech process that we best
understand.
7.1.1 The vocal organs
We generate speech by the coordinated use of various anatomical articulators known col-
lectively as the vocal organs. Figure 7.1 shows a mid-sagittal section of the head showing the
vocal organs. If we consider just the tongue in isolation for a moment, it should be clear that it
can move in all three dimensions, and can create a complex variety of movements and trajecto-
ries. The movement and position of the tongue directly affects the speech produced and even a
movement of a few millimeters can dramatically alter the speech sound produced. While the other
organs are somewhat simpler, it should be clear that a substantial number of vocal organ configu-
rations are possible, and each gives rise to a different sound. This alone gives some insight into
the complexities of speech.
7.1.2 Sound sources
Nearly all sounds in English are created by by air moving from the lungs through the vocal organs
to the lips and then outwards. This flow is called an egressive pulmonic air stream; egressive,
because the air is flowing outwards, and pulmonic because the source is the lungs. During the](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-172-320.jpg)
![Section 7.1. Articulatory phonetics and speech production 151
quencies known as harmonics. These are found at multiples of the fundamental, so if we have
a glottal sound with a fundamental of 100Hz, we will find harmonics at 200Hz, 300Hz, 400Hz
and so on. In general the harmonics are weaker than the fundamental, but nonetheless are heard
and can be thought of as giving the sound its basic timbre. A musical analogy might help here.
One can create a sound of the same fundamental frequency by either bowing or plucking a violin
string. In both cases the fundamental frequency is the same, however the different in action cre-
ates a different pattern of harmonics and this is why the timbre of bowing and plucking sounds
different.
If the glottis is opened slightly further, periodic vibration will cease. Instead a non-periodic
turbulent air flow will be created, which generates a different type of sound, termed noise. Here the
term “noise” is used in the technical sense (as opposed to a general definition which is a synonym
for “sound”) to mean the random sound created, somewhat like an untuned analogue radio. This
is the natural way to generate the [h] sound in HEDGE, but is also the way in which whispering
works. If the vocal folds are brought together completely, the air stream stops and no sound is
produced. By carefully controlling the timing of this opening and closing, it is possible to produce
a glottal stop, which is the normal way to realise the [t] in words such as BUTTER in some accents
of English (e.g. Glaswegian or Cockney).
Sound can be created from sources other than the glottis. If the glottis is open, it will not
generate any sound, but will allow the air stream to flow through. It is then possible to use com-
binations of the tongue, lips and teeth to form a constriction, and thereby generate a non-glottal
sound source. The [s] sound in SIT is generated this way - by speaking this, and extending the
length of time the [s] is spoken for, you can feel that the glottis is not vibrating and that the tongue
is very close to the roof of the mouth. Noisy sounds can also be created by holding the teeth close
to the lips (as in the [f] sound in FALL) or by placing the tongue near the teeth as in the [th] sound
in THANK. Because there is no glottal periodic vibration in these sounds, they are called unvoiced.
It is possible to have sounds which are both voiced and noisy, such as the [z] sound in ZOO. In
these cases, the glottis operates as with vowels, but a further narrow constriction is created in the
mouth to generate a secondary, noisy, sound source.
Sounds such as [s], [f] and all vowels can be spoken as a continuous sound, and are therefore
called continuants. By contrast stops are sounds which have a relatively short duration which
can only be spoken as “events” and can not be spoke continuously. The [p] sound in PEN or [t]
sound in TIN are stops. Stops are produced by creating a closure somewhere in the vocal tract so
that all air flow is blocked. This causes a build up of pressure, followed by a release where the
air suddenly escapes. Because the sound is produced in this way, the sound must have a finite,
relatively short duration and hence these sounds cannot be produced continuously.
7.1.3 Sound output
For most speech sounds, the acoustic signal leaves via the mouth - such sounds are termed oral.
For a few however, the sound leaves via the nose; such sounds are called nasal. In English the first](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-175-320.jpg)
![Section 7.1. Articulatory phonetics and speech production 153
sound, the velum blocks the entrance to the nasal cavity, and so this plays no part in determining
the sound. In nasal sounds however, while the oral cavity is blocked, it is not blocked at the
entrance, rather it is blocked at some other point (at the lips in an [m] sound and near the alveolar
ridge in an [n] sound). During the production of a nasal, air enters the oral cavity, is reflected and
eventual escapes from the nose. The shape of the oral cavity plays an important part in determining
the sound of the nasal, even though no air escapes from it directly. Figure 7.3 shows vocal tract
configurations for these output types.
7.1.4 The vocal tract filter
The diversity of sound from the source and output are further enriched by the operation of the
vocal tract. The vocal tract is collective term given to the pharynx, the oral cavity and the nasal
cavity. These articulators can be used to modify the basic sound source and in doing so create a
wider variety of sounds than would be possible by the source alone. Recall that all voiced sounds
from the glottis comprise a fundamental frequency and its harmonics. The vocal tract functions
by modifying these harmonics which has the effect of changing the timbre of the sound. That is, it
does not alter the fundamental frequency, or even the frequency of the harmonics, but it does alter
the relative strengths of the harmonics.
In general it is the oral cavity which is responsible for the variation in sound. The pharynx
and nasal cavity are relatively fixed, but the tongue, lips and jaw can all be used to change the
shape of the oral cavity and hence modify the sound. The vocal tract can modify sounds from
other sources as well by the same principle.
This model, whereby we see speech as being generated by a basic sound source, and then
further modified by the vocal tract is known as the source/filter model of speech. The separation
into source and filter not only adequately represents the mechanics of production, it also repre-
sents a reasonable model of perception in that it is known that listeners separate their perception
of the source in terms of its fundamental frequency from the modified pattern of its harmonics.
Furthermore, we know that the main acoustic dimension of prosody is fundamental frequency,
whereas the main dimensions of verbal distinction are made from a combination of the type of
sound source (but not its frequency), and the modification by the vocal tract. The mathematics of
both the source and the filter will be fully described in Chapter 10.
7.1.5 Vowels
All vowel sounds are voiced and hence have the same sound source (the glottis). What distin-
guishes one vowel sound from another is the shape of the oral cavity and to a lesser extent the
shape of the lips and duration with which the sound is spoken. The oral cavity and lips operate
to create different shaped cavities through which the source sound must pass. By moving the jaw,
lips and tongue, different filter effects can be created, which serve to modify the harmonics of the
glottal sound source and produce the wide variety of vowel sounds. It is important to note that](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-177-320.jpg)
![Section 7.1. Articulatory phonetics and speech production 155
the tongue with in relation to the front of the mouth defines another dimension, which we can call
front/back. Consider the difference in tongue position in the vowel in HEAT and HOOT - while
this is slightly more difficult to feel or measure than in the case of height, it should at least be clear
that the tongue is in a different position and that the jaw remains relatively fixed. Figure 7.4 shows
the articulator positions for HEAT, HOOT and HOT. A third dimension is called lip rounding or
simply rounding. This effect is caused when the lips are protruded from their normal position.
Some vowels (e.g. HOOT) have noticeably rounded vowels, while others are noticeably unrounded
(e.g. HEAT). A final basic dimension of vowel classification is length. This feature is very accent
specific, but in many accents the vowel in HEAT is considerably longer than the one in HIT.
In languages such as English, there is a particular vowel sound known as schwa. This is
often called the neutral vowel in that its position is neither front nor back, high nor low, rounded
nor unrounded. It is first vowel in ABOUT, SEVILLE, COLLECT, and the second vowel in SARAH,
CURRENCY, and DATA. Most other sounds are simply called by how they sound; schwa is given
its special name because it is difficult to pronounce in isolation: if one attempts this, the sound that
comes out is normally too strong sounding. Schwa is spoken with a tongue position which is close
to that when the tongue is in its rest position, and as such the tongue doesn’t create any particular
cavities or shapes in the vocal tract; rather the vocal tract is in the shape of a fairly uniform tube.
Some vowels are characterised by a movement from one vocal tract position to another. In
HEIGHT for instance, the vowel starts with a low front mouth shape and moves to a high position.
Such vowels are termed diphthongs. All other vowels are called monophthongs. As with length,
whether or not a vowel can properly be considered a diphthong is heavily accent dependent. In
particular the vowels in words such as HATE and SHOW vary considerably in their vowel quality
from accent to accent.
7.1.6 Consonants
Consonant sounds are more heterogeneous than vowels in that they are produced by a wider variety
of sound sources. While consonants share some inherent characteristics, it is not inappropriate to
define consonants as the set of speech sounds that are not vowels.
Consonants can be classified in terms of voicing, manner and place of articulation. As
we have mentioned, voicing describes whether the sound source is produced by the vibration of
the glottis or by another means. Consonants such as [f] and [s] are unvoiced while [v] and [z] are
voiced. This can be seen by speaking a [v] continuously and turning it into an [f] - it should be
possible to do this without altering the shape of the vocal tract and simply changing the vibration
mode of the glottis.
The first important category of sounds are known as oral stops. In English, stops occur in
voiced/unvoiced pairs, so that [b] and [p] use the same manner and place and of articulation, and
differ only as to whether the glottis is vibrating or not. The three nasals (the last sounds in THEM,
THIN and THING) are often also described as nasal stops. While it is possible to continuously utter
these three phones, they usually follow the pattern of the stop in normal speech, with a closure and](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-179-320.jpg)
![156 Chapter 7. Phonetics and Phonology
(a) Alveolar constriction (b) Labio-dental constric-
tion
Figure 7.5 Place of articulation for alveolar and labio-dental fricatives
release. During the production of nasals, the mouth is closed, so the sound only escapes from the
nose. The position of the tongue in the mouth is important however - the nasal tract is fixed and
so does not distinguish the three nasals. The production of nasals can be thought of as the glottis
sending a periodic sound into the mouth which is modified (filtered) by the oral cavity and then
passed out through the nose, which performs further modification.
Fricatives are generated by creating a narrow constriction in the vocal tract, by placing (say)
the tongue near to the roof of the oral cavity, thus creating a turbulent, noisy sound. Fricatives can
be voiced (e.g. [v]) or unvoiced (e.g. [f]). In voiced fricatives, there are two sound sources, the
glottis, which produces a periodic sound source, and the constriction which adds turbulence to
this. To demonstrate that this is indeed the case, try saying [v] and [f] with different pitches. This
is possible with [v] but not with [f]. The final class of true consonants is the affricates, such as
the first and last sound in CHURCH. These are spoken as a stop followed by further frication.
What then distinguishes the unvoiced fricatives such as [f] and [s]? The answer lies in where
the constriction occurs. For [f] the teeth combines with the lips to form the constriction - this is
termed labiodental. In [s] the sound in generated by the tongue coming into close proximity with
the roof of the mouth near the alveolar ridge and hence this sound is called alveolar. There are
many other possible positions, and each will produce a slightly different sound. While in reality
a continuum of constriction positions can occur, for convenience it is useful to identify and name
a few key positions for constrictions. These include bilabial in which both lips combine to form
the restriction, and dental in which the tongue combines with the teeth. The alveolar position is
where the tongue approaches the roof of the mouth in a forward position. The tongue can create a
constriction at positions further back, which are termed post-alveolar, palatal, velar and uvular.
The place of articulation can be described on a scale starting at the lips and moving to the near the](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-180-320.jpg)
![Section 7.1. Articulatory phonetics and speech production 157
back of the vocal tract. The important point is not so much which articulators (teeth, lips, tongue)
are involved, but rather where the constriction occurs, as this governs the shape of the oral cavity.
One important point worth noting about fricative sounds is that the constriction as a sound
source does not define the difference between an [f] sound and a [s] sound. In fact the sound
source produced by all fricatives is the more or less the same. Rather, it is the fact that the con-
striction gives rise to a vocal tract configuration, and it is this acting as a filter which modifies the
sound source and thereby produces the difference in sound. As we shall see in Chapter 11, it is not
the case that sound always moves forward in sound production. In dental fricatives, while some
portion of the sound source is immediately propagated from the mouth, the remainder is propa-
gated back into the vocal tract, where it is modified by the filter, after which it is then reflected and
propagated forward. It is by this mechanism that the vocal tract generates different sounds even if
the sound source is near the mouth opening. The term obstruent is used as a collective term for
fricatives, affricates and oral stops to indicate their noisy character.
The approximants are a class of sounds which are interesting in that they share many of the
properties of both vowels and consonants. All approximants are voiced and all are produced in
roughly the same manner as vowels, that is by varying tongue, lips and jaw position. [j] and [w]
are know as glides and have similar sounds and means of articulation to the vowels in HEAT and
HOOT. In some analyses, diphthongs are described not as compound vowels but as a combination
of a vowel and a glide, as the glide comes after the vowel it is termed an off-glide. It is also
possible to have on-glides, so CURE is described as a [y] sound followed by a vowel. Glides also
function as proper consonants, for example as the first consonant in YELL , WIN and so on. If
these functioned as vowels rather than consonants, the perception would be that these words have
two syllables rather than one. The approximants [r] and [l] are known as liquids and act more like
normal consonants, but still have many unusual properties (see Section 8.2.4).
7.1.7 Examining speech production
In traditional phonetics, the tools available to the phonetician were extremely limited and much
of what we have so far described was developed with virtually no equipment. Progress was made
by simply “feeling” the position of the articulators when speaking, or by looking down someone’s
throat as they spoke (which is not very helpful for examining nasal stops). The ability of these
phoneticians to construct a model of speech production with such limited means is quite astonish-
ing, and by and large, they got the basics of the model correct (or correct in terms of our current
understanding). One aspect which was not however realised at this time was the complexities of
the syntagmatic dimension, meaning the influence that sounds had with one another when spoken
as a sequence. Furthermore the role that “top-down” knowledge was playing in the phoneticians’
examination was also largely unrecognised.
As recording and examination apparatus improved it became clear that what had seemed
like clearly identifiable distinct sounds of speech where in fact quite hard to classify based on the
articulatory evidence alone. We will address this question fully in Section 7.3.2, but for now let](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-181-320.jpg)
![158 Chapter 7. Phonetics and Phonology
us simply note that it was found that what were thought of as single sounds (e.g. [p]) in fact had a
huge range of possible articulatory realisations.
Over the years a number of analysis tools have been developed to examine the production
process. The first was the use of x-ray still photography, and then x-ray motion pictures. These
techniques, especially the x-rays movies clearly showed the various articulators in motion. Unfor-
tunately, the availability of x-ray movies is extremely limited due to the now understood harmful
effects of x-ray exposure and hence this technique can not be used much. More recently a number
of techniques have been developed. These include:
electropalatography This takes the form of a plate fitted to the roof of the mouth which measures
contact between the tongue and palate.
electoglottograph or laryngograph This is a device which fits around the neck and measures the
impedance across the glottis. From this, a signal measuring glottal activity can be found.
airflow measurement A mask and measuring apparatus can be used for measuring air flow from
the mouth and separately from the nose. This is particularly useful in that we can measure
the output from the nose and mouth separately, and we can do this in terms of air flow signals
rather that the pressure signals which microphones record.
magnetic resonance imaging (MRI) This technique, originally developed for medical use, can
generate 3-D images of the head and so measure the position of the articulators.
electromagnetic articulograhpy (EMA) This is a device which works by attaching small elec-
tronic coils to the tongue and other articulators. During speech, the movement of the coils
causes fluctuations in a surrounding magnetic field, which allows articulator movement to
be tracked.
A description of the operation of many of these devices can be found in Ladefoged [272].
Examples of studies using EMA and MRI can be found in [498], [380], [332], [13], [331].
7.2 ACOUSTICS PHONETICS AND SPEECH PERCEPTION
We term output of the speech production process the speech signal. This acoustic signal travels
through air to the listener, who can then decode this to uncover the message that the speaker
uttered. Speech signals can be recorded by means of a microphone and other equipment, such that
with a little care, we can store a perfect representation of what the speaker said and the listener
will hear. The field of acoustic phonetics studies speech by analysis of acoustic signals. The idea
here is that since this is what the speaker receives when listening, all the required information for
perception should be present, and so by suitable analysis we can discern the processes of human
speech perception itself. That said, it must be emphasised that to date we have only a very limited
understanding of speech perception. We know some of the initial processes (which for instance
separate pitch from other aspects of the signal), but virtually nothing about how people actually
distinguish one sound from another. Hence speech analysis should be thought of as an independent](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-182-320.jpg)
![162 Chapter 7. Phonetics and Phonology
i I ae
u
First Formant
400 800 1200
Second
Formant
500
1000
2000
3000
4000
Figure 7.8 Chart of measured mean positions of vowels plotted in terms of first and second formant
positions. The chart shows that to a large extent, the two formant positions separate most vowels. This
has led to the general assumption that F1 and F2 positions are used to discriminate vowels. This chart
is shown in the form standard in phonetics; note that the axes have been specifically set up to make
this plot easier to interpret.
its own oral cavity shape however, and the resonances in this are the main distinguishing feature
between [m] and [n]. In stops, it is often possible to see the distinct phases of closure and the
subsequent burst. While subtle, it is possible to tell one stop from another from the resonant
patterns in the burst and in the immediately neighbouring vowel. Approximants look like weak
vowels, which is what we would expect.
7.3 THE COMMUNICATIVE USE OF SPEECH
We have shown in the previous sections how the vocal organs can organise into a rich variety
of configurations which in turn can produce a rich variety of speech sounds. We have also seen
that the differences in vocal organ configurations can produce patterns which are discernible in
the acoustic signal and representations derived from it. We now turn to the question of how this
capability can be used to communicate.
7.3.1 Communicating discrete information with a continuous channel
Recall that natural language is a discrete symbolic semiotic system. In this system we combine
words in different ways to produce a near limitless number of sentences, each with a separate](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-186-320.jpg)
![Section 7.3. The communicative use of speech 163
meaning. One might imagine a system of speech communication whereby each word was rep-
resented by a single unique speech sound, so that DOG would have a high central vowel, CAT a
alveolar fricative and so on. The problem with such a system would be one of distinctiveness: if
we had 10,000 separate words we would need 10,000 different speech sounds. It is unlikely that
such a system would work simply because it would not be possible for listeners to accurately dis-
tinguish all the separate sounds, nor would it be likely that speakers would have enough precision
to utter each sound exactly. To get round this problem, we can make use of a much smaller set
of sounds, but use multiple sounds in sequence for each word. As the set of fundamental sounds
is small it should be easier to distinguish one from the other; and so long as we can “segment” a
word into a sequence of separate sounds, we should be able to create a large number of distinct
words with a relatively short sequence for each.
This is of course how speech communication actually works; we have a small number of
units which we can distinguish from one another and we use these in different sequences to create
the form of a large number of words. We call these units phonemes. The number of phonemes
varies from language to language, but all languages use a roughly similar set of units ranging in
size from about 15 to about 50. There is clearly a tradeoff between the number of phonemes we
use and the length of words; languages with small numbers of phonemes will need on average to
use longer sequences of these phonemes to produce distinct forms.
The question now is, out of all the possible vocal organ configurations, which do we use
for this process? First, let us say that the evidence shows that we do no simply pick a arbitrary
set of 40 particular sounds from the entire set and make use of these alone. We know from our
principle of semiotic contractiveness that it makes sense to pick a set of sounds that are readily
distinguishable from one another; hence we don’t find languages which only have fricatives or only
have front vowels. Secondly we want sounds, or more precisely a system of sound sequencing,
that produces sequences that are relatively easy to speak and relatively easy to “segment” when
listening. (For example, we would not expect a language to have sequences of velar stops followed
by bilabial nasals followed by oral stops, as these sequences are difficult to say). The choice of
phonemes in a language is therefore a tradeoff between picking a large number of sounds (short
words) and a short number (easier to identify each), sounds which are distinctive from one another,
sounds which are easy to speak in sequence and sounds which are easy to segment when heard in
sequence.
Even given a set of sounds that obey these criteria, it is not the case that each phoneme has
a unique articulatory position or acoustic pattern. To take an example, when speaking the word
SHOE, most speakers utter the first sound (represented by the symbol [M]) sound with a mouth
position in which the lips are pursed or rounded. Compare this with the lip position in SHINE
and we see that the lips are in another position, which is more spread and in a position somewhat
similar to a smile. However, despite the fact that the acoustic and articulatory differences between
the initial sounds in these two words is quite big, they are perceived cognitively as a single sound.
The reason why this variation occurs is quite obvious; SHOE has a rounded vowel, and so all the
speaker is doing is anticipating speaking the rounded vowel by already having the mouth in this](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-187-320.jpg)
![164 Chapter 7. Phonetics and Phonology
position before hand. This is therefore a consequence of wanting to produce a sequence which
requires less effort to speak. This does not produce confusion because while the rounded [M] is
different from the normal [M], the rounded [M] is not confusable with any other phoneme, and
hence the listener should not have difficulty in identifying the correct phoneme. Note that it is
quite possible that the rounded and unrounded versions of [M] could be different phonemes in
another language; it is just a property of English that they are grouped together.
Hence the language system makes use of a set of contrasting sounds as its basic units, but
within each unit a considerable degree of variability is allowed. These properties are entirely
to be expected from the semiotic principles of contrast and sequencing, and the communicative
principles of ease of production and ease of perception.
7.3.2 Phonemes, phones and allophones
We will now define four terms commonly used in phonetics. The term phoneme is used for the
members of the relatively small (between 15-50) set of units which can be combined to produce
distinct word forms. Phonemes can have a range of articulatory configurations, but in general the
range for a single phoneme is relatively limited and occupies a contiguous region in the articulatory
or acoustic space. The term phone is used to describe a single speech sound, spoken with a single
articulatory configuration. The term allophone is used to link the phoneme and phone: different
ways of realising a single phoneme are called allophones. In our example, we identified a single
phoneme at the beginning of the two words SHOE and SHINE. As each of these has a different
articulatory position, each is a separate phone, but as they both “belong” to the same phoneme
they are termed allophones of this phoneme. Finally, we will use the term segment as a general
term to refer to phonemes, phones and allophones.
Phones can be thought of as units of basic speech production and perception whereas phonemes
are specific to a particular language. The way phonemes and phones are connected varies from
language to language and is not necessarily a simple relationship. The easiest way to demonstrate
this for speakers of English involves demonstrating a case where two sounds are phonemes in
English but allophones in a different language. The classic case of this is the [r] vs [l] distinction
in South East Asian languages; all English speakers readily tell these apart and usually have no
idea they are even related phonetically. In Japanese, there is no such distinction, and they are allo-
phones of a single phoneme (by convention usually represented [r]), such that Japanese speakers
speaking English often seem to get [l] and [r] “wrong”. To many speakers of English, this seems
incredible that such a “basic” distinction can be missed, and this unfortunately becomes the sub-
ject of many a lame comedy parody. In fact mastering the allophonic and phonemic distinctions is
one of the most difficult tasks when learning a new language as our whole perceptual mechanism
seems geared to treating all allophones in our native language as the same.
The above in fact shows the most accepted way of determining the phoneme inventory for
a language. This is based on the principle of the minimal pair, and in effect says that if one can
find two distinct words (such as RUSH and LUSH) that differ only in one sound, then those sounds](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-188-320.jpg)
![166 Chapter 7. Phonetics and Phonology
æ
V
*℄
•
•
•
•
T
•
'• u
•
5•
o
•
•
=
•
•
•
•
Œ
•
a•
œ
•
•
ø
•
e•
y
•
i•
Figure 7.9 IPA Vowel chart. The position of each vowel indicates the position of the tongue used to
produce that vowel. Front vowels are to the left, back vowels to the right. Where a unrounded/rounded
pair occurs, the unrounded version is on the left.
Bilabial Labiodental Dental Alveolar Post-alveolar Retroflex Palatal Velar Uvular Pharyngeal glottal
Plosive p b t d P c , k g q b
Nasal m 4 n 9 7 8 :
Trill B r R
Tap/Flap D F
Fricative C f v S s z M ` L _ CX jX X [ K # h $
Lateral fricative / 1
Approximant Y G H j 6
Lateral approximant l 0 L
Figure 7.10 IPA Consonant chart. Where there is a unvoiced/voiced contrast, the unvoiced symbol
is shown on the left.
the union of all known phonemic boundaries.
Yet another way of thinking of phones and phonemes is in terms of cognitive and physical
spaces. Cognitively, we represent all the instantiations of a particular phoneme as being the same;
in our SHOE/SHINE example, the first sound is clearly the same in the sense that most people are
completely unaware that there is even any difference until the differences are explicitly pointed
out. By the same measure, although the physical difference between say [p] and [t] may be very
slight, people readily agree that they are different sounds. It is important to realise that we don’t
group allophones together because we can’t tell the difference, but rather because they group via
the cognitive map of the phonetic space. To see this, we can again repeat the experiment where we
use an unrounded Mat the start of SHOE and a rounded Mat the start of SHINE. Most people can tell
the difference here, and while maybe not being able to attribute this difference to the articulation
itself, can none the less hear the difference. So it is not the case that somehow the ear or low level
perceptual hearing filters these changes, as we can hear them if we try.
It is now timely to comment on the practice of phonemic notation. Unlike the IPA there is](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-190-320.jpg)
![168 Chapter 7. Phonetics and Phonology
different notation system for phonemes than phones helps emphasise the difference between the
two. This is helpful for our purposes as it makes it clear that phonemes are a set of contrastive
linguistic units, and not entities which are expected to have well defined or invariant physical
manifestations. Notationally, it is conventional to use square brackets [ ] when describing phones
and segments, and slash “brackets” / / when describing phonemes. Hence for the word HELLO we
have a phone sequence [h l oV] and a phoneme sequence /h ax l ow/.
7.3.3 Allophonic variation and phonetic context
As has already been mentioned, phonemes do not represent a single point in phonetic space; if
they did then each phoneme would have a unique acoustic pattern and the process of synthesis
and recognition would largely be trivial. We have already touched on the subject of allophonic
variation, which we define as the realisation of a phoneme as a range of possible phones. We have
already discussed anticipatory rounding for M, but other anticipatory affects are quite common. For
example in the [k] of {COT, /k aa t/} and {KIT, /k ih k/ } we see that in COT the tongue position
for [k] is in a back position, whereas in KIT it is much further forward. This can be explained by
the position that the tongue will be in during the following vowel.
The allophonic possibilities for a phoneme can depend on its position in relation to other
phonemes. For example, the [n] in {NIT, /n ih t/} needs to sound different from the [m] in {MIT,
/m ih t/}, but the [n] in {HAND, /h ae n d/} does not need to contrast with [m] as there is no word
/h ae m d/ that it could be confused with. It does need to be distinct from say [r], [sh] and [v]
otherwise it could be confused with HARD, HASHED and HALVED. This shows that the [n] needs
to still be a nasal in that it still needs to be identified with respect to some other consonants; it
doesn’t need to be identified with respect to other nasals and so its distinctiveness with respect
to [m] and [ng] need not be so precise as in other positions. As there are constraints on what
phoneme sequences occur; only distinctions between possible phonemes at a given point need to
be maintained. The variability mainly arises from the fact that it is physically easier for a speaker
to produce a given phoneme one way in one position and a slightly different way in a different
position.
It turns out that a great degree of verbal variability can be explained once we take phonetic
context into account. By this we mean that if we consider the identity of the phonemes surround-
ing the one we are interested in; the identity of those surrounding phonemes helps explain why
we see the variation in the speech. For example, we find in general that nasals before stops often
share the place of articulation with the stop; so we naturally get words such as {HUNT, / h ah n t/}
and {HUMP, / h ah m p/}, where the [n] and [t] are both alveolar, and the [m] and [p] both bilabial;
but we don’t get words /h ah m d/ and / h ah n p/, where the nasal has one place of articulation and
the following stop has another. This phenomenon is called assimilation, reflecting the idea that
the place of articulation of one phone assimilates to the other.
Another common effect is called colouring, where the identity of the phoneme in question
isn’t affected or limited, but where the phoneme takes on some characteristics of the a neighbour-](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-192-320.jpg)
![Section 7.3. The communicative use of speech 169
ing phoneme. A common type of colouring is nasalisation, in which a vowel followed by a nasal
itself becomes somewhat nasalised (that is the velum is not completely closed during the produc-
tion of the vowel). Colouring in vowels frequently happens when they are followed by nasals and
also to some extent [l] and [r] (in rhotic accents). In English, this is again usually anticipatory.
Other effects include elision or deletion in which a sound is “deleted” and epenthesis or insertion
is where an extra sound is inserted. For these phenomena, consider the words MINCE and MINTS;
they sound indistinguishable and clearly a process of elision or epenthesis has occurred. Interest-
ingly, it is not clear which has happened; either we could have elision, in which both would have
the phoneme sequence /m ih n s/, or epenthesis in which the sequence for both would be /m ih n t
s/.
Basic phonetic context on its own is not enough to explain all the variation. Consider the
behaviour of [t] in the words TOP, LAST and STOP. In the first two, the [t] is normally aspirated,
that is, there is a clear burst of energy after the release, which is heard of as a short section of
noise. In STOP there is no burst; the vowel starts immediately after the release. This cannot be
explained simply by context alone; after all the [t] is aspirated when preceded by a [s] is LAST and
is also aspirated when followed by a vowel in TOP. These and other examples require reference to
other structural factors such as the position in the syllable and word.
7.3.4 Coarticulation, targets and transients
If contextual allophonic variation was all we had to contend with, the phonemic to acoustic process
would still be quite simple; all we would need is a model whereby each phoneme had a range
of possible realisations; the process of conversion from one domain to the other would still be
relatively trivial.
In fact, even if we take into account any allophonic variation that may occur, we find that the
articulatory and acoustic patterns for a single phoneme are quite complex. This is because it is not
the case that a single allophone is spoken with a single articulatory position; rather the articulators
are constantly in motion. This is readily discernible from any of the acoustic or articulatory rep-
resentations already encountered. The articulators require time to move from one position to the
next; and while this can be done quite quickly, it is usually the case that by the time the articulators
have reached the position required to speak one phoneme the speaker is already moving them to a
new position to speak the next phoneme. This means that the articulators are constantly moving,
and that the pattern of movement for a particular phoneme is heavily dependent on the phonemes
preceding and following. This phenomenon is known as coarticulation indicating that to a certain
extent, two neighbouring phonemes have a joint articulation. As the number of possible contexts
a phoneme can occur in is very large, this massively increases the possible variability in phonetic
patterns.
A impressionistic model of the process can be thought of as follows. Each allophone has a
set of canonical articulator positions. When each phoneme is “input” to the production apparatus,
the articulators start to move to those canonical positions and because of this “aiming” effect,](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-193-320.jpg)
![Section 7.4. Phonology: the linguistic organisation of speech 173
As a final example, consider the example of MINTS and MINCE that we mentioned before.
While the canonical pronunciation of these words can be different (/m ih n t s/ and /m ih n s/),
when transcribing speech, we are in the position of having two words which sound the same.
We could use a single canonical pronunciation for both, but this would mean that in a process of
word formation (described below in Section 7.4.2) we would need some process or rule which
would delete the /t/ of MINT when the plural morpheme was added. Or we could use different
transcriptions for each, meaning that we would have to acknowledge that two transcriptions can
sound the same. Or, two words could have the same narrow transcription but different broad
transcriptions; but again we would have to decide what this was and whether the [t] should be
deleted. The main point is that there is no single sure way to determine the transcription for real
speech. While useful in various ways, all transcriptions are an enforced symbolic descriptions of
what was said and none reflect the “full reality” of the message and signal from every perspective.
7.3.7 The distinctiveness of speech in communication
Given these two problems of segmentation and classification, one can rightly ask, how then is it
possible that humans can perform these tasks and thereby decode the speech when listening? The
answer again lies in our model of semiotic communication. We find that people make heavy use
of top down information when listening to speech; such that, using their internal grammar and
knowledge of the conversation and context, they form strong hypotheses of the words that they
expect to hear. This limits the choice of phoneme sequences which could agree with these words,
and by effectively narrowing the search in this way they make the problem tractable. This is also
how modern speech recognisers work. They are composed of two main parts; an acoustic model
which models the acoustic patterns for each phoneme, and a language model which dictates the
probability that a word (and hence phoneme) will occur in a given context.
We find that the degree of “clarity” or distinctiveness between phonemes in speech therefore
varies depending on the conversational context. In situations where the listener has a strong idea
of what will be said next, we find that speakers tend to be less distinct as the chance of confusion
is low. Hence we find that speech in conversations between friends and family can exhibit quite
indistinct acoustic patterns. Conversely, we see that in situations where the listeners cannot easily
guess what will be said next, we find the level of distinctiveness increases. Lectures or news
broadcasts are good examples of this type of speech.
7.4 PHONOLOGY: THE LINGUISTIC ORGANISATION OF SPEECH
Now that we have described the production process and the means by which discrete information
can be encoded in a continuous signal, we turn to the issue of how speech sounds are organised
linguistically. So far we have simply said that words are made from sequences of phonemes, but a
closer examination will show however that it is far from the case that any sequence of phonemes](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-197-320.jpg)
![Section 7.4. Phonology: the linguistic organisation of speech 175
0 → s 1
1 → p 2
1 → t 2
1 → k 2
2 → r 5
1 → p 3
3 → l 5
1 → k 4
4 → w 5
From examination of phone sequences, we can gradually build a grammar which generates
all and only those sequences which occur. Table 7.12 defines one such set of rules for a grammar of
the legal phoneme sequences for most single syllable English words. The set of terminal symbols
of this grammar is the set of phonemes, and the set of non-terminals is defined as a set V =
{0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16}. A state and transition diagram of this is shown
in Figure 7.14. Legal words can be created by starting at state 0, and following any path until an
end state is encountered. This grammar can be extended to polysyllabic words by the addition of
arcs from the terminal states back to the start states. Some care is needed however to ensure that
phenomena that occur only at the starts and ends of words are excluded from the grammar which
defines consonant sequences in the middles of words.
The grammar in Table 7.12 captures many facts that we have already mentioned, such as the
fact that /ng/ cannot occur word initially and so on. While our finite state grammar may accurately
describe what sequences can occur, its form is very “verbose” in that there are many regularities
which the grammar does not explicitly express. For instance after a /s/, only the unvoiced stops
occur, but we have no way of expressing this directly, we have to list these stops [/p/, /t/ and /k/]
explicitly.
We can look at this problem by taking a different approach as to how we define segments
(which in the following can be either phonemes or phones). In our account of segments and their
properties, we made use of such terms as “nasal”, “high”, “stop” etc, as these helped us categorise
the different ways in which sounds are generated. One of the central ideas in modern phonology
is that these terms are not simply labels for groups of somewhat similar segments but act as a
set of more basic building blocks. Hence we define a set of phonological primitives known as
distinctive features, and from arrangements of these we form segments. This system is sometimes
view as an additional level in the sentences/words/phonemes hierarchy, and indeed it follows on
from this in that there are fewer distinctive features that phonemes, and arrangements of these few
distinctive features give rise to a greater range of phonemes. One significant difference though is
that distinctive features do not arrange in sequences to form segments; rather their values are all
realised at the same time instant. In this way we can see that the process of phoneme formation
is one of “mixing” as opposed to “sequencing”. One analogy therefore is the way that the full
spectrum of colour can be formed by mixing the primary colours. Another comes from atomic](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-199-320.jpg)
![176 Chapter 7. Phonetics and Phonology
0 → p 1
0 → k 1
0 → b 1
0 → g 1
0 → f 1
1 → l 7
1 → r 7
0 → t 2
0 → th 2
0 → d 2
2 → w 7
2 → r 7
0 → p 7
0 → t 7
0 → k 7
0 → b 7
0 → d 7
0 → g 7
0 → ch 7
0 → dz 7
0 → m 7
0 → n 7
0 → f 7
0 → th 7
0 → s 7
0 → sh 7
0 → v 7
0 → dh 7
0 → z 7
0 → zh 7
0 → l 7
0 → r 7
0 → w 7
0 → y 7
0 → h 7
0 → E 7
0 → s 3
3 → p 4
3 → t 4
3 → k 4
4 → r 7
3 → p 5
5 → l 7
0 → s 6
6 → p 7
6 → t 7
6 → k 7
6 → m 7
6 → n 7
6 → l 7
6 → r 7
6 → w 7
7 → a 9
7 → e 9
7 → i 9
7 → o 9
7 → u 9
7 → uh 9
7 → ii 8
7 → ei 8
7 → aa 8
7 → uu 8
7 → ou 8
7 → oo 8
7 → ai 8
7 → au 8
7 → oi 8
7 → i 8@
7 → e 8@
7 → u 8@
7 → 7 @@
8 → 8 @
8 → E 9
9 → t 14
9 → b 14
9 → d 14
9 → g 14
9 → ch 14
9 → dz 14
9 → ng 14
9 → f 14
9 → th 14
9 → sh 14
9 → v 14
9 → dh 14
9 → z 14
9 → zh 14
9 → l 10
9 → s 10
10 → p 15
10 → t 15
10 → k 15
9 → l 11
11 → m 15
9 → n 12
9 → l 12
12 → d 15
9 → f 16
9 → p 16
9 → k 16
9 → m 13
13 → p 15
Figure 7.12 Complete grammar for syllable structure using classical finite state grammar rewrite
rules
physics; once atoms where thought the primary and hence “indivisible” building blocks of nature;
but subsequently electrons, protons and neutrons were discovered. Various arrangements of these
three types of particle gives rise to more than the 100 known elements. Likewise, phonemes were
thought of as the “atoms of speech”, but now we believe that a more parsimonious model can be
built with reference to more basic units.
When using distinctive features, a common approach is to define each segment as a feature
structure, of the type we introduced in Section 4.5. So we can define a segment [t] as
VOICED false
PLACE alveolar
MANNER stop
From this, “t” simply becomes a label which we attach to this particular set of values. This
notion has many attractive qualities. Firstly we can argue that it is somewhat more scientifically
economical to have a basic set of say 12 primitives, and then to form a larger set of 40 segments
from these, rather than have just a flat list of segments. Using our physics analogy we can say that
regarding classification of the elements, instead of just saying that we have hydrogen, helium, etc,
we instead say that these are made from protons, neutrons and electrons, and that the number of
protons and electrons fully defines the difference between each element.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-200-320.jpg)
![Section 7.4. Phonology: the linguistic organisation of speech 181
VOICING true
#
+ PLURAL →
VOICING true
#
z
VOICING false
#
+ PLURAL →
VOICING false
#
s
In fact, we can do better than this, by making use of some further devices common in phonological
theory. One is the principle of markedness which is a linguistic term for the concept of default.
The idea is that given a binary or other distinction, one form of behaviour is normal, unremark-
able and therefore unmarked, whereas the other form is noticeable, stands out and is therefore
marked. In our example, we could say that the unmarked plural morpheme is simple /s/, but some
special behaviour is invoked if this is added to a word that ends in a voiced consonant. Hence we
would not need a rule for joining /s/ to many words; this would be the default rule that happens
when any morphemes are joined (i.e. no changes occur). Hence we only need a rule for voiced
cases, which would be:
VOICING true
#
+ s →
VOICING true
#
z
We might well ask why it should be that /s/ is chosen as the default rather than /z/ (note that we
shouldn’t be misguided by the spelling convention here) and this is in fact a prime concern in
developing phonological theories, namely that we choose “natural” rules rather than simply ones
that describe the date.
At this stage it is worth examining why we have chosen to describe the plural formation
effect as a phonological one rather than a phonetic one. Wouldn’t it be possible to simply say
that this is a matter of allophonic variation or coarticulation? While the rule is clearly based on
articulatory principles, it is included as part of the phonological because it is seen as a rule of
English rather than a general process of speech. It is quite possible to overrule this effect and say
/d ao g s/ and /k ae t z/. More importantly however, we see that there are many non-plural English
words that end in /s/ which don’t undergo this rule, for example {BASIS, /b ey s ax s/} and {SENSE,
/s eh n s/}. It no more difficult or natural to say PENS as /p eh n z/ than PENCE as /p eh n s/. This
demonstrates that this plural formation rule is one of grammar and phonological organisation and
not some speech encoding effect.
English is rich with such word formation rules. For example, one common process, known
as palatalisation, describes the process whereby consonants move from their original place of
articulation to a palatal place of articulation, usually resulting in a [sh] phoneme. So if we take
examples such as:](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-205-320.jpg)
![Section 7.4. Phonology: the linguistic organisation of speech 183
vocalic
consonantal
high
back
low
anterior
coronal
round
tense
voice
continuant
nasal
strident
Figure 7.16 Binary features in the Sound Pattern of English
7.4.3 Distinctive Features and Phonological Theories
In defining a feature system properly, one must formally define the attributes and the type of values
they can take. So far we have described a somewhat traditional set of features (manner, place
etc) each of which can take a multiple values (e.g. dental, alveolar, or palatal) and have made
use of descriptive devices, such as “assimilation” to describe how sounds change in particular
environments. In phonology researchers attempt to improve on this, by defining features and rule
mechanisms in an attempt to explain these sorts of effects in a more elegant and parsimonious
manner.
The Sound Pattern of English [91] (known as “SPE”) was a seminal landmark in phonology,
in that it was the first full account of how to describe English phonology formally, that is, with
precise features and rules. SPE uses a binary feature system, in which each phone is described
by 13 features each of which takes a +/- value. The SPE binary feature set is shown in Table
7.16. While some of these have the familiar interpretations, a few are new. ANTERIOR is used for
consonants and indicates a front constriction, CORONAL, also used for consonants, indicates an
arched tongue. Combinations of these are enough to distinguish most English consonants. TENSE
is used for vowels in a very similar way to length and STRIDENT is used to indicate a “noisy”
phone such as [s] or [f].
SPE showed how many seemingly desperate phenomena could in fact be explained by rules
that made direct access to the feature set. Nearly all rules were posited in terms of context-
sensitive rules (these rules are used to define a third fundamental type of grammar in addition to
finite state rules and context free rules that we saw before.). For example, the tensing rule that
describes how word patterns such as DIVINE, DIVINITY and PROFANE, PROFANITY arise is given
as:](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-207-320.jpg)
![184 Chapter 7. Phonetics and Phonology
V → [−tense] / CVCV
meaning that the vowel V in question, becomes lax (the opposite of tense) when it occurs before
a CVCV sequence (such as the /n ih t iy/ of PROFANITY). Even for different feature sets, models
and theories, rules of this type are very common. It is worth noting however that the use of
context-sensitive rules here is merely a notational convenience; most phonological rules do not
use recursion, and so the above can readily be implemented with finite state rules, if required for
practical purposes.
Despite its success in many regards, several problems immediately appeared with SPE. One
of the main problems is that if there are 13 binary features, this should give rise to 213 = 8192
possible phones. Not only is this considerably greater than most phonologists would agree exist,
it is clear that many combinations (e.g. [+tense, +anterior, +high]) simply don’t correspond to
any known or even theoretically possible phone. For this and many other reasons, a huge flurry
of activity started with the aim of defining more succinct and more “natural” systems of features,
and accompanying rules.
In the 1970s, two theories broke from the mold of SPE by proposing that sound patterns
could better be understood if more elaborate rule and representation mechanisms were used. Au-
tosegmental phonology [180] uses two (or more) tiers, in which we have a phoneme tier (roughly
similar to SPE) and a tonal tier. The idea is that rules can be written to operate on just one tier or
the other, which is much simpler that would be the case if the phonemes and tones had to be repre-
sented on a single tier. The other theory is that of Metrical Phonology [284] which was designed
to explain phonological phenomena of stress and intonation (described below). Together, these
theories and some that follower were often called non-linear phonology (this is a different use
of the term “non-linear” to that in mathematics), because their representations were not sequences
of segments as in traditional phonetics, phonology and SPE. Since then a bewildering variety of
other theories and models have been produced. Among these are Natural Phonology [137] which
argued against the sometimes very abstract positions taken in theories such as SPE (for instance
the example we gave where the underlying representation of SIGN would require a [g] segment).
Despite the large number of different theories, some trends have stood out in how phonology
has developed. One strand of work has attempted to move away from the context-sensitive rule
formalisms described above. Mainly this follows a general move in this direction in other fields of
linguistics such as syntax, and is motivated by a desire to avoid the complications of determining
what these rules really stand for cognitively. For instance, it is often difficult to believe that a
speaker or hearer has to apply a whole sequence of rules in just the right order to generate the
surface form of a word. Instead, a series of constraints are proposed, from which new words can
be formed. Theories of this school include Dependency Phonology [145], [144], and Govern-
ment Phonology [248]. Many theories use feature systems more elaborate than those expressible
naturally in feature structures. The idea here being that features form hierarchies, and in doing so,
we can use these hierarchies to naturally limit the number of feature combinations and rules, thus](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-208-320.jpg)
![Section 7.4. Phonology: the linguistic organisation of speech 185
avoiding the “segment explosion” problem of SPE. The theories of Articulatory Phonology [70]
and Feature Geometry [100], [306], [386] are based on this idea.
In addition to the issues discussed in this chapter, it should be understood that the field of
phonology tackles many issues which aren’t directly relevant to speech synthesis needs, such as
say finding the set of language universals. Optimality Theory is one such model which has
gained particular attention in recent years [362]. In addition to this, the emphasis on phonol-
ogy even within directly relevant topics is often somewhat different from what we require from
an engineering perspective. In particular the old issue of “economy” comes up again and again,
with phonologists striving for every more natural and economical representations, without usually
much interest in assessing the processing requirements incurred. The general trends in phonol-
ogy are clear though; a move away from finite state, linear models towards more complex feature
organisations, a move from ordered context-sensitive rules to systems which use constraints and
dependencies, and a move from procedural systems to declarative systems. Somewhat ironically,
while most phonologists trace their work back to SPE, the “formal” side of this approach has
largely been forgotten, and few works in phonology since have anywhere near the level of ex-
plicitness and comprehensiveness of that work. A notable exception comes from the work of
Computational Phonology [43].
7.4.4 Syllables
One of the most important aspects of phonology concerns the structural representation of sound
patterns above the phoneme. We have already seen that morpheme boundaries are significant (e.g.
in the difference between the realisation of PENS and PENCE) as are word boundaries. In addition
to these, we find it is very useful to make use of a third unit, the syllable as this also helps explain
many of the effects and patterns in speech.
As we have already seen, the syllable provides a solid basis for defining the phonotactic
grammars of what constitutes a legal word. If we define a grammar for a syllable, it is an easy
matter to define a grammar for any words by simply including more syllables and any possibly
special sequences for word beginnings and ends. That said, it is interesting to ask real evidence
do we have for syllables as real entities. We seem to instinctively know that SET, FROM and
BATH have one syllable; that TABLE, WATER and MACHINE have two, and OPERATE, COMPUTER
and ORGANISE have three. These judgments seem intuitive and are not dependent on any expert
knowledge. Secondly, if we consider singing we see that in normal singing each note is sung with
one syllable. If this principle is broken the singing can sound awkward if syllables are inappropri-
ately squeezed in or elongated over several notes. Thirdly, syllables seem to act as the basic unit
of many aspects of prosody. Syllables carry stress such that in THE PROJECT the first syllable of
PROJECT is stressed, but in PROJECT THAT LIGHT the second gets the stress. There are no other
possibilities - it is not possible to stress the /p r/ and not the vowel, or any other individual phone.
So what exactly is a syllable? We will take the view that the syllable is a unit of organisation.
As such, all words are composed of a whole number of syllables, and all syllables are composed](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-209-320.jpg)
![Section 7.4. Phonology: the linguistic organisation of speech 189
sented only by its phonemes. There are main related noun/verb pairs in English which are dis-
tinguished by stress, such that we have {RECORD NOUN, / r eh1 k ax0 r d/} contrasting with
{RECORD VERB, / r ax0 k ao1 r d/} ; and likewise {PROJECT NOUN /p r aa 1 j eh2 k t/} contrast-
ing with {PROJECT VERB, / p r ax0 j eh1 k t/ }. That said, there are very few genuine minimal
pairs in English. Some have gone to considerable lengths to argue that stress can in fact be derived
from just the phonemes (this is advocated in SPE). Valiant though these attempts are, there are
simply too many words with unpredictable stress patterns for us to be able to derived stress from
phonemes for all words.
Stress can be graphically represented in the sort of tree used to describe syllable constituent
structure. In our previous example the nodes at the syllable level were unlabelled; we can now
show the same tree but with stress labels at the nodes:
2
Rhyme
Nucleus
ih
Coda
n
0
Onset
f
Rhyme
Nucleus
ax
Coda
r
1
Onset
m
Rhyme
Nucleus
ey
0
Onset
sh
Rhyme
Nucleus
ax
Coda
n
An alternative to the use of absolute levels of stress such as primary and secondary, is the
theory of metrical stress [284], [286] in which stress is purely relative. In this, a binary tree is
used to represent each word, and within this, each syllable is defined as either strong or weak. So
INFORMATION would be defined as:
W
S
in
W
for
S
S
ma
W
tion
The strongest syllable is the one which only has S nodes above it; everything else is defined relative
to this.
If lexical stress is a relative phenomenon, what of single syllable words? As there are no
other syllables to contrast with, the stress status of such words is technically undefinable. However,](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-213-320.jpg)
![190 Chapter 7. Phonetics and Phonology
by convention, most of these words are given a primary stress or strong stress in their syllable. The
exception is content words, which are given a primary stress if in their full form (/f r ah m/) or no
stress in their reduced form (/f r ax m/).
In order for us to perceive stress differences, stress patterns must of course manifest them-
selves in the acoustic signal. How is this achieved? Firstly, taking a contrast between “normal”
syllables and their stressed version, we can say that in the case of stress, the syllable is spoken
with greater articulation effort, that is, the articulators are more likely to move to their canonical
positions, and in so doing, produce a syllable with many of the idealised sounds which we would
expect from our discussion of articulatory phonetics. As a by product of this, we can see the that
the syllable seems more “distinct” in the spectrogram; the formants of the vowel are clearly ob-
servable, and it is often easier to ascertain the identity of the vowel from the formants of stressed
syllables than for unstressed ones. Likewise with the consonants, stops often have more distinct
closure and burst patterns in stressed syllables and so on. Following from the great articulation
effort, stressed syllables are general longer in time than normal versions. The source acts differ-
ently in stressed syllables also. The voice quality is likely to be more clearly periodic (not creaky
or breathy). Also, the pitch of the syllable is likely to be higher in a stressed syllables than in the
syllables immediately surrounding it. This effect is difficult to determine in isolation however, as
the intonational patterns of the utterance interact strongly here (described in more detail in Section
9.1.
By contrast, in reduced syllables, the opposite effects seem to occur. The articulation effort
is diminished to the extent that any distinctiveness in the vowel quality is absent - all vowels are
either /ax/ or /ih/. Consonants in reduced syllables tend to be less clear. The duration of reduced
syllables can be very short indeed, even to the extent of being non-detectable in faster or more
casual speech. Finally, it is unlikely that the pitch in an unstressed syllable with be higher than
its surrounding syllables (note however, that the presence of a reduced syllable does not cause a
decrease in pitch).
7.5 DISCUSSION
The study of phonetics has a long and rich history [97]. Just as the world’s most famous detec-
tive is the fictional Sherlock Holmes, so indeed is the world’s most famous phonetician. Today the
Henry Higgins of Pygmalion and My Fair Lady would probably appall any self-respecting linguist
or phonetician with his dogmatic and prescriptive view of how English should be spoken. Nev-
ertheless, Higgins probably does share some genuine features in common with the early British
school of phonetics.
The study of modern phonology is widely regarded to have started with the “Prague School”
of linguists, of which the main figures were Nikolay Trubetzkoy and Roman Jakobson. They
established the principle of the formal (or scientific) study of phonology and the notion of the
phoneme. Most importantly, they showed the semiotic basis of phonological contrast with the](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-214-320.jpg)
![Section 7.5. Discussion 191
idea being that the linguistic use of sounds could only be studied as within a system and not in
individual, purely phonetically terms. Jakobson collaborated with two other notable figures in
speech and linguistics, Morris Halle and Gunnar Fant to produce a coherent system of distinctive
features. As already mentioned, the Sound Patter of English was a seminal work, both in its scope
and its affect in proposing formal methods for phonological work.
7.5.1 Further reading
The most popular work in introductory phonetics is undoubtedly Peter Ladefoged’s classic a
course in phonetics [271], and is highly recommended. Another excellent book, which extends
the subject and gives more detail on acoustic issues is Clark and Yallop [97]. A more advanced
and comprehensive book is Laver’s Principles of Phonetics [275].
There are many excellent introductory text books on phonology including Giegerich [174],
Katamba [245], Lass [274] and Anderson [17]. Clark and Yallop [97] is again particularly recom-
mended as it gives a thorough introduction to phonology, does so with reference and comparison
to phonetics, and discusses the issues involved from the type of traditional scientific perspective
familiar to most engineers and scientists. Unfortunately, nearly all these works are quite introduc-
tory and few venture beyond SPE type phonology. While there are plenty of books and articles
which cover more recent work, they typically focus on one particular theory or model, making it
hard to perform the type of comparison that would be necessary for us to adopt any one of these
for speech technology purposes. Books and articles on the various schools of phonology are listed
above.
7.5.2 Summary
Phonetics
• Speech sounds can be described with reference to either their articulatory or acoustic fea-
tures.
• Vowel sounds are described with the primary dimensions of high/low and front/back, and
secondary dimensions of rounding and length.
• Consonants are described in terms of voicing, manner and place of articulation. The main
values manner can take are stop, fricative, nasal, affricate and approximant.
• Computer analysis can be used to study the acoustic patterns in speech. The primary repre-
sentation is that of the spectrogram.
• The resonances of the vocal tract are called formants, and these are thought to be the primary
means by which listeners differentiate between different vowel sounds.
Communication with speech
• A phoneme is a linguistic unit of speech defined on the semiotic principle of contrast.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-215-320.jpg)
![Section 8.1. Pronunciation representations 195
this with just [t] (the IPA symbol for an unaspirated [t]), using [th] for cases where aspirating
exists. Context will still play a part, because coarticulation (as opposed to allophonic variation)
still operates, but the degree of contextual effect will be lessened. As many phonemes have more
than one allophone, the number of base types may well increase (to say 60), but because we can
do without a left and right context for each, then the overall number will be considerably lower.
8.1.3 Difficulties in deriving phonetic input
There is however a drawback with this approach. The problem with phone representations was
discussed previously in Section 7.3.2 where we stated that although phone descriptions could be
very useful in certain circumstances, they should be treated with caution as they are a discrete
representation enforced on a continuous space; and as such only crudely reflect the actual pro-
cesses going on. To give an example, let us consider again an example of place of articulation
assimilation that occurs with words such as
TEN + BOYS → t e n + b o i z → t e m b o i z
where the alveolar place of articulation of the /n/ of the first word assimilates to the bilabial place
of articulation of the /b/ of the second word (bilabial). There is no doubt that that something
is going on here; the [n] certainly seems somewhat different from a “normal” [n], and this can
be verified from both acoustic and articulatory measurements. This is not the issue; rather, it is
whether the [n] has turned into an [m]. Such a statement says that there are only two choices; an
[n] or an [m] and in certain situations the [n] keeps its identity and in others it changes. Consider
the situation of saying TEN BOYS with a pause between the two words; it should be clear that
the [n] sounds like a “normal” [n] and that if we said the (made up word) TEM (/t e m/) as TEM
BOYS with a short pause, we can tell the difference between the two. Now if we say a sequence
of utterances TEN BOYS starting with a long pause and gradually shortening it, until there is no
pause, we see that there is a gradual shift in the nature of the /n/, until, when we talk quickly, it
seems impossible to make the distinction between TEN and TEM. This shows that assimilation is
certainly taking place, but that it is not the case that an underlying [n] is at one point realised by
the phone [n] but at a certain points suddenly becomes a different phone [m]. The process is a
continuous one of change, and it is quite inappropriate to model it as a discretely. We see exactly
the same effects with /r/ colouring, nasalisations and reduction.
In a way, we are forcing a discrete level of representation onto a range of effects that have
a substantial continuous component; and this is damaging in all sorts of ways. As an example,
consider again the phenomena of reduction; whereby we have AND represented underlyingly as [a
n d], but where this undergoes reduction, resulting in a range of possible phonetic forms including
[a n], [n d], [n], [n] and so on. It may be clear that some level of reduction has occurred, but at
what stage do we delete the [d]? Where do we draw the boundary between a full [a] and it being
reduced to a schwa? Studies of hand human labelling have shown that even trained phoneticians](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-219-320.jpg)
![Section 8.1. Pronunciation representations 197
Onset
t
Rhyme
Nucleus
aw
Coda
p
Onset
l
Rhyme
Nucleus
a
Coda
s t
From these trees, we can access the information we require; the phonemic context can be
read directly by accessing a phoneme’s neighbours; the syllable structure is presented as a tree,
and from this we can distinguish between /s t/ in a syllable onset in STRENGTH, and offset in
LAST.
8.1.5 Abstract phonological representations
We have just seen that despite potential issues with data sparsity, in general, higher level phonemic
representations are preferred over lower level phonetic ones. In pursuing this approach however,
we should ask ourselves whether the phonemic representations we have been using are in fact
optimal; might there not be an even more abstract, and less linear representation that would help
the dimensionality problem? Progress in modern phonology has in fact moved away from using
linear sequences of phonemes or phoneme like units as the “deep” or canonical representation;
freed from the constraint that representations must be linear, the more structured “non-linear”
representations reviewed in Section 7.4.3 could be considered instead.
Ogden et al [334] discusses just this issue and point outs that the problem with specifying
phonemic style representations is that this concentrates overly on the paradigmatic dimension; the
phonemic system is based on what contrasts occur at a particular point, and the issue of what
phonemes follow what other phonemes is left to a different, phonotactic part of the phonology. In
fact, we see that it is indeed possible to construct representations in which both the normal notion
of phonemic contrast, and a mechanism for governing legal sequences are handled in a unified
way. While phonemic representations are less susceptible to borderline decisions than phonetic
representations, they are not entirely free of problems.
At present, we know of no unit selection techniques which make use of such abstract phono-
logical representations, but this possibility is sure to be explored eventually. While there is every
chance the field will indeed move in this direction, because of the lack of practical proof we will
continue with phonemic representations for the remainder of the book. We should point out that
phonemic representations do have one significant advantage over virtually any other type of pho-
netic or phonological representation in that they are generally easy for people to work with. Overly
phonetic representations can be difficult in that it takes time and effort to fill in the phonetic detail;
overly abstract representations require a significant knowledge and understanding of the partic-
ular theory which underlies them. With a little training, most people get the hang of phonemic
transcription and are able to read and generate transcriptions. This is important for text-to-speech](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-221-320.jpg)
![198 Chapter 8. Pronunciation
because at some stage, someone must of course actually enter all the lexicon entries by hand. In
doing so, we wish to have a system which is both quick (there might be hundreds of thousands
of entries) and accurate, and the phonemic system seems to fulfill these criteria reasonably well.
That said, we should make the point that it is perfectly possible to enter the pronunciation for a
word in the lexicon in terms of phonemes, and then by rule convert this to either a deeper or more
superficial representation.
Finally, using a linear phonemic representation has the benefit that it makes automatic database
labelling considerably easier. In Chapter 17 we shall consider the issue of labelling a speech
database with phoneme units, both by hand and by computer. As we shall see, it is easier to per-
form this labelling if we assume a linear pronunciation model, as this works well with automatic
techniques such as hidden Markov models.
8.2 FORMULATING A PHONOLOGICAL REPRESENTATION SYSTEM
Having in principle chosen a system which uses phonemes and syllable structure, we now turn to
the issue of how to actually define this for a given speaker, accent or language. It is essential that
this is done with care; if we decide say to use a single phoneme /ch/ instead if /t/+/sh/, we have to
consistently use this, such that we represent CHURCH as /ch u r ch/ and not /t sh u r ch/ or /ch u r t
sh/ and so on.
The following sections explain how a phonological system is developed. The first thing that
we must decide is which accent we are dealing with from the language in question. In general it
is not possible to define a phoneme inventory which can cover multiple accents, so whatever will
we develop will be particular to that accent. One solution is to this is to use the actual speaker the
TTS system will be based on, and from analysis of his or her phonemic contrasts ensure that the
correct pronunciation system is found. This solution can require too much effort, so more general
accents often have to be defined. In this chapter, we will conduct an exercise on developing
a representation for just two accents of English, namely General American and British English
received pronunciation (RP). The principles involved however are the same for most other accents
and languages.
It does seem somewhat wasteful that separate pronunciation systems and therefore lexicons
are required for British English and American English; while differences of course exist, pronun-
ciations in the two accents aren’t completely different. Fitt [161], [162], [163] proposed a solution
to this where a more abstract pronunciation drawn from a large set of phonemes was used as a
base lexicon. From this, filters could be used to generate accents for any accent of English.
8.2.1 Simple consonants and vowels
Here we will adopt the standard approach of using the principles of minimal pairs and phonetic
similarity to develop a phoneme inventory (see Section 7.3.2). This is not an exact procedure and](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-222-320.jpg)
![Section 8.2. Formulating a phonological representation system 199
it is possible to come up with two equally valid phoneme inventories for a given accent. There
is therefore some discretion in determining a phoneme inventory; the key point is not to find
“absolute truth” of what constitutes the set of sound contrasts; but rather to define an inventory
that suits our purposes, and then standardise on this; so that exactly the same conventions are used
in all parts of the system. Experience has shown that “drift” in phoneme conventions, whereby say
we transcribe BOTTLE as /b aa t el/ in one place and as /b aa t ax l/ in another can have a disastrous
effect as the data, rules, synthesiser and lexicon can become misaligned. The crucial point is to
chose a standard as best we can and stick to just that standard everywhere in the system.
Recall that the principle of minimal pairs states that it is not the similarity or dissimilarity of
two sounds per se which is the issue, but rather whether they distinguish two separate words. By
using this test, it is straightforward to determine a set of contrasts in a particular position. So if we
take the form [X i p], we can find a set of contrasts as follows:
p PIP m MIP l LIP
t TIP n NIP r RIP
k KIP s SIP w WHIP
b BIP sh SHIP y YIP
d DIP z ZIP h HIP
The /X i p/ pattern doesn’t give the full list of possible contrast because not every possible legal
phoneme sequence actually ends up as a real existent word. But if we take some further patterns,
we can elicit further contrasts:
X /X eh t/ X /X ih n/ X /X ay n/
p PET sh SHIN v VINE
b BET f FIN dh THINE
d DEBT th THIN f FINE
g GET s SIN s SIGN
m MET sh SHINE
n NET
When we consider one particular pattern, the analysis seems clear: PET and BET are different
words, so /p/ and /b/ must be different phonemes. It is reasonable to ask though, how we know
that the sound at the start of PET is the same sound as at the start of say PIT? There is no known
deterministic or objective way to prove that this is the case, all we can do is rely on judgment and
state that in terms of articulation, the two sounds are both produced with the lips, are unvoiced
are stops and so on, and that if we examine them acoustically we see they are similar. But they
are rarely so similar to be taken as identical; their articulation and acoustic patterns are affected
by the following vowel and this does have some effect. To a certain extent, there is a leap of faith
required to say that PET and PIT start with the same phoneme. This problem is even more acute
when we consider sound contrasts in different positions:](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-223-320.jpg)
![Section 8.2. Formulating a phonological representation system 201
such as THE, THEIR, THEM, THEN, THERE, THESE, THEY and THIS. When we take the remainder
of the words in which /dh/ is found we see that it is found in many “old” or traditional sounding
English words such as FATHOM, HEATHEN, HITHER, LATHE, LITHE, LOATHE, NETHER, SCATHE,
SCYTHE, SEETHE, SHEATHE and TETHER. In fact, it only occurs in a relatively few “normal”
English words such as MOTHER, LEATHER and HEATHER. Clearly /dh/ is a phoneme, and for
purposes of defining our phoneme inventory, it gets a place along with the other fricatives. There
are consequences that arise from its use though, in that because of the dominance of its used in
function words, it is difficult to obtain “typical” characteristics of the phoneme (say in terms of its
average duration) that are not swamped by the fact that it is used as a function word (which are
often very short, regardless of what phonemes make them up).
Now consider /zh/, the voiced equivalent of /sh/. This occurs in words such as BEIGE /b ey
zh/. Again, while there are few direct minimal pairs, some exist (BEIGE vs BAIT or BANE) and
so we classify /zh/ as a phoneme. Problems again arise from patterns of use. It has been argued
(e.g. [97]) that /zh/ is a relative newcomer to the English language as in the past, words such as
TREASURE and MEASURE were pronounced /t r eh z y uh r/ and /m e z y uh r/ and that it is
relatively recently that these have been spoken as/ t r eh zh er/ and /m eh zh er/. /zh/ presents the
same problems that arose with /dh/, but in this case they are even more difficult because of the very
few words naturally occurring with /zh/. These problems with /zh/ and /dh/ can lead to problems
in database collection and analysis (see Section 14.1).
The sounds [ng] and [h] are noteworthy because they are severely restricted as to the posi-
tions they can occur in. Primarily, /h/ can not occur at the end of a syllable or a word, and /ng/
can not occur at the beginning of a syllable or word. There are further issues with /ng/ in that as
it only occurs in a post-vocalic position, it suffers from the loss of contrast that affects all nasals
in this position. So while it is clear that SIN, / s ih n/, SIM, /s ih m/ and SING, /s ih ng/ all end in
different sounds, it is less clear what the correct representation for words like THINK should be.
In words such as this the last stop is in a velar position, and the nasal takes on a velar quality also
(this is similar to the TEN BOYS effect we discussed earlier). The issue is then whether we should
use a velar nasal /ng/ or a dental nasal /n/; there is no contrast possible in this position and so it is
a question of convention which we should pick. As with other such issues, the main goal is to aim
for consistency; if /ng/ is used here then it should be used in all other post-vocalic pre-velar stop
positions.
8.2.3 Diphthongs and affricates
So far we have concentrated on the paradigmatic dimension, by contrasting (say) HAT, HEAT and
HOT and showing that in a frame of /h X t/ we can distinguish a number of phonemes. We were able
to do this because a word such as HOT, clearly divides into three segments. We assign provisional
phonemic status to each, swap in other sounds and thereby determine the system of phonemic
contrasts. In many words however, it is not so easy to segment the word into distinct parts, and
in such words we face a question of not only what the contrasts are, but how many positions are](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-225-320.jpg)
![Section 8.2. Formulating a phonological representation system 203
speech. Setting aside the phonemic question, consider for a moment how this differs phonetically
across languages. In English, an alveolar approximant /r/ ([G] in IPA) is most common and it is
quite common for this to have a retroflex (curling of the tongue) quality as well. In contemporary
Scottish English, a tapped alveolar (D) is quite common, and is separate again from the “classical”
Scottish trill [r] (now usually only seen on stage or in film). In French and many accents of German
a uvular [J] sound is the norm, with other accents of German having a more alveolar quality. In
Spanish, words such as PERRO are formed with a trilled [r], and this forms a minimal pair with the
tapped [D] of words such as PERO. It seems incredible given all this variety, that somehow all these
sounds are grouped together across languages and represented in writing as an r. Do they even
have anything in common? While we will not go into cross language issues here, it is worth noting
that words are quite commonly shared across languages, for example the words EMAIL, APRIL and
METRE. In the case of METRE, we see that it contains an r in all languages and is pronounced in a
wide variety of ways. If we study the acoustics however, we see that these different “types” of /r/
do share somewhat more in common in that they have characteristic energy bands at certain parts
of the spectrum.
When we consider the phonemics of /r/, we find that it has an interesting effect. First we
distinguish between rhotic accents which have /r/ in post-vocalic position (in English, these are
Scottish, Irish, General American) and non-rhotic accents which have no realised /r/ (most Eng-
land English, Australian and some New York accents). For example, in Irish English, there is no
“long a” sound [aa], which is found in RP BATH /b aa th/; this word is pronounced /b a th/. How-
ever, in words which have an /a/ and a /r/ following it, the /a gets lengthened, and phonetically
we have something like [h aa r d]. Being a phonetic realisation effect, this doesn’t concern us
unduly, but it does show how significant (and unexpected to the unwary) the presence of an /r/ can
be. More significant are the non-rhotic accents. In these, it is not the case that the /r/ is simply
not present; rather (historically or by having some deep presence) it has the affect of altering the
quality of the vowels before it. So in a word like HER, which might be represented as /h uh r/ in
a rhotic accent, we do not get /h uh/ but rather /h @@/ which has an extended vowel. It is almost
as if the /r/ has turned into a /@/ and is still present. Because of this, we are faced with a choice
between transcribing non-rhotic FIRE as /f ai @/ or invent a new symbol /ai@/ which would be
regarded as a triphthong.
8.2.5 Defining the full inventory
From the above we should have sufficient knowledge to construct a phoneme inventory for any
accent or language we are interested in. In practice we find that we don’t always have to design
a phoneme inventory from scratch; for many languages or accents the existence of a significant
computer speech research community means that phoneme inventories already exist. As these
communities often develop and distribute resources using these phoneme inventories, it can be
highly beneficial to adopt one of these rather that start from scratch. In the past, databases of
speech which contained hand labelled phone or phoneme transcriptions were highly valued, but](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-227-320.jpg)
![204 Chapter 8. Pronunciation
as phonetic/phonemic transcription can now be done automatically, today by far the most useful
resource is in fact a comprehensive lexicon. As the development of a large ( 100,000 word)
lexicon can take years of work, any existing resources greatly help. Lexicons are in general tied
to a specific accent of the language, and as such come with a phoneme inventory pre-defined.
So in many cases a phoneme inventory may exist for the accent we are interested in. By and
large, most phoneme inventories developed for computational work roughly follow the guidelines
just described. As just pointed out, there are however several ways to resolve some of the issues,
and so it is worth examining any predefined phoneme inventories in light of the above discussion
to assess suitability. Most predefined phoneme inventories have not been developed with TTS in
mind, rather they were developed for purposes of speech recognition or are just “general” phoneme
inventories developed for a variety of uses. Experience has shown that minor modifications in
predefined phoneme inventories can sometimes be beneficial for TTS purposes. The most common
problem is that sometimes the phoneme inventory has some “phonetic” features (such as markers
for aspirated or unaspirated, or a different label for the stop closure than the stop release), but these
can normally be removed with some automatic processing.
Table 8.1 gives a definition of a complete phoneme inventory based on a modification of
the TIMIT phoneme inventory. This was originally designed for use with the acoustic-phonetic
database of the same name and now one of the most well known phoneme inventories for General
American [173]. Most of the modifications we have made have been to eliminate some of the
purely phonetic distinctions in the original TIMIT inventory1. The main peculiarity of the TIMIT
inventory relates to the fact that sequences of schwa followed by /r/ are represented by a single
phoneme /axr/, and the vowel in words such as BIRD is treated similarly and represented as /b er
d/. All other combinations of vowel and /r/ are treated as two separate phonemes. Table 8.2 gives
a phoneme inventory for the British English RP accent, based on the University of Edinburgh’s
MRPA (machine readable phonetic alphabet) system [276]. The consonant system is more or
less the same; this is what we would expect from the facts of English accent variation. The
vowel system is different, and in particular the treatment of “r” is different because RP is a non-
rhotic accent. The pseudo-diphthong vowels /i@/, /u@/ and /e@/ are used in cases where there
1 For the record, the differences between this and the standard phone inventory are:
1. The /axr/ unstressed schwa+/r/ phoneme has been replaced by /er/. /er/ and the old /axr/ are now distinguished
by stress alone
2. the close phones bcl, tcl etc have been removed
3. the nasal tap /nx/ has been merged with the /n/ nasal
4. /hv/ has been merged with /hh/
5. the glottal stop phone /q/ has been removed
6. the syllabic consonants /en/, /em/, /eng/ and /el/ have been replaced by /ax/ + consonant sequences, e.g. /en/ →
/ax n/
7. /ix/ has been replaced by either /ax/ or /ih/, /ux/ by /uw/ and /ax-h/ by /ax/](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-228-320.jpg)
![Section 8.3. The Lexicon 211
encountering exactly the same words in different samples. Hence memorising common examples
is not in itself harmful [76].
8.3.2 Lexicon formats
In speech technology, it is quite often to find existing lexicons in what we can term simple dic-
tionary format. Such lexicons can be stored as simple ascii files, where each entry starts with
the orthography, followed by the pronunciation and possibly other information. The entries are
ordered by the orthography such that a few typical entries would look like this:
deep d iy1 p
deer d ih1 r
defend d ax0 f eh1 n d
define d ax0 f ay1 n
degree d ax0 g r iy1
delicate d eh1 l ax0 k ax0 t
delight d ax0 l ay1 t
demand d ax0 m ae1 n d
density d eh1 n s ax0 t iy2
It is clear that this format explicitly gives priority to the written form, as this is used to order and in-
dex the entries. Dictionary formats can be problematic though, for the reason that the orthographic
form of a word is not unique; homographs are fairly frequent and these must somehow be handled.
Two common amendments to this style are to have multiple lines with the same orthographic form
project p r aa1 jh eh2 k t
project p r ax0 jh eh1 k t
or to use more than one entry for the pronunciation on each line
project p r aa1 jh eh2 k t, p r ax0 jh eh1 k t
Even if this is done, we still somehow need to know that /p r aa1 jh eh2 k t/ is the noun form of
project, it is not sufficient to say that project simply has two different pronunciations. While
POS markers could be added, we soon find that when we also consider the issue of pronunciation
variants (which are arbitrarily different ways of pronouncing the same word) or orthographic vari-
ants (different ways of spelling the same word) we see that the simple dictionary style comes under
considerable pressure, and looses its main positive feature, that it is easy to read and process.
A solution to this is to build the lexicon as a relational database. In this, we have exactly
one entry for each word. Each entry contains a number of uniquely identified fields, each of which
has a single value. For for a simple word, the entry may just contain two fields, ORTHOGRAPHY
and PRONUNCIATION. It is a simple matter to add more fields such as POS or SYNCAT. Each
entry in a relational database can also be seen as a feature structure of the type we are now familiar
with, and because of this similarity, we will use the feature structure terminology for ease of](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-235-320.jpg)
![218 Chapter 8. Pronunciation
most people can correctly read aloud bobstay, cringles and jibe even they don’t know
what these words mean. Other terms are ore difficult however, such that most struggle with how
to read aloud fo’c’sle boatswain and gunwale if they are not familiar with the terms.
For such groups of words, the only answer is to consult a knowledgeable source; either a book on
sailing or a sailor in person.
In considering this issue, we come across an interesting philosophical issue regarding the
pronunciation of words. Quite simply, how do we know what the correct pronunciation for any
word is? Saying that word pronunciations are properly defined in traditional dictionaries avoids the
point; this would imply that words not in a dictionary have no standard pronunciation, and misses
the fact that dictionary writers try and listen to real speakers to elicit the actual pronunciation [484].
The best we can do is say that the “correct” pronunciation for a word is X because that is what
the population of speakers believe. For most common words, this is unproblematic; and everyone
of a particular accent will say the words WALK, DOG and HOUSE the same way. Problems do
arise though because people do not in fact always pronounce words the same way, even within a
single accent. The issue is whether we can establish a correct pronunciation, relegating the others
to some “incorrect” status, or whether we have to accept multiple correct forms. Establishing the
boundaries here is difficult, it seems equally absurd to insist on a “gold standard” of a single correct
pronunciation, or an “anything goes” policy where all pronunciations are equally legitimate. There
is a personal element in this as well; it doesn’t bother me whether someone says CONTROVERSY
as /k o1 n t r ax0 v er2 s iy/ or /k ax0 n t r o1 v ax0 s iy2/, but somehow when someone says /n
uw1 k uw2 l ax0 r/ for NUCLEAR this just seems wrong. To take another example, it is now quite
common to hear American English speakers to pronounce the “silent” /l/ in words like CALM and
SALMON - while this may seem clearly “wrong”, how can it be if millions of people now follow
this habit?
Difficulties most arise with regard to rare names. In most European countries, the native
pronunciation of names is well established, and even if a name is borrowed from one language to
make a new name in another language, we can often find the correct pronunciation in the original
language. The situation in the United States is significantly different, in that the sheer diversity
of names outweighs any other country. In the US, we see firstly that many names have moved
considerably from their original pronunciation, but secondly many names are genuinely made up
or so altered from the original that they effectively became new names2. This situation creates
many cases where there is only a handful of families with a particular name, making it difficult for
the lexicon developer to find the pronunciation. Even more problematic is that we find cases where
a handful of families share the orthography for their name but have different pronunciations. It is
often unknown even if these names share a common origin. How then do we determine the correct
2 Note that there is a consistent myth in that peoples; names were changed by officials during transit through the Ellis
Island immigration service. What evidence we have suggests that this is not in fact the major source of new names,
as some degree of consistency had to be kept between embarkation in the origin country and immigration procedure.
Many minor changes in spelling and so on certainly happened, but it was not the case that immigration officials simply
invented new names because they thought the real name was too awkward to handle.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-242-320.jpg)
![Section 8.4. Grapheme-to-Phoneme Conversion 219
pronunciation for such a name? As Spiegel [409] says:
Complicating the issue is that everyone is a pronunciation expert for their own name,
and (only slightly less so) for the people they know. Sometimes people are unaware
of other (valid) pronunciations for their names, even when the pronunciation is more
common.
Hence even if we decide that the correct pronunciation for smith is /s m ih1 th/ we will still
find enough people who complain that the system got it wrong because their name is pronounced
/s m ay1 th/. There is no objective real solution to this problem as this is one of system usage and
user requirements. The real point is that developers who deploy TTS systems should be aware that
the may not be any objective truth as to whether a system is saying a name correctly or not.
8.4 GRAPHEME-TO-PHONEME CONVERSION
A grapheme-to-phoneme (G2P) algorithm generates a phonemic transcription of a word given
its spelling. Thus it generates a sequence of phonemes from a sequence of characters3.
8.4.1 Rule based techniques
Grapheme-to-phoneme conversion by rule is perhaps the classic application of traditional knowl-
edge based rules in TTS. Indeed, in the past it was thought that this was how humans performed
the task; when presented with an unknown word they would apply a set of rules to find the pro-
nunciation. The most common approach is to process the character sequence left-to-right, and for
each character apply one or more rules in order to generate a phoneme. It should be clear that
these rules can’t operate in isolation, otherwise a t character would always result in a /t/ phoneme.
Rather the character context of the character is used and in this was we can generate /th/ when t
is followed by h.
Most approaches use context sensitive rewrite rules of the type introduced in Section 5.2.4.
Recall that these are of the form
A → B/d/C
which reads as “A is converted (rewrites to) d when preceded by B and followed by C”. In any well
developed system, multiple rules will match a given context, and so a means of choosing amongst
rules is required. A common technique for this is to order the rules and apply each in turn, so that
3 In this book, the use of the term grapheme-to-phoneme conversion relates specifically to algorithms which perform
just this task; elsewhere this term is sometimes used for the process which uses algorithms and the lexicon, and indeed
sometimes used for the entire process of text analysis. The term letter-to-sound rules is also encountered, but this
is increasingly misleading as most modern techniques don’t used rules in the traditional way. A final complication
regarding the terminology of this field is that strictly speaking a grapheme is a minimal unit of spelling rather than a
single character, so in that the th is described as a single grapheme. As our input is always a sequence of characters,
we are really performing character-to-phoneme conversion.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-243-320.jpg)
![220 Chapter 8. Pronunciation
the first matching rule is the one chosen. As this is a purely hand written technique the developer
has to order these rules him or herself. A common technique is to order the rules with the most
specific first. Rule systems of this kind include [8], [307], [155], [226], [136].
These systems can often produce quite reasonable results for highly regular systems with
a shallow orthography (e.g. Spanish). The results are generally very poor though for languages
like English and French, which has increasingly led to the adoption of data driven and statistical
techniques [300].
8.4.2 Grapheme to phoneme alignment
All the data driven techniques described below use algorithms to learn the mapping between char-
acters and phonemes. They use an existing lexicon as training data, and often words in this are
held out as test data and so these data driven techniques are only usable in cases where there a
comprehensive lexicon is available. In general words do not have a one-to-one correspondence
between characters and phonemes: sometimes two or more characters produce a single phoneme
(e.g. gh → /f/ in rough) and sometimes they don’t produce a phoneme at all (e.g. the e in hate
→ /h ey t/). A pre-requisite for training then is to align the characters with the phonemes for each
entry in the lexicon.
Most techniques for performing this are based on a dynamic time warping, (DTW) tech-
nique [343]. This works by assigning a cost to every phoneme that can align with a grapheme.
“Good” correspondences are given low costs while “bad” correspondences are given high costs.
The DTW technique is a dynamic program algorithm which effectively searches all possible align-
ments and picks the one which has the lowest total cost. This is a significantly easier task than
either writing the rules by hand or aligning every word in the lexicon by hand. For example, the
character s maps to the phoneme /s/ with a low cost, but maps to /n/ with a much high cost. In
many cases there is an obvious phoneme which maps from the character, but in other cases, e.g.
h this is used to represent many different phonemes (/sh/, /th/, /dh/ etc) in writing and so has a
number of low costs.
A probabilistic version of this was used by Jelinek [239], who used the hidden Markov
model Baum-Welch training technique (described in Section 15.1.8 to learn the probability that a
character would generate a phoneme. This was extended by Taylor [433] who showed that it was
possible to do this from a flat start meaning that no initial seeding is necessary. In general, the
alignment method is independent of the actual algorithm used, so that any alignment technique
can be used with any of the algorithms described below.
8.4.3 Neural networks
The first data driven G2P algorithm was the NetTalk algorithm developed by Sejnowski and Rosen-
berg [397], [396], [377]. This work became genuinely famous, not for its raw performance or use
in practical text-to-speech, but because it was one of the first real demonstrations of a neural net-](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-244-320.jpg)
![Section 8.4. Grapheme-to-Phoneme Conversion 221
work in operation. A popular demonstration of the system speaking the pronunciations it had
learned added to the fame of the system as it showed how the system performed better with each
training iteration. NETtalk uses a feed-forward network consisting of an input layer, a hidden
layer and an output layer. The input considers a character at a time, and uses a context window of
three characters before and three after. This results in a total of 7 groups of units in the input layer
of the neural network. Each group has a series of nodes which can be used in a binary coding
fashion to represent each character (For example, if we have 5 nodes, then a would be 00001, b
would be 00001, c would be 00001, and z would be 11010. A single group of units is used in the
output and a single phoneme is therefore generated.
A number of further studies have also used neural networks for G2P [191],[240] but this
technique’s importance is mostly due to it being an powerful demonstration in the early days of
machine learning.
8.4.4 Pronunciation by analogy
Pronunciation by analogy is a popular data driven technique first introduced by Dedina and
Nusbaum [127] and popularised by Damper and colleagues [422],[122],[123], [300], [27]. Its
motivation comes from studies of human reading and draws from studies by Glushko [175] and
Kay and Marcel [247]. The basic idea is that given an unknown word, a human considers known
words which are orthographically similar, and adapts a pronunciation of the nearest word to use
for the unknown word. For example, if we are confronted with a new word zate, we consider
that while we don’t know the pronunciation of this, we do know of many similar words date,
late, mate, all of which have the same ending pronunciation. In addition, we also know that
most words that start with z generate a /z/ phoneme, and so we apply both these analogies and
decide that zate should be pronounced /z ey t/. Marchand and Damper [300] give a review of
this technique and describe it as belonging to the family of machine learning techniques known as
memory based learning [119].
The algorithm works by comparing the characters in the input word with each entry in the
lexicon. For each matching substring, a lattice is built with the input characters and the out-
put phonemes (which are found by the alignment process). The lattice comprises a path of the
characters c1,...,cN and a set of arcs which span character sequences in the path. Each arc
represents one match between a sequence of characters and a sequence of output phonemes. Hence
in through we would have an arc from c0 (a dummy start character) to c2, showing the match
between the letters t h and the phoneme /th/. In addition a frequency count is kept of how many
times this match is found in the lexicon. This lattice then represents a summary of all the matches
in the lexicon for this word. Finally, a decision function is used to find the single best sequence of
phonemes.
The decision function can take a number of forms, and finding better methods of designing
this function is the main goal of research in pronunciation by analogy. Marchand and Damper
[300] give a review of some techniques. The original approach is to find the shortest paths through](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-245-320.jpg)
![222 Chapter 8. Pronunciation
the lattice, meaning choosing the matches which have the longest individual spans. If more than
one shortest path is found, then the one with the highest frequency count is used. Other approaches
include calculating the product of arc frequencies, the standard deviation of the path structure, the
frequency of the same pronunciation, the number of different pronunciations generated, and the
minimum of arc frequencies.
8.4.5 Other data driven techniques
A number of other data driven techniques have also been used to G2P conversion. Pagel et al
[343] introduced the use of decision trees for this purpose, a choice which has been adopted as
a standard technique by many researchers [249], [198]. The decision tree works in the obvious
way, by considering each character in turn, asking questions based on the context of the character
and then outputing a single phoneme. This can be seen as a way of automatically training the
context-sensitive rules described above.
Other data driven techniques include support vector machines [114], [111], [18], transfor-
mation based learning [67], [519] and latent semantic analysis [38]. Boula de Mareüil et al [62]
describe a formal evaluation of grapheme-to-phoneme algorithms and explains the complexities
involved in ensuring that accurate and meaningful comparisons can be performed.
8.4.6 Statistical Techniques
The statistical approach has been somewhat of a latecomer to grapheme-to-phoneme conversion,
perhaps because of the success of other data driven techniques such as pronunciation by analogy
or the impression that context-sensitive rewrite rules are adequate so long as they can be automat-
ically trained, e.g. by a decision tree. In recent years however a number of approaches have been
developed which give a properly statistical approach.
As mentioned previously, Jelinek [239] proposed a system to align graphemes and phonemes
using the Baum-Welch algorithm original developed for speech recognition. His intention was to
expand the lexicon to include unknown words that would then be used in a speech recogniser.
His use of the statistical approach was limited only to the alignment step however, the actual
grapheme-to-phoneme conversion was performed by other techniques. Taylor [433] took this work
to its logical conclusion by using HMMs to perform the whole task. In this approach, states repre-
sent the phonemes, each of which has a discrete output probability distribution over the characters
which form the observations. The system therefore mimics a “writing” process where the under-
lying form (phonemes) generates the surface representation (characters). G2P is then the opposite
process of recovering the phonemes from the characters. If performed with a standard HMM, the
results are quite poor but with various refinements the technique can be made to perform as well as
any other. These include using context in the models, using an n-gram over the phoneme sequence
and by performing pre-processing steps to make the alignment task easier. The main advantage
of this technique is that no explicit alignment is required; the system implicitly generates a “soft”](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-246-320.jpg)
![Section 8.5. Further Issues 223
alignment during training. The use of the n-gram phoneme constraint also means that even if the
answer is wrong, the generated phoneme sequence should be a likely one, and not simply a random
sequence that is un pronounceable.
A more common approach is to use technique proposed by Galescu and Allen where a search
is performed using a joint n-gram model over both the grapheme and phoneme sequences [44],
[172], [223], [86]. Typically, the alignment is performed in the normal way. This training data
then gives a collection of short sequences of phonemes and graphemes which are aligned. For
example, the word PHONEME gives { ph, f }, { o, ow }, { n, n }, { e, iy }, { m e, m }. Once the
entire training lexicon has been converted to this format an n-gram is calculated over all sequences.
At run time, a simple search gives the most likely sequence given the characters. The technique
is completely general and allows for a variety of refinements in terms of n-gram smoothing and
estimation techniques. It avoids the pre-processing steps of the HMM technique described above
but does does however require a separate alignment step.
8.5 FURTHER ISSUES
8.5.1 Morphology
It is common in linguistics to say that words are not atomic entities with regard to syntax or
semantics but rather are composed of morphemes. If we consider the words BIG, BIGGER and
BIGGEST we see there is a clear connection between the three words. The issue that concerns us
is whether we should consider all of these as separate words, and store them in the lexicon, or just
store the root form BIG and then use rules to derive the other forms.
There is no right or wrong answer to this; the fundamental issue of one of size vs speed. If
we have plenty of storage space and require quick processing, it is better to store all words. If
storage space is limited it may be better to store only the root and generate other examples no
the fly. In modern TTS systems, the first approach is generally taken as memory is cheap, and
even for large lexicons, the size of these is swamped by the size of the unit selection database. It
is traditional to defined two types of morphology; inflectional, where (in English) we typically
add well defined suffixes to words to make other words in the same class (as above). This is in
contrast to derivational where we make new words that often have different parts of speech or
different meanings, such that we form CREATION from CREATE. While in general it is possible
to store all common derived and inflected forms of words, the set of possible words is very large
indeed; for example we can attach the suffix -ESQUE to virtually and noun or proper noun (e.g.
TAYLORESQUE).
We can deal with this in a number of ways, but the most common approach is to store all
roots and common inflections and derivations in the lexicon. When other inflections or derivations
are encountered, these are treated as unknown words and passed to the grapheme-to-phoneme
convertor. This then can attempt a basic morphological decomposition which attempts to find](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-247-320.jpg)
![224 Chapter 8. Pronunciation
known morphemes within the word. Either this succeeds for all parts of the word, in which case
the pronunciation can be formulated by using these parts, or only some of the morphemes are
identified, in which case a mixture of a lookup table for the known morphemes and a grapheme-
to-phoneme conversion for the unknown morphemes can be used. The difference may be slight
for English, but this approach is certainly more attractive for languages with rich morphology such
as German [370], [460], [483].
8.5.2 Language origin and names
A further clue to how to pronounce a word is given by the origin of the word. While we know that
late is pronounced /l ey t/, and that the final e is not pronounced; we also know that latte is
pronounced /l ah t ey/ with a pronounced final e. The reason is of course that LATTE is an Italian
word and so receives an “Italian” pronunciation. We should note however that the word is not
pronounced the way an Italian person would say it, but rather an English pronunciation using an
approximation of the Italian phoneme sequence. Quite how “natively” words such as this should
be spoken is by and large a cultural issue, many native English speakers pronounce the final /t/
in RESTAURANT many do not. Given, there have been approaches which have used analysis of
character sequences to find the language of origin, typically using n-grams trained on separate
lexicons from each language [94], [289].
A final point concerns the types of words that are most likely to be used by grapheme-
to-phoneme algorithms. The approach taken by most machine learning research is to arbitrarily
divide the lexicon into training and test data, for example by assigning every tenth word to the test
set. When using G2P algorithms for lexicon compression purposes this is a reasonable thing to do,
but when using the G2P algorithm for rare words that are unknown to the system, this becomes
less valid. In a system with a large, well developed lexicon, the words most likely to be processed
by a G2P convertor are vary rare, new formations (e.g. email) loan words or proper names. The
chances are that G2P algorithms will fare worse on these than normal words, so care should be
taken in assuming results taken from a normal word test set apply to the words we are likely to
find in real usage. Taking this into account, a better approach would be to design or train G2P
systems on just those words which we are most likely to see. For instance Spiegel [409] describes
a system which uses an extensive set of rules to model proper names. The system includes all
the types of names found in the United States and is designed primarily for directory enquiry and
other telephony applications.
8.5.3 Post-lexical processing
Some systems use the output from the lexical lookup process and/or grapheme-to-phoneme con-
version directly as the input to the synthesizer. In some systems however, an additional step often
known as post-lexical processing is performed which modifies the output in some way. The
most significant factor in determining how much post lexical processing is the specific form of](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-248-320.jpg)
![228 Chapter 9. Synthesis of Prosody
of F0 at phrase beginnings. Finally even if intonation is not used for any affective or augmentative
purpose, all utterances have characteristic intonation patterns which must be modelled correctly if
the speech is to sound natural.
9.1.1 F0 and pitch
In this book we have used the terms F0 (fundamental frequency) and pitch interchangeably, as
there are few cases where the difference matters. Strictly speaking, pitch is what is perceived,
such that some errors or non-linearities of perception may lead this to be slightly different to
fundamental frequency. Fundamental frequency is a little harder to define; in a true periodic
signal this is simply defined as the reciprocal of the period, but as speech is never purely periodic
this definition itself does not suffice. An alternative definition is that it is the input or driving,
frequency of the vocal folds.
In prosody, F0 is seen as the direct expression of intonation; and often intonation is defined
as the linguistic use of F0. The relationship between the two is a little more subtle than this though
as it is clear that listeners do not perceive the F0 contour directly, but rather a processed version of
this. The exact mechanism is not known, but it is as if the listener interpolates the contour through
the unvoiced regions so as to produce a continuous, unbroken contour.
9.1.2 Intonational form
There is probably no area of linguistics were there is less agreement than when it comes to the
nature of intonational form. In the following sections we will cover many of the well known
models and theories, and it will be clear that there is enormous variation in how researchers have
chosen to represent intonation. Disagreements occur in all other areas of linguistics; it is just
the degree to which the theories differ in intonation that is so noticeable. For instance, when we
consider the theories of verbal phonology described in Section 7.4, we do in fact find that there are
quite a number of approaches to the problem, but all researchers at least agree that there is such a
thing as phonology.
We can give an overview of the situation as follows. The function of intonation is to express
emotion and attitude, to augment the verbal component by firstly emphasising or marking words,
and secondly by giving phrases and utterances characteristic patterns which show their structure
and type. Although rarely expressed as such, this much is uncontentious1 At the other end of
the scale, it is widely accepted that F0 is the main acoustic manifestation of intonation. The
disagreement stems around how any intermediate intonational form should be represented. In
some early studies, researchers did actually try to map directly from intonational semantics to
acoustics [107], [287], but these studies were criticised by many, summarised clearly by Ladd
[269]. Basically, in measuring say the F0 of anger, we must of course do this with reference
1 Any disagreement would be based around the notion of whether there was a pure linguistic intonational core, and a
separate para-linguistic or extra-linguistic component.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-252-320.jpg)
![Section 9.1. Intonation Overview 229
to particular speakers and if we did this, the difference in basic physiological behaviour between
speakers might well swamp any subtleties in the effects of anger itself. Put more plainly, regardless
of how high or low a speakers voice, a woman, child or man plainly exhibits clear regularities in
how these effects are expressed, and we are missing something if we state that the mean value of
anger is say 150Hz. This leads us to the conclusion that there must be some sort of intonational
form that speakers of a language/accent share, and which is readily identifiable to those speakers.
This is obviously somewhat similar to the idea of the form of words and phonemes in verbal
language, and often just this parallel is drawn such that we are trying to find the intonational
equivalents of notions such as word, phoneme, phone and allophone.
The basic problem is that while nearly everyone agrees that some representation of prosodic
form must exist, we have no equivalent test to the minimal pair test for phonemes. Recall from
Section 7.3.2, that this said that if two sounds could be found that distinguished two words, then
we could call these phonemes. If however we had two sounds which clearly sounded different (say
dark and light [l] in English), but which did not distinguish words, then these were not phonemes
but allophones. Crucially, this test lies on the extremely convenient ability of listeners to tell
whether two words are the same or not - this can be done without even knowing what the words
mean. In intonation, and other aspects of prosody, we simply do not have such a test, and the
“semantic space” while perhaps simpler overall, doesn’t arrange itself into nice discrete categories
the way the words in a lexicon do.
We should realise from the start however that not every model was designed with the same
goals in mind; for instance, the British school (explained in Section 9.3.1) evolved from old
style prescriptive linguistics, and even quite recent publications were concerned with the teaching
of “correct” intonation patterns to non-native learners of English. The aims of autosegmental-
metrical (AM) (Section 9.3.3) school were to provide a theory of intonation that worked cross
linguistically such that phenomena from African and Asian tone languages, pitch accent lan-
guages such as Swedish, and intonation languages such as English could all be described with a
unified model. An aim of the Fujisaki model (Section 9.3.5) was to create a model that followed
known biological production mechanisms, whereas the aim of the Tilt model (Section 9.3.6) was
to create a model solely for engineering purposes.
The models differ significantly in what they take as the primary form of intonation. In the
AM model this is quite abstract while in the Tilt model this is quite literal or “acoustic”. These
differences in primary form should not be taken to mean that the proponents of these models do
not believe that there should be more abstract or more concrete representations, just that the “best”
representation happens to lie where they describe it. In the many synthesis schemes based on the
AM model there are other, more phonetic or acoustic levels, and in the Tilt model there is always
the intention that it should serve as the phonetic description of some more abstract higher level
representation.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-253-320.jpg)
![Section 9.2. Intonational Behaviour 231
Y(P1,P2,...PN) → F01,F02,...F0M,
These models usually work on a principle whereby they encode or stylize the real F0 con-
tour. In other words, they give a tractable simplification of the F0 contour in which the salient
linguistically important aspects of the contour are preserved, while less necessary parts are dis-
carded. At its simplest, the parameters would be a discrete list of target points in the contour.
From these, a continuous F0 contour can be synthesized. If we keep the number of points constant
for each linguistic unit, then we have a situation where the parameters of the model align nicely
with the linguistic description, while hopefully still allowing accurate synthesis of the contour.
The values of these parameters can then be learned by a machine learning algorithm or written by
hand, as we now have a much simpler situation where we can align input and output pairs. The
models described below generally take a more sophisticated approach, but in essence the principle
is the same: to make the mapping between the discrete prosodic form units and the F0 contour
more tractable.
9.1.4 Micro-prosody
Micro-prosody is a term used to denote some of the lowest level, more detailed aspects of prosody.
Often this includes the interaction between verbal effects and prosodic effects on the main acoustic
components of prosody such as F0. It has been widely observed [405], [270] that the phones of an
utterance have an affect on the F0. While these effects are less noticeable than the effect of prosody
proper, they need to be handled if a complete model of F0 synthesis or analysis is to be performed.
Typical effects include a sudden raising or lowering of F0 at a boundary between a voiced and
unvoiced section of speech. The degree to which micro-prosody needs to be synthesized is debat-
able. Most intonation algorithms synthesise smooth contours and simply ignore micro-prosody
altogether and unit selection systems have the fine micro-prosodic detail already included in the
units they use. If however any automatic analysis of F0 is to be performed, the micro-prosody
must be handled. One common approach is simply to use an algorithm to smooth the F0 contour,
in an attempt to make it appear more like the stylised intonation contour that we believe represents
the underlying intonation pattern. Modern statistical techniques can handle miro-prosody more
directly, by considering it as noise in the data.
9.2 INTONATIONAL BEHAVIOUR
In this section we describe some basic intonational phenomena and attempt to do so in a theory
neutral way.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-255-320.jpg)
![232 Chapter 9. Synthesis of Prosody
9.2.1 Intonational tune
The intonation tune can be broadly described as the core pitch pattern of an utterance. Tunes differ
from one another in type and in association. By using different types of tunes, the speaker can
express for example an emotion such as surprise, disbelief or excitement. The association of the
tune connects the prosodic part of intonation (the tune itself) to the verbal part, by associating parts
of the tune with particular words. By shifting association from “john” to “match” in examples 118
and 119, one can convey different effects. By varying the type of tune, one can also express
different effects as in example 120.
(118) John went to the match (as opposed to Harry)
(119) John went to the match not the theatre
(120) John went to the match (disbelief: but he hates football !
Describing tune type is perhaps the most difficult issue in prosody. Tune schemes can be
broadly divided into those which classify tunes using dynamic features (rises and falls) and those
which use static features (tones). Theories also vary in the size of the units they use. Global
descriptions make use of a few basic patterns that cover the entire phrase, atomistic theories make
use of smaller units that combine together to form larger patterns. Jones [242] is at the global
end of the scale, the British school [333], [193] uses sub-phrase units (see Section 9.3.1 , while
the AM school [352], [351] and the Dutch school [446] use units which are smaller still (Sections
9.3.2, 9.3.3.
Much of the discussion on the subject of tune centres around how to describe pitch accents.
A pitch accent is commonly manifested in the F0 contour as a (relatively) sudden excursion from
the previous contour values. This is where association comes in, as pitch accents only occur
in conjunction with prominent syllables, and in doing so attract attention to that syllable. Pitch
accents can only occur in association with prominent syllables (see Section 6.3 on prominence),
but need not occur on all prominent syllables.
Most work agrees that the prosodic phrase as described in Section 6.2 is the basic domain of
intonational tune patterns. Hence the phrase not only groups words together, it also serves as the
start and end of each section of tune. In many models, we have a nuclear accent, which occurs
once per phrase.
The other main area of interest in tune description concerns what happens at the ends of
intonation phrases. Often F0 is low at a phrase boundary, but in many circumstances F0 is high.
For instance, if another phrase directly follows the current one, a continuation rise may be present.
If the tune is that of a yes/no question, the final pitch may also be high. The British school deals
with these effects by using different nuclear accent and tail configurations. The AM model makes
use of high and low boundary tones which distinguish the different types of contour.
It would be impossible to show all the possible types of intonational tune for English, but six
common tunes which vary because of their nuclear accent types are shown in Figures 9.1 to 9.6.
These examples are not comprehensive and other theories may classify these contours differently.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-256-320.jpg)
![Section 9.2. Intonational Behaviour 233
Fo
time
anna came with manny
*
Figure 9.1 Typical Fall Accent. “Anna came with Manny.” The nuclear fall is on the stressed
syllable of “manny” denoted by an *. The fall is a commonly found intonation accent and is often
used in neutral declarative situations.
Fo
time
anna came with manny
*
Figure 9.2 High Fall, “Anna came with Manny !”. This shape corresponds to a British “high
fall”, +raised or pitch level 4. In this particular utterance there is still a single intonation phrase,
and the word “anna” also has an accent, but this accent is pre-nuclear. Some may argue that there
is no phonological distinction between fall and high fall, and that the high fall is really just an extra
prominent fall.
The examples merely demonstrate some of the intonational effects that can be produced.
9.2.2 Downdrift
It has been observed by many researchers that there is often a gradual downdrift in the value of
F0 across a phrase [446], [447], [351], [107], [285], [167]. How downdrift (often referred to as
declination) is dealt with by different theories varies widely. Ladd [267] gives a review of some
of the different theories.
Many treat downdrift as an automatic physiological effect arising from changes in sub-glottal
pressure during the course of an utterance [287], [107]. This account gives the speaker little
conscious control over declination. The approach of the Dutch School [446], has been to use three](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-257-320.jpg)
![234 Chapter 9. Synthesis of Prosody
Fo
time
it was amazing
*
Figure 9.3 Rise fall accent, “It was Amazing !”. Spoken with a sense of wonderment, this accent
is similar to a fall, but with a much larger preceding rise. The peak value of the F0 contour is also
later than with a simple fall accent.
Fo
time
*
with anna
manny come
did
Figure 9.4 Fall rise accent, “Did Manny come with Anna ?” A peak in the F0 contour occurs in
the stressed syllable of “manny” (*). After coming down from the peak, the contour rises slowly and
finishes with a sharp rise at the end of the phrase. This type of accent is often used for simple yes/no
questions.
parallel declination lines, which refer to a baseline, a mid-line and a line corresponding to the top
of the speaker’s normal range. The contour must follow one of these lines or be rising or falling
between them. Fujisaki’s model is more flexible in that the rate of declination and initial starting
value can be varied, but the overall effect is still automatic [167]. Liberman and Pierrehumbert
[285] show that the final F0 value for utterances is invariant under a wide range of utterance
lengths and pitch ranges which is inconsistent with the view that declination slope is constant.
They propose an exponential decay downdrift effect, with the additional feature of final lowering
at the end of the phrase. Figure 9.7 shows three different views of declination.
A major claim of Pierrehumbert’s thesis [351] was that more than one factor was responsible
for the downdrift of F0 contours. As with many other theories, she proposed that the phonetic dec-
lination effect exists, but also argued that the major contribution to the downdrift of utterances was
downstep which was a phonological effect and therefore controllable by the speaker. Figure 9.8](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-258-320.jpg)
![238 Chapter 9. Synthesis of Prosody
this is only for convenience, in reality the pitch accents rarely have distinct beginnings and ends.
The most common type of pitch accent takes the form of a peak in the F0 contour. The
actual shape of this map be a rise followed by a fall, or a rise followed by a relatively level section
and then a fall. The degree of the excursion is important, and generally speaking the higher the
excursion the more prominent the pitch accent will sound. There are many other types of pitch
accent also. One common type is the simple fall, where the contour is level, falls over a short
region and then continues at a reasonably level rate. Another important type is the low pitch
accent, or rising pitch accent. The local shape of this is a little harder to define, but basically this
type of accent is marked by a negative excursion in the F0 contour. This can take the shape of a
fall followed by a rise, or just a specific low point in the contour. This is often associated with
question intonation, or cases where the speaker is doubting the listener and so on. Beyond these,
theories differ as the range of additional types. The important point about pitch accent shape is
that too often these are thought of as simple peaks in the contour; falls and low pitch accents are
just as important.
The pitch accent’s setting in the overall contour is also vital. This is governed by two factors;
its height, meaning how high or low it occurs in the speaker’s range and its alignment with the
phonetic content. In terms of height, we find a few characteristic patterns. In general, pitch
accents height decreases as a function of time throughout the utterance, as we just mentioned on
the section about downdrift. We also find that as with the size of the excursion, the accent height
is correlated with the degree of prominence (in fact in many models, accent height and excursion
size are governed by the same parameters).
The issue of alignment is more subtle. While intonation does has a form of its own, its
interaction with the verbal phonetic content of the utterance is important to how it is produced and
perceived. We have already seen that association is a key link between the prosodic and verbal
component as it determines which syllable or which word is accented. In addition to this, the
precise position of the accent in time with respect to the accented syllable is very important. Some
models [260], [259], [258], talk of early accents and late accents, and some argue that exactly
the same intrinsic pitch accent shape can give rise to quite different overall perceptions if only the
alignment differs. Quite detailed studies of alignment have been conducted [465], [266] and [132],
and found that this is a complex issue, and it is not even clear just what should be aligned with
what. Some suggest that the start of the vowel in the accented syllable should act as one anchor
point, with the peak or mid point of the pitch accent acting as the other. Many other alignment
models are however possible and in general the interactions between the intonation and phonetics
are quite complicated.
The other major intonational effect is the behaviour at phrase boundaries. The F0 at these
points can be systematically raised or lowered for specific effect. The term boundary tone is
often used as the equivalent intonation unit to pitch accent. Strictly speaking, this term is only
applicable to AM models, but it has gained wide enough usage that we use it here to describe
general behaviour at boundaries regardless of whether the particular model uses tones or not.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-262-320.jpg)
![Section 9.3. Intonation Theories and Models 239
9.3 INTONATION THEORIES AND MODELS
9.3.1 Traditional models and the British school
The British School of intonation includes contributions made as far back as Palmer [344]. Other
major contributions in this school have come from O’Connor and Arnold [333], Crystal [117], and
Halliday [193]. All these variants on Palmer’s original theme use dynamic features such as rise
and fall to describe intonation.
In the account given by Crystal, the most important part of the contour is the nucleus which
is the only mandatory part of an intonation phrase. The nuclear accent can take one of several
configurations, e.g. fall, fall-rise, low rise. Other parts of the contour are termed the tail (T) which
follows the nucleus, the head (H) which starts at the first accented syllable of the intonation phrase
and continues to the nucleus, and the pre-head (P) which precedes the head. The intonation phrase
has a grammar of (P) (H) N (T), where the brackets denote optional elements.
The relationship between the form and acoustics of this school is the most loosely defined
of all the models described here; this is hardly surprising however as none of the originators of
this system had the technology to analyse F0 contours in detail. The description of form is related
to actual contour shapes that are found, but the descriptions should not be interpreted too literally.
Both “fall” and “rise-fall” accents have rises followed by falls, the difference being that the rise
in the fall accent is much smaller and earlier in the syllable than the rise in the rise-fall accent.
Halliday describes his tones using rise and fall terminology, but does away with the standard
naming system preferring simply to name his tones 1, 2, 3, 4 5.
Some more formal descriptions have been proposed for use with the British School phonol-
ogy. In particular, two models which have been designed for speech synthesis purposes are those
of Isard and Pearson [231] who use Crystal’s phonology and Vonwiller et al [479], who use Hall-
iday’s. Both these synthesis models use the full range of the British school tune descriptions and
Isard and Pearsons’s scheme is capable of variations in prominence and pitch range.
9.3.2 The Dutch school
The Dutch School [446], [447] is based on the principle of stylization of F0 contours. Stylization
in the Dutch system involves taking a F0 contour and attempting to fit a series of straight lines
as closely as possible to the original contour. This stage is useful as it reduces the amount of
data needed for further analysis: a small number of straight lines are easier to deal with than a
continually varying F0 contour. From these stylizations, a series of basic patterns can be found -
this process is called standardization.
The version of the theory presented in t’Hart and Cohen [446] describes contours in terms
of three declination lines - high, middle and low. Pitch accents are realised by rising and falling
between these declination lines. An example of a stylized and standardized contour is shown in
Figure 9.9 (from Willems [494]).](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-263-320.jpg)
![240 Chapter 9. Synthesis of Prosody
0.0 0.3 0.6 0.9 1.2 1.5
100
200
300
500
50
and the weather turned out fine
Fo (Hz)
time (seconds)
Figure 9.9 Example of a standardized contour in the Dutch system. The dotted lines denote the
three declination lines and the thicker solid line shows the path of the F0 contour. The first excursion
to the top declination line is a head accent (British school). The second accent which rises to the top
line and then falls to the baseline is a fall accent. The rise at the end is a continuation rise.
Because of the stylization process, the continuously varying nature of the F0 contour is
eliminated, and because of the standardization process, the contour description is further reduced
into a small number of units (rises, falls etc). The idea of F0 stylisation has proved popular with
many models and techniques outside the Dutch school [481], [121].
9.3.3 Autosegmental-Metrical and ToBI models
Ladd terms this group the autosegmental-metrical (AM) model of intonation. Its origins are
in Liberman’s [284] and Bruce’s [286] work, but its first full exposition was in Pierrehumbert’s
seminal thesis [351]. From that basic model, a number of alternatives and amendments have been
proposed, but the basics remained unchanged. Most significantly, the intonation part of the ToBI
scheme comes directly from Pierrehumbert’s work in this school. Before these early publications,
it was rare to associate the word “phonology” with prosody, but these early works promoted the
idea that intonation and prosody could be described and modelled with much the same technical
apparatus as normal phonology. Hence this school is often referred to as intonational phonology
[266], [269].
The AM model describes intonation as a series of high and low tones. By using a system of
diacritics which distinguish tones located on accented syllables from those occurring at boundaries
and between accents, Pierrehumbert argued that intonation could be described by patterns of two
basic tones, which she called H (high) and L (low). Pitch accents can be represented as either a
single or double tone. Every pitch accent has a starred tone (*) which signals that it is that tone
which is directly associated with the accented syllable. The possible pitch accents are H*, L*, H*
+ L, H + L*, L + H* and L* + H. At phrase boundaries, boundary tones can be found, which](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-264-320.jpg)
![Section 9.3. Intonation Theories and Models 241
are marked with a (%). Phrase tones are used to show path of the contour from the last (nuclear)
accent to the phrase boundary, marked with a (-).
Unlike the British school analysis, there is no strict division of the contour into regions
such as head and nucleus. Both nuclear and pre-nuclear accents can be any one of the six types
described above. The nucleus accent is distinguished because the phrase and boundary tones that
follow it allow a much larger inventory of intonational effects.
Each tone forms a target from which F0 contours can be realised by using interpolation
rules. As with many other theories, the AM model retains the idea of a declination baseline,
but says that the downdrift commonly observed in F0 contours is mainly due to the phonological
effect of downstep which again is controllable by the speaker. In her original work, Pierrehum-
bert proposed that the downstep effect is triggered by the speaker’s use of a sequence of H L H
tones, using evidence from African tone languages as justification (see Figure 9.8 for examples of
downstepping and non-downstepping contours).
The history of the AM model is particularly interesting in that it has a dual character of being
intended as a “pure” linguistic theory in the MIT, Chomsky/Halle phonology tradition, and as a
working model for the Bell Labs TTS system. It is important to realise that from a theoretical
linguistic point of view, the model as just described is not intended to be a “complete” model of
intonation as used in human communication. Rather, the AM school divides the general intonation
system into a part which is properly linguistic, a part which is para-linguistic and a part that is
extra-linguistic. The distinction is seen as important within this school as the properly linguistic
part, i.e. that described above, is the main area of interest as it forms part of the general grammar of
a speaker’s language faculty. For a complete description however, we require the para-linguistic
effects too, which include for instance the degree of prominence of a pitch accent, and extra-
linguistic effects such as global F0 settings. For our purposes, the distinctions between these
subsystems are less important, but it is important to realise that the system of H* + L accents and
so on is only a partial representation of an utterance’s intonation.
9.3.4 The INTSINT Model
The INTSINT model of Hirst et al [212], [213], [211], was developed in an attempt to provide
a comprehensive and multi-lingual transcription system for intonation. The model can be seen
as “theory-neutral” in that it was designed to transcribe the intonation of utterances as a way of
annotating databases and thereby providing the raw data upon which intonational theories could be
developed. Hirst has described the development of INTSINT as an attempt to design an equivalent
of IPA for intonation. As stated in Section 6.10, there is no such thing as completely theory neutral
model as all models make some assumptions. Nevertheless, INTSINT certainly fulfills its main
goals of allowing a phonetic transcription of an utterance to be made without necessarily deciding
which theory or model of intonation will be subsequently used.
INTSINT describes an utterance’s intonation by a sequence of labels each of which repre-
sents a target point. These target points are defined either by reference to the speaker’s pitch range,](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-265-320.jpg)
![242 Chapter 9. Synthesis of Prosody
in which case they are marked Top (T), Mid (M) or Bottom (B), or by reference to the previous tar-
get point, in which case they are marked Higher (H), Same (S) or Lower (L). In addition, accents
can be marked as Upstepped (S) or Downstepped (D).
Hirst [212] describes this system in detail and shows how it can be applied to all the major
languages. Several algorithms have also been developed for extracting the labels automatically
from the acoustics and for synthesizing F0 contours from the labels2. Applications of this model
to synthesis include Veronis et al [475].
9.3.5 The Fujisaki model and Superimpositional Models
Fujisaki’s intonation model [166] takes a quite different approach to the models previously dis-
cussed in that it aims for an accurate description of the F0 contour which simulates the human
production mechanism. Fujisaki’s model was developed from the filter method first proposed by
Öhman [335].
In the model, intonation is composed of two types of components, the phrase and the accent.
The input to the model is in the form of impulses, used to produce phrase shapes, and step functions
which produce accent shapes.
This mechanism consists of two second order critically damped FIR filters (these are intro-
duced fully in Section 10.4). One filter is used for the phrase component, the other for the accent
component. The F0 contour can be represented by equations 9.1, 9.2 and 9.3.
lnF0(t) = lnFmin +
I
∑
i=1
Api Gpi (t −T0i)+
J
∑
j=1
Aaj (Gaj (t −T1j )−Gaj (t −T2j )) (9.1)
where
Gpi (t) =
α2
i te−αit for t ≥ 0
0 for t 0
(9.2)
and
Gaj (t) =
min[1−(1+βjt)e−βjt,θ] for t ≥ 0
0 for t 0
(9.3)
2 In addition to being an expert on multi-lingual intonation, Hirst is also know for being able to do a particularly fine
rendition of “Don’t Cry For Me Argentina”.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-266-320.jpg)
![244 Chapter 9. Synthesis of Prosody
time
Fo
time
Amplitude
(b)
(a)
Figure 9.11 Three accents of differing duration. Graph (a) shows the path of the F0 contour, graph
(b) shows the input step function for the second accent. As the accent duration increases, the accent
becomes less like a peak and develops a flat top.
much higher than α, which gives the filter a quicker response time. This means the shape produced
from the accent component reaches its maximum value and falls off much more quickly than
the phrase component. Phrases are initiated by impulses, which have the effect of resetting the
baseline when fed through the filter. Figure 9.10 shows the F0 contour produced by the phrase
component.
A number of amendments have been proposed to the Fujisaki model aimed at giving it more
flexibility for modelling languages other than Japanese for which it was originally designed. For
instance, Van Santen et al [465], [467], [411] have proposed a model which allows more than
one type of phrase contour, and allows the modelling of micro-prosodic effects as well as detailed
alignment of the contour to the phones. The principle of superimposition is attractive in itself
as it seems intuitive to model different parts or functions of the F0 contour separately and then
combine these to produce the final contour for the utterance. This has lead to other types of
superimpositional models, for instance the superposition of functional contours model of Bailly
and Holm [31], discussed in detail in Section 9.6.4.
9.3.6 The Tilt model
The Tilt intonation model developed by Taylor [434] [439] [436] was developed with the explicit
intent of creating a practical, engineering model of intonation. To this extent, issues of purely
linguistic concern (such as phonological rules in the AM model) or biological plausibility, as in the](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-268-320.jpg)
![246 Chapter 9. Synthesis of Prosody
In the Tilt model, there are two types of event accent and boundary (these are effectively the
same as the pitch accents and boundary tones in the AM model). Each event comprises two parts,
a rise and a fall, as shown in Figure 9.12. Between events straight lines called connections are
used. All variation in accent and boundary shape is governed by firstly changing the relative sizes
of the rise and fall components, and secondly by how the event aligns with the verbal component
(basically, whether it is higher or lower in the pitch range, or occurs earlier or later with respect
to the accented syllable). This then gives a total of six parameters, 4 describing the shape of
the event and 2 describing how it aligns. This basic model (called the rise/fall/connection model
[439]) models contours accurately but is not particularly amenable to linguistic interpretation. The
Tilt model itself is a transformation of the 4 parameters for each event (rise and fall amplitude, and
rise and fall duration) into the 3 Tilt parameters. Amplitude and duration are given by
Amplitude = A = |Arise|+|Afall|
Duration = D = |Drise|+|Dfall|
The Tilt parameter itself is used to define the general shape of the event, independent of its am-
plitude, duration or alignment. To find the value of Tilt, we first calculate an interim value, Tilt
amplitude, found by taking the ratio of the differences of the rise and fall amplitudes divide by
the sum of the rise and fall amplitudes:
tiltamp =
|Arise|−|Afall|
|Arise|+|Afall|
and likewise for duration
tiltdur =
|Drise|−|Dfall|
|Drise|+|Dfall|
these two quantities are highly correlated, which allows us to combine them into the single final
Tilt parameter:
tilt =
tiltamp +tiltdur
2
Figure 9.13 shows a number of event shapes for different values of the Tilt parameter.
9.3.7 Comparison
Many other models have been proposed in addition to the above, but the models described are
perhaps the most popular and at least cover the range of models available. An entire book could
be devoted to discussing the philosophical, historical, theoretical and practical issues surrounding
why one model has the features it does and why it may or may not be better than another model.
For now, however we will simply describe some of the differences in the model to help better
understand why these differences occur.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-270-320.jpg)
![248 Chapter 9. Synthesis of Prosody
tonational units. Often a grammar is used which states which units can follow each other, but so
long as the sequence lies within the grammar, any unit can follow any other.
Being brief, the main reason for supporting the superimpositional model is that F0 contours
do seem to exhibit patterns of global behaviour, where phrases define particular F0 patterns. The
main reason for supporting linear models comes from speech production concerns. Here, we see
that if we support a superimpositional model then the speaker has to pre-plan the utterance some
way ahead of time. This doesn’t seem to correspond to known facts in speech production, where
we know for example that a speaker can change the course of the F0 contour at any time.
Finally we should note that some papers have shown how labels in one model can be trans-
lated to labels in another [436], [374].
9.4 INTONATION SYNTHESIS WITH AM MODELS
We now turn our attention to this issue of intonation synthesis itself. In Chapter 6 we described a
variety of techniques for generating the prosodic form representations of phrasing and prominence
from text. We only made passing reference to the issue of generating intonational form from text
as this really required a thorough discussion on the issue of intonational form first. With this job
now done, we can describe techniques for both generating intonational form descriptions from text
and for generating F0 controls from intonational form.
While it is possible to build synthesis algorithms with any of the models outlined in Section
9.3.3, we will focus our attention on the most popular techniques, which divide into those based
on AM models (described here) and those based on acoustic models (described next).
9.4.1 Prediction of AM labels from text
We approach the task of predicting intonation from text in much the same way as we predicted
phrasing and prominence from text. With the AM model, our job is to assign an intonational label
to each syllable, and no matter what algorithm is used, this is formulated as a mapping from a
set of features describing the syllable and its context to a label. A typical approach of this type
is given by Lee [280] who trained a decision tree to generate ToBI labels using linguistic features
such as position in phrase, part of speech tags and syntactic information. In general though, any
of the standard text prediction machine learning techniques introduced in Chapter 5 can be used
on this task, and the approach is by and large the same as for prominence prediction and phrasing
prediction. Many other approaches for the prediction of AM and other intonation representations
from text have been proposed, including HMMs [325], decision trees/CART [513], [378] and rules
[462], [365].](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-272-320.jpg)
![Section 9.4. Intonation Synthesis with AM models 249
9.4.2 Deterministic synthesis methods
Given a representation of intonational form predicted from text, we next attempt to synthesize
an F0 contour from this. In this section we will examine two rule based deterministic proposals
for doing this, the first developed for use the with original Pierrehumbert model and the second,
developed more specifically for use with ToBI.
Anderson, Pierrehumbert and Liberman [16] proposed a synthesis scheme (hereafter called
APL) which generated an F0 contour from an arbitrary AM tone description. Recall that a feature
of the AM model is that it is linear in the sense that there is no global pattern to intonation,
each intonational unit in the sequence operates largely independently. This is reflected in the
synthesis algorithm where an acoustic model made of target points is used. In the algorithm, each
intonational event generates a set of target points which define the basic shape of the event. Pitch
range in the model is expressed by two lines, the reference line which expresses the default or “at
rest” value of F0, and the baseline which is the minimum F0 value, and the value that F0 falls to
at the end of a phrase which is terminated by an L- L% boundary tone. Declination as a global
trend is absent from the model; rather, downdrift is synthesized by a trend where each successive
pitch accent is in general placed lower in the pitch range. Between pitch accents, phrase accents
and boundary tones, linear interpolation is used to “fill in” the F0 values. An important nuance
in the algorithm is its treatment of phrase final behaviour. Following extensive studies reported in
previous work [285], Anderson et al added a component to the algorithm which dropped the F0
contour to the baseline after the nuclear accent, regardless of whether the model itself would have
predicted such a value. This greatly increases the naturalness of the contours produced as they
have a proper impression of finality compared to other synthesis algorithms they investigated.
A more recent algorithm designed to work with the ToBI scheme was proposed by Jilka,
Mohler and Dogil [241]. This in many ways extends the original work by Anderson et al as it too
synthesises the ToBI events as isolated units, places them in context in the utterance, and then uses
interpolation to cover the regions between. The algorithm improves on the original work in that it
pays particular attention to where target points should be placed with respect to their syllables, and
for instance models in some detail the fact that accents that occur early in phrases are often later
with respect to their syllable position than accents which occur later in phrases (its as if the phrase
boundaries are repelling the accents in both directions). In this algorithm, a topline and baseline
are used as the reference points. When downstepping occurs, it triggers a lowering in the topline,
and this is the general mechanism by which downdrift occurs. Example rules for the position of
the target points are given below:](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-273-320.jpg)
![250 Chapter 9. Synthesis of Prosody
Intonational event context resultant time position
H* First syllable in phrase 85%
Last syllable in phrase 25%
One syllable phrase 50%
All other cases 60%
L* First syllable in phrase 85%
Last syllable in phrase 25%
One syllable phrase 50%
All other cases 60%
L+H* (H part) First syllable in phrase 90%
Last syllable in phrase 25%
One syllable phrase 70%
All other cases 750%
L+H* (L part) Normal case 0.2s before H*
If voiceless region 0.2s before H* (reference point) 90 %of voiced region to left of referenc
9.4.3 Data Driven synthesis methods
Black and Hunt [51] described a data-driven technique for synthesising F0 contours from ToBI
labels. In fact, their work can be seen as a data-driven version of the APL algorithm, in which the
target points are learned from the data rather than specified by hand. The approach was based on
learning the mapping between ToBI labels, which had been marked on a database by hand, and F0
contours extracted automatically. The APL model defines target points at very specific locations,
but these are difficult to determine bottom-up so the Black and Hunt approach instead uses three
target points for each ToBI label. These are place the beginning, middle and end of the syllable
associated with the label. This gives us a representation where the input is a ToBI label and a set
of features describing the linguistic context of the label, and the output is three target points. In
this form, the mapping is now amenable for use by a standard machine learning algorithm.
Black and Hunt in fact used a linear regression technique, with features such as lexical stress,
numbers of syllables between the current syllable and the end of the phrase, identity of the previous
labels and so on. Once learned, the system is capable of generating a basic set of target points for
any input, which we then interpolated and smoothed to produce the final F0 contour. Other data
driven techniques such as CART have proven suitable for synthesizing from AM representations
[292], [340].
9.4.4 Analysis with Autosegmental models
The biggest difficulty with the AM models is the difficulty in labelling corpora. Several studies
have been conducted in labelling with ToBI with the general conclusion that while labellers can
often identify which syllables bear pitch accents, they are very poor at agreeing on which particular](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-274-320.jpg)
![Section 9.5. Intonation Synthesis with Deterministic Acoustic Models 251
accent is correct. As we explained in Section 6.10, the inter-labeller agreement with models such
as ToBI is worryingly low, and in many practical database labelling projects many of the pitch
accent distinctions have been dropped [488], [487].
Two major problems stem from this. Firstly, any database which has been labelled with ToBI
will have a significant amount of noise associated with the pitch accent label classes. Secondly,
for any large scale machine learning or data driven approach, we need a considerable amount of
labelled data to the extent that it is impractical to label data by hand. As we shall see in Chapters
15 and 16, virtually all other aspects of a modern data driven TTS system’s data are labelled
automatically, and so it is a significant drawback if the intonation component can not be labelled
automatically as well. Because however the level of human labeller agreement is so low, it is very
hard to train a system successfully on these labels; we can hardly expect an automatic algorithm
to perform better than a human at such a task.
One solution that is increasingly adopted is to forgo the distinction between label types
altogether, see for instance [488]. While the break index and boundary tone components are often
kept, only a single type of pitch accent is used; in effect the labellers are marking whether a word
is intonationally prominent or not. However, it should be clear that such an approach effectively
reduces ToBI to a data driven system of the type described below.
9.5 INTONATION SYNTHESIS WITH DETERMINISTIC ACOUSTIC MOD-
ELS
Here we describe how to synthesise F0 contours using the Fujisaki and Tilt models and as we
will see, the approach is similar in both cases. Synthesis from the Fujisaki or Tilt parameters is
of course given by the models themselves, and so presents no problems. For TTS purposes then,
the issue is to generate the Fujisaki or Tilt parameters from the text. Two basic approaches can be
taken in this regard; either we try and generate the parameters directly from the text, or we take
a two stage approach where we first generate an abstract, discrete prosodic form representation
from the text, and then proceed to generate the model parameters from these.
9.5.1 Synthesis with superimpositional models
The Fujisaki model is most commonly used with Japanese, but has been used or adapted to many
other languages. In Japanese, we find that the range of pitch accent phenomena is narrower than
in languages such as English, which means that the model’s single type of accent is particularly
suited. In addition, the nature of intonation in Japanese means that accents are marked in the
lexicon, which greatly simplifies the problem of prominence prediction. Hence a simple approach
to this, which uses accent information from the lexicon alone, is often sufficient. A common
approach therefore is to determine phrase breaks and prominent syllables from the text, and then
phrase by phrase and syllable by syllable generate the input command parameters for the Fujisaki](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-275-320.jpg)
![252 Chapter 9. Synthesis of Prosody
model using one of the standard machine learning techniques.
Hirose et al [206], [205] describes such an approach for Japanese. In doing so, he makes use
of the basic lexicon tone distinctions in Japanese to given where accents occur, thus simplifying
the prominence and accent prediction process considerably. To generate the accent commands, he
uses parts of speech, syllable structure, relative positions of these, numbers of proceeding phrases
and punctuation. Syntactic units are also optionally used. A number of neural network algorithms
are used, including recurrent networks. In addition, multiple regression analysis in a vein similar
to the ToBI Black and Hunt [51] technique was investigated, in most cases the results were very
similar regardless of which algorithm is used (this of course follows the pattern for other prosody
prediction problems).
One significant advantage that Hirose makes note of is that even when the prediction algo-
rithm generates incorrect parameters for the Fujisaki model, the result is still a “natural” contour
because the model itself is tightly constrained. In other words, the contour produced is some con-
tour, just not the right one. By comparison, when prediction is poor with target models, the result
may be a contour that bears no similarity to the type that a human could produce.
9.5.2 Synthesis with the Tilt model
Dusterhoff et al [146] described a three stage technique for synthesising intonation with the Tilt
model. First, they predict where accents are placed; this is effectively a prominence prediction
algorithm of the type described in Section 6.8. Next, for each accent, the three intrinsic and 2
extrinsic Tilt parameters are generated. This is done with a decision tree, using features such as
syllable position in phrase, length of phrase and so on. Finally, the Tilt model itself generates
the F0 contour. In general though there are many similarities between synthesis with Tilt and the
Fujisaki model, so a technique which works with one can probably be adapted to the other.
9.5.3 Analysis with Fujisaki and Tilt models
An advantage that the Tilt model has over many others is that automatic analysis is significantly
easier using this model. Taylor [436] describes a complete technique for first finding pitch accents
and boundary tones and then determining the Tilt parameters from the local F0 contour in the
surrounding region. Many algorithms have also been developed for automatically finding the
Fujisaki model parameters from data [165], [71], [314], [7], [379].
9.6 DATA DRIVEN INTONATION MODELS
A common criticism of the Fujisaki model [285], [436] is that there are many contours which
the model can not generate. These include gradual low rising contours which occur when a low
accent occurs early in a phrase, and the effect of phrase final lowering where the contour falls past
the normal baseline to a final baseline in declarative utterances. Suggestions that fix the problems](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-276-320.jpg)
![Section 9.6. Data Driven Intonation Models 253
have been made, for instance by including phrase-final negative phrase commands and so on, but
such suggestions detract from the elegance of the original model. The Tilt model can be criticised
as over generating in that there are contours it can generate which are never observed. When we
consider simpler stylisation models we see that these models often massively over generate.
To solve these problems, we can of course refine the models, but this can result in endless
“tinkering” in which the models loose their originally elegance and simplicity. The basic problem
is that when we design a model, we wish it to be elegant and easy to interpret, but we always
run the risk that some intonational phenomena that we haven’t taken into account may appear and
be outside the model’s capabilities. We then face the choice of adding to the model, which often
appears as a “hack” or designing a completely new model.
We can however consider a quite a different approach where we completely abandon the idea
of creating an explicit model, but instead infer the model from data. Bailly [29] has considered
this issue and asked whether explicit intonation models are required at all; perhaps we should
concentrate on databases and labelling, and leave the specifics of the model to be determined by
modern machine learning techniques. As we shall see in Chapter 13, it is convenient to describe
synthesis techniques in terms of three generations, where the first generation used explicit hand
written rules, the second generation uses basic and quite constrained data driven approaches, and
a third generation which used general data driven and statistical approaches. The same pattern can
be seen in intonation synthesis, where for example the APL algorithm is first generation and the
Black and Hunt decision tree ToBI algorithm is second generation. The techniques described here
can be considered third generation as they match closely the approaches used in the HMM and
unit selection synthesis techniques described in Chapters 15 and 16.
In the following sections, we give an overview of the techniques used, but as many of these
follow the synthesis techniques described in Chapters 15 and 16, we wait until then to give a fully
formal account.
9.6.1 Unit selection style approaches
The unit selection synthesis technique described in Chapter 16 uses an entirely data driven ap-
proach, whereby recorded speech waveforms are cut up, rearranged and concatenated to say new
sentences. Given the success of this approach in normal synthesis, a number of researchers have
applied these algorithms to F0 synthesis [296], [310], [311].
F0 contours extracted from real utterances are by their very nature perfect. The basic idea
therefore is to collect a database of naturally occurring F0 contours, and use these at synthesis
time for new utterances. It is extremely unlikely that one of the complete contours in the database
will be exactly what we want; for this to be the case we would have have to have exactly the same
text for the synthesis and database utterance. At a smaller scale, say at the phrase, word or syllable
level, it is however possible to find exact or close matches and so the database processed to create
a set of F0 units. Each unit has a set of features describing it, and at run time, these features are
compared with features generated by the prosodic prediction. Sometimes exact matches are found,](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-277-320.jpg)
![256 Chapter 9. Synthesis of Prosody
Markov models (HMMs) as these are in general easier to train and allow more complexity with
regard to noise/covariance terms.
HMMs are however similar to dynamic system models in certain ways in that they have
a hidden state space, within which the model dynamics operate, and a projection from this to
an observation space. We have used HMMs before for various tasks including part-of-speech
tagging and phrase-break prediction. In those cases, the output distributions were discrete and
gave the probability of seeing (say) a word given a POS tag. It is also possible to have HMMs
with continuous output distributions which give the probability of seeing a continuous value given
a state. This sort of HMM can therefore be used for F0 generation. In the dynamic system model,
the hidden state space is a continuous trajectory, and this seems very natural for F0 contours.
HMMs by contrast have a discrete state space, and when we move from one state to another during
generation, the observation probabilities suddenly change. In this mode an HMM will generate
extremely unnatural contours where we have a series of constant F0 values, a jump to a different
series of constant F0 values and so on. No real F0 contours have behaviour anything like this. If
we however adopt the formulation of Tokuda et all [453], we can generate from the HMM states
while also taking into account the learned dynamics of F0 behaviour. Inanoglu [230] describes a
system where F0 for various emotions can be accurately generated by HMMs (diagrams of this
are shown in Figures 15.14 and 15.15). This method of HMM synthesis is the subject of Chapter
15 and so we leave further explanation of HMM F0 synthesis until then. We finish however with
the observation that by using this formulation, extremely natural F0 contours can successfully be
generated with HMMs.
9.6.4 Functional models
Given the considerable difficulties in constructing a model of prosodic form, never mind any
difficulties in using such a model, we can ask whether an explicit model of prosodic form is
really necessary. While experience has shown that some notion of prosodic form is necessary
(see Section 9.1.2), it may not in fact be necessary to model this explicitly; rather we can learn a
model of prosodic form and keep it as a hidden part of a more general model [501], [502]. The
superposition of functional contours (SFC) [216], [321], [217], [31] is a model in which the
mapping from the meaning and linguistic structure to F0 is attempted in a unified fashion. The
SFC model is based on the idea that there are a number of identifiable prosodic functions3 and
that each has a characteristic prosodic pattern. As the scope of each function is variable, and as we
can have several independent functions operating on a particular part of the utterance (for instance
syntax and emotion), then we need to combine these effects together to form a single prosodic
pattern.
In the SFC model, the syllable is used as the unit of prosodic control and each is described
by three F0 values and a duration value representing the prosodic lengthening to be applied. Each
3 Here the word “function” means “purpose” and is a combination of affective and augmentative prosody.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-280-320.jpg)
![Section 9.7. Timing 257
prosodic function is assigned a generator, whose purposes is to generate the prosodic output for the
particular sequence of syllables given when that prosodic function is active. A training database
therefore comprises a set of utterances, each marked with the multiple prosodic functions and the
4-parameter descriptions for each syllable. Each generator is then trained by an iterative procedure
which learns globally what prosodic patterns each generator should produce. In the original paper
Bailly and Holm used neural networks as the generators, but in principle, any trainable model
can be used for this, including the dynamic system models and HMMs described above. The
advantage of this approach is that realistic and natural prosody can be generated without the need
for determining which syllables receive pitch accents and what type they should be. All this is
modelled and optimised implicitly in the model.
9.7 TIMING
Many aspects of intonation, prominence and phrasing are manifested in differences in the timing
of speech. It is therefore the task of the timing module in a TTS system to assign a temporal
structure to the utterance. The first question then to ask is how exactly should we specify these
timing values? Most commonly this is formulated in terms of unit durations, although other
models assign timing values to significant points in the utterance [35]. The question then is, what
type of units should we assign these durations to? It is well known from experimental phonetics
that the temporal properties of speech are complex. If one talks more quickly, it is not the case that
every part of the utterance contracts by a constant factor. Emphasis generally lengthens sections
of speech, but in very particular ways. At a micro-level, we know that the articulators all have
specific dynamic properties and when a phone is spoken at a faster rate, some articulators may
undershoot their target more than others.
A completely comprehensive approach to timing would therefore attempt to model all these
factors, but in general it is very hard to determine any reliable anchor points at a level lower than
the phone. In practical TTS, the complex sub-phone timing patterns are often ignored as firstly
it is difficult to predict the interactions and secondly it is subsequently difficult to know what to
do with the sub-phone information once obtained. In practice then the most common approaches
are to predict syllable durations or phone durations. Syllables are attractive as they seem to be
the natural units of prosody; we have seen elsewhere that they are the units of prominence and
therefore pitch accent placement. In addition, we usually think of the syllable as being the natural
unit of rhythm. Phones are attractive as they are the smallest units we can predict and therefore
allow the most detail in prediction.
9.7.1 Formulation of the timing problem
Unlike intonation, timing synthesis can timing synthesis can in fact be formulated as a regression
algorithm directly as our goal is to predict a single continuous value from a set of features. Re-](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-281-320.jpg)
![Section 9.7. Timing 259
9.7.3 Klatt rules
One of the most widely set of deterministic rules was developed by Dennis Klatt [253], [10].
Although designed for use with the MITalk formant synthesiser, these rules can be used with any
synthesis technique.
In this, the basic properties of each type of phone are given by two values, a minimum
duration M and an inherent duration . The inherent duration is a sort of average, but is best
thought of as the duration of the phone spoken in a neutral canonical context. The final is then
given by
Duration = [(inherent duration −minimum duration)×A]+minimum duration (9.8)
The term A is calculated by the successive application of a set of rules, where each firing rule
multiples A by a factor. The factors are calculated as follows:
clause-final lengthening If a segment is the vowel or in the coda of a clause-final syllable, then
A = 1.4A.
non-phrase-final shortening If a segment is not in a phrase final syllable A = 0.6A. If the seg-
ment is a phrase-final postvocalic liquid or nasal A = 1.4A.
non-word-final shortening If a segment is not in a word final syllable A = 0.85A.
Poly-syllabic shortening If a vowels is in polysyllabic words, A = 0.80A.
Non-initial-consonant shortening If a consonants is in non-word-initial position, A = 0.85A.
Unstressed shortening Unstressed segments are half again more compressible; so that minimum
duration = minimum duration/2. If the segment is unstressed or reduced, set A as follows:
• vowel in word medial syllable: A = 0.5A
• other vowels: A = 0.7A
• prevocalic liquid or glide: A = 0.1A
• all others: A = 0.7A
Prominence if the segment is in a prominent syllable, A = 1.4A.
Post vocalic context of vowels A vowel is modified by the consonant that follows it in the fol-
lowing ways:
• no following consonants, word final: A = 1.2A
• before a voiced fricative: A = 1.6A
• before a voiced plosive: A = 1.2A
• before a nasal: A = 0.85A
• before a voiceless plosive: A = 0.7A
• before all others voiced fricative: A = A
These effects are less in non-phrase final positions, in which case set A = 0.7+(0.3A)](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-283-320.jpg)
![260 Chapter 9. Synthesis of Prosody
Shortening in clusters • vowel followed by a vowel: A = 1.2A
• vowel preceded by a vowel: A = 0.7A
• consonant surrounded by consonants: A = 0.5A
• consonant followed by a consonant: A = 0.7A
• consonant preceded by a consonant: A = 0.7A
In addition, Klatt also states that pauses should be 200ms long, and that non-reduced vowels
preceded by voiceless plosives should be lengthened by 25ms.
9.7.4 Sums of products model
Van Santen [411], [464] proposed the sums-of-products model which can be seen as a semi-
trainable generalisation of the Klatt model described above. Van Santen makes a few key obser-
vations regarding the nature of duration prediction. Firstly he notes that many possible feature
combinations will never be observed in the training data (this of course is the same sparse data
problem we have come across before). Crucially however, he notes that duration modelling can
be seen as a well behaved regression modelling problem, in which certain properties of the prob-
lem allow us to make assumptions that can help ease data sparsity. Specifically, he states that to
a large degree the factors influencing duration are monotonic (or “directionally invariant” in his
terms). In practice this means, while we know that emphasis and phrasing both affect the length
of a phone, and that while strictly speaking we can not consider these features as being indepen-
dent, their general influence on the duration is well behaved, such that emphasis always lengthens
a phone as does phrasing; it is not the case that the effects of these features ever reverses. This
means that it is possible to build a model where the interactions of these features do not have to be
comprehensive.
In general, models of this type will not work because they don’t model interactions be-
tween the features. In van Santen’s approach however, the process of building the model is to
use knowledge to design the features. Hence the features actually used in the model will often
be transformations of the ones generally available in the utterance structure. This process of hand
design helps produce features which can operate more independently than the original features.
The parameters can be set in a number of ways; how this is done is independent of the model itself.
A sampling approach is one possibility.
9.7.5 The Campbell model
Campbell [83] proposed a model that uses the syllable as the fundamental unit of duration. In this
a syllable duration is predicted from a set of linguistic features after which the individual phone
durations within the syllable are then calculated. This approach has the attraction in that it is more
modular, where we have one component modelling the prosodic part of the duration and another
modelling the phonetic part.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-284-320.jpg)
![Section 9.8. Discussion 261
The prosodic part of the Campbell model uses a neural network to calculate a syllable du-
ration. This is an attractive approach as the neural network, unlike the Klatt or sums-of-products
model, can model the interactions between features. A somewhat awkward aspect of the model
however is the fact that the syllable duration itself is of course heavily influenced by its phonetic
content. If left as is, the phonetic variance in model duration may swamp the prosodic variance
that the neural network is attempting to predict. To allow for this, Campbell also includes some
phonetic features in the model. Campbell maps from syllable to phone durations using a model
based on his elasticity hypothesis which states that each phone in a syllable expands or contracts
according to a constant factor, normalised by the variance of the phone class. This operates as
follows. A “mean” syllable is created containing the correct phones with their mean durations.
This duration is then compared to the predicted duration, and the phones are either expanded or
contracted until the two durations match. This expansion/contraction is performed with a constant
variance, meaning that if the vowel is expanded by 1.5 standard deviations of its variance, the
constant before it will be expanded by 1.5 standard deviations of its variance.
The problem of the absolute syllable duration can be easily solved by having the neural
network predict this z-score instead of an absolute duration, an approach followed by [456]. This
then frees the neural network from phonetic factors completely and allows it to use only prosodic
features as input.
In general this model is quite effective as it solves the feature explosion problem by positing
a modular approach. One significant weakness however is that the elasticity hypothesis is demon-
strably false. If it was true then we would expect the z-scores for all the phones in a syllable to be
the same, but this is hardly ever the case; depending on context, position and other features the z-
scores for phones across a syllable vary widely. This is a problem with only the second component
in the model and a more sophisticated model of syllable/phone duration interaction could solve
this. In fact, there is no reason why a second neural network could not be used for this problem.
9.7.6 Other regression techniques
As we might expect by this stage in the book, most of the usual classification/regression algo-
rithms have been applied to the duration prediction problem. These include decision trees [372],,
neural networks for phone prediction [109], [157], genetic algorithms [319] and Bayesian belief
networks [182]. Comparative studies of decision trees and neural networks found little difference
in accuracy between either approach, [473], [187], [72].
9.8 DISCUSSION
In this chapter, we have considered techniques for synthesizing the acoustics of prosody. Any
discussion on this topic will draw more or less the same conclusions as given in the discussion in
Chapter 6, where we explained that there was an enormous range of models, theories and explana-](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-285-320.jpg)
![262 Chapter 9. Synthesis of Prosody
tions for prosodic phenomena. The consequence of this is that developing engineering solutions
to practical prosodic problems was considerably more difficult than with the verbal component
of language. Hence the problem is complicated because there is no commonly agreed way of
representing the linguistic form of the phenomena we are attempting to synthesize.
While many of the AM models and deterministic acoustic models provide useful and ad-
equate representations for intonation, the trend is clearly towards the data driven techniques de-
scribed in Section 9.6. These have several advantages; besides bypassing the thorny theoretical
issues regarding the “true nature” of intonation, they have the ability to automatically analyse
databases, and in doing are also inherently robust against any noise that can occur in the data,
whether it be from errors in finding F0 values or from other sources.
The purpose of all the prosody algorithms described in this chapter is to provide part of the
specification which will act as input to the synthesizer proper. In the past, the provision of F0
and timing information was uncontested as a vital part of the synthesis specification and most
of today’s systems still use them. As we shall see in Chapters 15 and 16, however, some third
generation systems do not require any acoustic prosody specification at all, making used of higher
level prosodic representations instead. Rather than use F0 directly, stress, phrasing, and discourse
information are used. While such an approach completely bypasses all the problems described in
this chapter, it does have the consequence of increasing the dimensionality of the feature space
used in unit selection or HMM synthesis. It is therefore a practical question of tradeoff whether
such an approach is better than the traditional approach.
Most work on the acoustics of prosody has focused on the modelling of F0 and timing. This
has mainly been because these are the parts of prosody which are most readily identifiable and
separable from the influence of verbal factors. In addition, these are the aspects of speech which
signal processing algorithms find easiest to modify. It is quite clear however that prosodic factors
have a strong influence on a range of other acoustic aspects of the signal. For instance at the ends
of phrases the vocal effort is considerably reduced, and the overall voice quality is quite different
than elsewhere. To date, only a few studies have considered these effects [424], [417], [183]. If
the prosodic influence on voice quality is not modelled within a TTS system the naturalness with
definitely suffer, but it is unclear whether these affects should be modelled explicitly, or implicitly
within a unit selection or HMM synthesizer.
9.8.1 Further Reading
Articles on all the models and techniques described here are readily available, but there are few
which consider all models together and offer comparisons. The best account of the AM model
is given by Ladd [269]. Ladd not only describes the model and its development, but also serves
as a solid introduction to the whole field of intonational phonology and the practical phonetics
associated with it. Accounts of the INTSINT model are given in Hirst [212], the Tilt model in
Taylor [436], the Fujisaki model in Fujisaki [166], the SFC model in Bailly and Holm [31], the
Dutch model in [447], the British school in O’Connor and Arnold [333].](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-286-320.jpg)
![Section 10.2. Digital signals 283
10.2 DIGITAL SIGNALS
We now turn our attention to digital signals. All the signals we have discussed up until now are
analogue signals in which the amplitude and time dimensions are represented by real numbers.
By contrast, Digital signals are represented by a sequence of integers. In the “real-world”, most
signals are analogue, including speech signals, where the amplitude is a measure of sound pressure
level or some other physical quantity. Digital signals are mostly found in computers.
10.2.1 Digital waveforms
A digital signal is represented by a sequence of evenly spaced points, the value at each point being
the amplitude, represented by an integer. This integer amplitude value can be interpreted as the
digital manifestation of sound pressure level, or voltage or some other such analogue quantity.
Formally, a digital sequence is represented as follows:
x[n] = ...,x−2,x−1,x0,x1,x2,... (10.26)
For digital sequences, it is usual to use n and write x[n], to distinguish from continuous time
t in analogue signals. Each point in the sequence is called a sample, and the distance between one
sample and the next is called the sample period, Ts. The sample rate or sample frequency is given
by Fs = 1/Ts. The sample rate limits the frequencies which the digital signal can describe. The
highest frequency that a digital signal can contain is called the Nyquist frequency and is exactly
half the sample rate.
The choice of sample rate is a compromise between wanting to have the highest possible
quality (high sample rate) and the lowest storage or transmission cost (low sample rate). Practical
data from human perception studies shows that the limit of human hearing is about 20,000Hz and
this figure would lead us to use sampling rates of 40,000 Hz. Indeed, compact discs and digital
audio tapes have sampling rates of 44,100Hz and 48,000Hz respectively. However, it can be shown
that for speech signals, very little useful speech energy is found in the higher frequencies, and so
sample rates of 16,000Hz are often used and found to be perfectly adequate for accurate represen-
tation of speech signals. Where space is tight, lower sampling rates are used, most commonly in
the telephone which has a standardised sample rate of 8000Hz. While speech at this sample rate
is normally quite intelligible, it is clear that it sounds much poorer than when higher sample rates
are used.
The bit range describes the dynamic range of a digital signal, and is given by the range
of the integer used to store a sample. Widely used values are 16 bits, which gives an absolute
dynamic range of 216, meaning the amplitude of the sample can range from -32,768 to 32,767.
Dynamic range is normally expressed in the logarithmic decibel scale. In digital signals, we take
the reference as 1 and therefore the dynamic range of a 16 bit signal is:
10log(216
) ≈ 96dB (10.27)](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-307-320.jpg)
![284 Chapter 10. Signals and Filters
To save on space and transmission bandwidth, telephones and other systems frequently use 8
bits (48db) and 12 bits (72db). At 8 bits, speech quality is noticeably worse than at 16 bits. 12 bits
is a reasonable compromise, but if space is not an issue (and with most types of speech synthesis
it is not), 16 bits is normally used.
10.2.2 Digital representations
We use somewhat different notation when dealing with digital signals. First, an integer n is used
instead of t for the time axis, where n is related to t by the sample period:
n = tTs (10.28)
From this, we can relate the continuous signal x(t) and digital signal x[n] by x[n] = x(n/Ts). We
use square brackets “[ ]” rather than parentheses (round brackets) “( )” when describing digital
functions. Strictly speaking, a signal x[n] is in the n-domain but we shall informally refer to it as
being in the time domain.
Secondly, it is common to use normalised frequency, rather than natural frequency. From
10.28 we see that the relationship between frequency, angular frequency and normalise angular
frequency is
ω̂ =
ω
Ts
=
2πF
Ts
(10.29)
Using n and normalised angular frequency ω̂ makes calculations easy as the equations are
the same regardless of sample rate. To convert to real frequency, the normalised frequency is
simply multiplied by the sample rate. In general, we use the same terms for amplitude and phase
for both analogue and digital signals.
10.2.3 The discrete-time Fourier transform
For a continuous signal, we can move from the time domain to the frequency domain via the
Fourier transform, but as we are now dealing with digital signals, this formula cannot be used di-
rectly. We will now derive an equivalent transform for digital signals. Starting with the previously
defined Fourier transform (Equation 10.21)
X(jω) =
Z ∞
−∞
x(t)e− jωt
dt
we can use this on a digital sequence with the relation x[n] = x(n/Ts). The Fourier transform
is calculated as a integral over a range, which can be interpreted as finding the area under the
curve over this range. The numerical method of the Riemann sum says that we can approximate a
continuous area function by adding up the areas of separate rectangles. The area of each rectangle
is the value of the function at that point multiplied by the width of the rectangle. In our case,](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-308-320.jpg)
![Section 10.2. Digital signals 285
the width of the rectangle is the sample period Ts. This gives a numerical version of the Fourier
transform:
X(jω) =
∞
∑
−∞
(x(nTs)e− jωt
)Ts
Ts is independent of the integral, we move it outside, and use x[n] in place of x(nTs) giving:
X(jω) = Ts
∞
∑
−∞
x[n]e− jωnTs
We can get rid of the Ts by making use of normalised frequency, ω̂ = ω/Ts. The result is the
discrete-time Fourier transform or DTFT:
X(ejω̂
) =
∞
∑
−∞
x[n]e− jω̂n
(10.30)
It should be clear that the DTFT is a transform in which time is discrete but frequency is contin-
uous. As we shall see in later sections, the DTFT is a very important transform in its own right
but for computational purposes, it has two drawbacks. Firstly, the infinite sum is impractical, and
secondly we wish to obtain a representation for frequency that is useful computationally, and such
a representation must be discrete.
10.2.4 The discrete Fourier transform
For the first problem, we choose a sufficiently large interval L such that the samples outside 0
n L are zero. For the second problem, we will use a similar operation to time domain sampling
and “sample” the spectrum at discrete intervals. Just as n = t.Ts, we define the spacing between
each discrete frequency as Fs/N = 1/(NTs) where N is the total number of frequency values we
want. Each angular frequency therefore lies at
ωk =
2π
NTs
giving the transform:
X(ejω̂k
) =
L
∑
n=0
x[n]e− j ˆ
ωkn
If we set L = N (discussed below), and use the above definition of ωk, we can write 10.2.4 in a
more compact usable form:
X[k] =
N−1
∑
n=0
x[n]e− j 2π
N kn
(10.31)](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-309-320.jpg)
![286 Chapter 10. Signals and Filters
which gives us the standard definition of the Discrete Fourier Transform or DFT.
The interpretation is the same as for the Fourier transform defined in section 10.1.6; For
each required frequency value, we multiply the signal by a complex exponential waveform of
that frequency, and sum the result over the time period. The result is a complex number which
describes the magnitude and phase at that frequency. The process is repeated until the magnitude
and phase is found for every frequency.
The DFT states that the length of the analysed waveform and the number of resultant fre-
quency values are given by N. This makes intuitive sense: if we only analyse a few samples, we
can only expect a somewhat crude version of the spectrum; by analysing more samples, we have
more information “to go on” and can fill in more detail in the spectrum: that is, the points in the
spectrum (ωk) get closer and closer. As N → ∞, we analyse all the waveform values, and the dis-
tance between the frequency points in the spectrum → 0. This is exactly the case that the DTFT
describes.
Both the DTFT and DFT have inverse transforms:
x[n] =
1
2π
Z π
−π
X(ejω̂
).ejω̂n
dω̂ (10.32)
x[n] =
1
N
N−1
∑
n=0
x[n]ej 2π
N kn
(10.33)
In practice an algorithm called the Fast Fourier Transform or FFT is commonly used to
calculate the DFT. This is not another type of transform, merely a fast way of calculating a DFT:
the result for the FFT is exactly the same as for the DFT. The FFT only works for values of N
which are powers of 2, hence it is common to find N set to 256, 512. If a DFT of a different N is
required, a time signal of the next highest power of 2 is chosen, and the extra values are set to 0.
So for example, if we have a signal of length 300, we set N = 512, “pad” the extra 212 values with
0, and the calculate a 512 point FFT. The use of the FFT is pervasive in practice, and one often
hears references to “the FFT of the signal” etc. We repeat, in terms of the outcome, the FFT and
DFT are exactly the same and hence taking “the FFT of the signal” is no different from “taking
the DFT”. An overview of the FFT can be found in Oppenheim and Schafer [342].
10.2.5 The z-Transform
We now introduce another important transform, the z-transform 3. The z-transform is perhaps the
most useful of all transforms in practical digital signal processing. It is defined as:
X(z) =
∞
∑
n=−∞
x[n]z−n
(10.34)
3 what? you can’t be serious - another transform!](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-310-320.jpg)
![Section 10.2. Digital signals 287
This should look familiar - if we use z = ejω̂, we get:
X(ejω̂
) =
∞
∑
−∞
x[n]e− jω̂n
which is the discrete-time Fourier transform, Equation 10.30. Just like the DTFT, the z-transform
is a transform which operates on a digital signal but which produces a continuous function.
An important property of the z-transform is its affect on a delay. If we define a signal
y[n] = x[n−nd]
that is, a delayed version of x[n]. By use of a dummy variable m = n − nd, we can find its z-
transform as follows
Y(z) =
∞
∑
n=−∞
y[n]z−n
=
∞
∑
n=−∞
x[n−nd]z−n
=
∞
∑
m=−∞
x[m]znd −m
=
∞
∑
m=−∞
x[m]z−m
znd
The summation in m is just the z-transform of x defined in m rather, than n and so we can write:
Y(z) = X(z)znd
(10.35)
Hence delaying a signal by nd multiplies its z-transform by znd . This important property of the
z-transform and is discussed further in section 10.3.5.
Why use yet another transform? The reason is that by setting z to ejω̂ we greatly simplify
the DTFT, and this simplification allows us to perform sophisticated operations in analysis and
synthesis with relatively simple mathematical operations. The key to the ease of use of the z-
transform, is that equation 10.34, can be re-written as:
X(z) =
∞
∑
n=−∞
x[n](z−1
)n
(10.36)
That is, the transform can be written as a polynomial in z−1, so that the techniques of polynomial
solving can be brought to bear.
A final transform, the Laplace Transform, is frequently used in continuous time signal
processing, and is the continuous time version of the z-transform. The z-transform and the Laplace
transform can be thought of as generalisations of the DTFT and Fourier transform respectively.
The Laplace transform is not required in this book, as all our signal processing is performed on
digital signals. We mention it here for completeness only.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-311-320.jpg)
![Section 10.3. Properties of Transforms 289
X(jω) =
Z ∞
−∞
(αx1(t)+βx2(t))e− jωt
= α
Z ∞
−∞
x1(t)e− jωt
+β
Z ∞
−∞
x2(t)e− jωt
= αX1(jω)+βX2(jω) (10.37)
10.3.2 Time and Frequency Duality
The Fourier transform and inverse Fourier transform have very similar forms. If we have a function
g(.) and define a signal as g(t) then its Fourier transform will be G(jω) = F {g(t)} (here we use
the notation F to denote the operation of the Fourier Transform). If we define the spectrum in
terms of our function as g(ω) and take the inverse Fourier transform, we get G(t) = 2πF {g(ω)}.
It should be clear that the form is the same apart from the 2π term, which is required because of
our use of angular frequency.
g(t)
F
−
→ G(w) (10.38)
g(ω)
I F
←
−
− G(t) (10.39)
Figure 10.12 demonstrates this property.
10.3.3 Scaling
We have already described the Fourier transform of a pulse of different durations in section 10.1.6.
The shorter duration pulse has a spectrum that is more spread out, and the longer duration pulse
has a more contracted spectrum. This is known as the scaling property, which states that: com-
pressing a signal will stretch its Fourier transform and vice versa. This can be formally stated
as:
x(at)
F
−
→
1
|a|
X(j(w/a)) (10.40)
10.3.4 Impulse Properties
From the scaling property, we know that as the duration of the pulse approaches zero, the spectrum
spreads out more and more. In the limit, we reach the unit impulse, which has zero width. This
is a particularly important signal. For digital signals, this is simply defined as a signal, δ[n], which
has a value 1 at time 0 and a value 0 elsewhere. For continuous-time signals, the definition is a
little trickier, but is described as a signal which has values of zero everywhere apart at time 0 (i.e.
zero width), and at that time it is described by an integral which defines its area (not height) as 1:
δ(t) = 0 for t 6= 0 (10.41)
Z ∞
−∞
δ(t)dt = 1 (10.42)](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-313-320.jpg)
![292 Chapter 10. Signals and Filters
10.3.8 Analytical and Numerical Analysis
The main practical difference between the transforms is whether an analytical or numerical method
is used in the calculation. For transforms and inverse transforms that contain integrals, it should be
clear that we must be able to find the analytical form of integral. This can be a blessing in that we
can use analytical techniques to find and succinctly describe the transforms of some signals, but
this can also be a difficulty, because first we have to be able to represent the signal as an analytic
function, and second we have to be able to find its integral. When dealing with arbitrary signals,
say from a recording, it is nearly certain that we will not know the form of the signal, and so a
numerical method must be used.
10.3.9 Stochastic Signals
All the transforms we have examined either integrate or sum over an infinite sequence. This will
only give a meaningful result if the sum of the signal over that range is finite; and it can easily
be shown that this is the case for all periodic and many non-periodic signals. There are however
signals for which this is not so, and in such cases we say that the Fourier (or other transform) does
not exist for these signals. One important class of such signals are stochastic signals, which are
signals generated by some random process. The “noisy” signals (for example found in fricatives)
that we have been mentioning up until now are such signals.
In trying to characterise these signals we face two problems. Firstly, we cannot exactly
describe them in the time domain as the very nature of the random process means that there is no
strict pattern in the samples. Secondly, as the transforms we have introduced can not be calculated,
we can not express these signals directly in the frequency domain either. We get round these issues
by focusing on averages within the signal, and specifically we use a measure of the self similarity
of a signal called its autocorrelation function. For a signal x[n], this is defined as
R(j) =
∞
∑
n=−∞
x[n]x[n− j] (10.51)
which is the expected value of the product of a stochastic signal with a time shifted version of
itself. This function does have a Fourier transform, and this gives the power spectral density of
the signal.
10.4 DIGITAL FILTERS
A filter is a mathematically defined system for modifying signals. In signal processing filters are
put to many uses; as their name might suggest they can “filter out” some unwanted portion of a
signal. In speech synthesis, we are interested in them mainly because of the acoustic theory of
speech production, (section 11.1), which states that speech production is a process by which a](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-316-320.jpg)
![Section 10.4. Digital Filters 293
glottal source is modified by a vocal tract filter. This section explains the fundamentals of filters
and gives insight into how they can be used as a basis for modelling the vocal tract. Digital filters
are of course filters which operate on digital signals. Here, we describe a specific but powerful
kind of filter, called the linear time invariant (LTI) filter.
Using the notation x[n] 7→ y[n] to describe the operation of the filter (“7→” is read “maps to”),
we state the time invariance property as:
x[n−n0] 7→ y[n−n0] for any n0 (10.52)
which states that the system behaviour is not dependent on time. A linear system has the property
of scaling, such that if x[n] 7→ y[n] then
αx[n] 7→ αy[n]
and superposition, such that if x1[n] 7→ y1[n] and x2[n] 7→ y2[n] then:
x[n] = x1[n]+x2[n] 7→ y[n] = y1[n]+y2[n]
These combine to form the linearity principle:
x[n] = αx1[n]+βx2[n] 7→ y[n] = αy1[n]+βy2[n] (10.53)
The linearity property combines with the principle of Fourier synthesis - as the filter linearly
combines the components in a waveform, we can construct and analyse a filter for components
separately and not have to worry about their interactions.
10.4.1 Difference Equations
Digital LTI Filters are often described in terms of their time domain difference equation, which
relates input samples to output samples. The finite impulse response or FIR filter has the differ-
ence equation:
y[n] =
M
∑
k=0
bkx[n−k] (10.54)
The infinite impulse response or IIR filter has the difference equation:
y[n] =
M
∑
k=0
bkx[n−k]+
N
∑
l=1
aly[n−l] (10.55)
As well as operating on the input, the IIR filter operates on previous values of the output and for
this reason these filters are sometimes referred to as feedback systems or recursive filters. The](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-317-320.jpg)
![294 Chapter 10. Signals and Filters
-2
-1
0
1
2
0 2 4 6 8 10
amplitude
time
/tmp/tmpos4p_z using 0:1
(a) Input signal
-6
-4
-2
0
2
4
6
0 2 4 6 8 10
amplitude
time
/tmp/tmpwO0Pqx using 0:1
(b) Output signal
Figure 10.13 Operation of an FIR filter on an example, finite waveform
FIR is a special case of an IIR filter, formed when the coefficients {al} are set to zero. (Note that
the {al} coefficients do not have a value for 0 as the output for y[n] must be calculated before it
can be used on the right hand side of the equation.)
Consider a simple FIR filter with M = 3 and bk = {3,−1,2,1}. As M = 3, this is known as
a third order filter. From the FIR definition equation, we can see that this will modify any input
sequence x as follows:
y[n] = 3x[n]−x[n−1]+2x[n−2]+x[n−3] (10.56)
Figure 10.13 shows an example signal x = {0,1,2,1,0,−1,−2,−1,0,0} and the output of the
filter for this example signal
10.4.2 The impulse response
While the above clearly showed how our filter created an output signal y[n] from our given in-
put signal x[n], it isn’t immediately clear how the filter will modify other, different input signals.
Rather than having to examine the output with respect to a wide range of different input signals,
it would be useful to have a characterisation of the effect of the filter that is independent of the
specific form of the input. Such a characterisation is called the response of the filter, and a particu-
larly useful response is provided by using a unit impulse as input. The output is then the impulse
response, denoted h[n].
A unit impulse, δ, is a special function which has value 1 at n = 0, and 0 elsewhere:
δ[n] =
n 1 n = 0
0 n 6= 0
(10.57)](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-318-320.jpg)
![Section 10.4. Digital Filters 295
-6
-4
-2
0
2
4
6
0 2 4 6 8 10
amplitude
time
/tmp/tmpszD3oV using 0:1
(a) FIR filter
0
0.5
1
1.5
2
0 5 10 15 20 25 30
amplitude
time
/tmp/tmpHwWEdb using 0:1
(b) IIR filter
Figure 10.14 Example impulse responses for FIR and IIR filters. Note while the number of coef-
ficients in the IIR filter is actually less than for the FIR filter, the response will continue indefinitely
(only values up to 30 samples are shown)
The output of the example FIR filter for an impulse is given in figure 10.14a, which shows that
the impulse response is simply the coefficients in the filter. Because of this, the duration of the
impulse will always be the order of the filter. As the order is finite, the impulse response is finite
which gives rise to the name of the filter.
Consider now an example first order IIR filter, with a1 = −0.8, and b[0] = 1. Its difference
equation is therefore:
y[n] = x[n]−0.8y[n−1] (10.58)
If we give this filter the unit impulse as an input, then we can calculate the output for any n
after 0. The output is shown in Figure 10.14b. First of all we observe that the output is not of
a fixed duration: in fact it is infinite as the name of the filter implies. Secondly we observe that
the response takes the form of a decaying exponential. The rate of decay is governed by the filter
coefficient, a1. (Note that if a1 1 the output will grow exponentially: this is general undesirable
and such filters are called unstable.)
Consider the impulse response of the second order IIR filter a1 = −1.8,a2 = 1, shown in
figure 10.15a. Just as in the first order case, this filter has an infinite response, but this time takes
the form of a sinusoid. This shows the power of the IIR filter - with only a few terms it can
produce a quite complicated response. The filter a1 = −1.78,a2 = 0.9, has its impulse response
shown in Figure 10.15b. We can see that it is a decaying sinusoid; we have in effect combined the
characteristics of examples 10.14b and 10.15a. The key point to note is that in all cases, the input
is the same, ie a unit impulse. The characteristics of the output are governed solely by the filter -
after all the input has no periodicity or decay factor built in. If we choose a slightly different set](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-319-320.jpg)
![296 Chapter 10. Signals and Filters
-6
-4
-2
0
2
4
6
0 10 20 30 40 50 60 70 80 90 100
amplitude
time
-4
-3
-2
-1
0
1
2
3
4
0 10 20 30 40 50 60 70 80 90 100
amplitude
time
Figure 10.15 Impulse response for 2 IIR filters, showing a sinusoid and damped sinusoid response.
of coefficients, we can create an output with a different period and decay factor.
Different filters can produce a wide range of impulse responses; the decaying sinusoid pat-
terns here only come from a specific type of second order IIR filter. In general though, the more
terms in the filter, the more complicated the impulse response will be. The impulse response is
important for a number of reasons. Firstly, it provides us with a simple way of examining the
characteristics of the filter. Because the input only lasts for a single sample, the output is not “clut-
tered” with the further interference from subsequent input samples. Impressionistically, exciting
the filter with an impulse response can be thought of as “giving the filter a nudge and seeing what
happens”. Secondly, and more importantly, we can use the principle of linearity and time invari-
ance to describe the responses of more general signals by using the impulse response, explained
next.
10.4.3 Filter convolution sum
The response for a non-unit impulse is simply the unit impulse response scaled. So using the
linearity principle, we can state that:
3δ[n] 7→ 3h[n]
As the unit impulse is simply a single sample, a full input sequence x[n], can be characterised by a
sequence of scaled unit impulses:
x[n] = ∑w[k]δ[n−k]
What are the values of w? As the delta function will set everything outside n−k to 0, it should be
clear that w[n] = x[n] for a given n.Therefore, we can state:
x[n] =
K
∑
k=0
x[k]δ[n−k] (10.59)](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-320-320.jpg)
![Section 10.4. Digital Filters 297
This may seem a rather trivial result, but it shows that a signal can be fully described by a set of
appropriately scaled impulse responses. The time invariance property of LTI systems tells us that
a shifted input will give a shifted response. For the impulse function, the following is therefore
true:
δ[n] 7→ h[n]
δ[n−1] 7→ h[n−1]
δ[n−2] 7→ h[n−2]
δ[n−k] 7→ h[n−k] for any integer k
The linear property states that αδ[n] 7→ αh[n] for any α and so the following is also true:
x[0]δ[n] 7→ x[0]h[n]
x[1]δ[n−1] 7→ x[1]h[n−1]
x[2]δ[n−2] 7→ x[2]h[n−2]
x[k]δ[n−k] 7→ x[k]h[n−k] for any integer k
We can use these two properties on equation 10.59 to find the output of the filter
∑
k
x[k]δ[n−k] 7→ ∑
k
x[k]h[n−k]
which gives us:
y[n] = ∑
k
x[k]h[n−k] (10.60)
which is known as the convolution sum. It shows that we can calculate the output for a filter for
any input, provided we know the impulse response h[n].
10.4.4 Filter transfer function
We will now examine the frequency domain description or frequency response of the LTI filter
by means of the z-transform. The IIR difference equation (10.55), expanded to a few terms, is:
y[n] = b0x[n]+b1x[n−1]+...+bMx[n−M]+a1y[n−1]+...+aNy[n−N]
By using the notation Z{.} to denote the z-transform, we can take the z-transform of both sides as
follows:
Y(z) = Z{b0x[n]+b1x[n−1]+...+bMx[n−M]+a1y[n−1]+...+aNy[n−N]}
The coefficients are not affected by the z-transform, so using the linearity of the z-transform, this
simplifies to:
Y(z) = b0Z{x[n]}+b1Z{x[n−1]}+...+bn−MZ{x[n−M]}+a1Z{y[n−1]}+...+aNZ{y[n−N]}}](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-321-320.jpg)
![298 Chapter 10. Signals and Filters
The z-transform time delay property (equations 10.35 and 10.46) states that Z{x[n−k]} = X(z)z−k
and so our equation becomes:
Y(z) = b0X(z)+b1X(z)z−1
+...+bn−MX(z)zM
+a1Y(z)z−1
+...+aNY(z)zN
If we group the Y(z) terms on one side, and factor the common Y[z] and X[z], we get
Y(z) =
b0 +b1z−1 +...+bMzM
1−a1z−1 −...−aNzN
X(z)
The term is known as the transfer function of the filter
H(z) =
Y(z)
X(z)
=
b0 +b1z−1 +...+bMzM
1−a1z−1 −...−aNzN
(10.61)
=
M
∑
k=0
bk
1−
L
∑
l=1
al
(10.62)
and fully defines the characteristics of the filter in the z-domain. From this, the frequency domain
characteristics of the filter can easily be found (explained below).
10.4.5 The transfer function and the impulse response
We have previously shown that the impulse response, which completely characterises the filter,
allows us to calculate the time domain output for any input by means of the convolution sum:
y[n] =
∞
∑
k=−∞
h[k]x[n−k] (10.63)
Section 10.3 showed how to derive the Fourier transform for a convolution and we can apply this
to a digital signal by using the z-transform in the same way.
Y(z) = Z
n ∞
∑
k=−∞
h[k]x[n−k]
o
=
∞
∑
k=−∞
h[k]
∞
∑
n=−∞
x[n−k]z−n
=
∞
∑
k=−∞
h[k]z−k
X(z) (10.64)
The ∑h[k]z−k term is simply the z-transform of h[n] defined in terms of k rather than n, so 10.64
simplifies to
Y(z) = H(z)X(z) (10.65)](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-322-320.jpg)
![Section 10.5. Digital filter analysis and design 299
which is the same result as equation 10.62. This shows that the transfer function H(z) is the
z-transform of the impulse response h[n]. This gives us the following two important properties:
1. a LTI filter is completely characterised by its impulse response, h[n] in the time domain.
The z-transform of the impulse response is the transfer function, H(z), which completely
characterises the filter in the z-domain.
2. The transfer function is a ratio of two polynomials, with the same coefficients as used in the
difference equation.
10.5 DIGITAL FILTER ANALYSIS AND DESIGN
Section 10.4 fully described the relationship between the time domain difference equations and
the z-domain transfer function of a LTI filter. For first order filters, the coefficients are directly
interpretable, for example as the rate of decay in an exponential. For higher order filters this
becomes more difficult, and while the coefficients {al} and {bk} fully describe the filter, they are
somewhat hard to interpret (for example, it was not obvious how the coefficients produced the
waveforms in Figure 10.15). We can however use polynomial analysis to produce a more easily
interpretable form of the transfer function.
10.5.1 Polynomial analysis: poles and zeros
Consider a simple quadratic equation defined in terms of a real variable x:
f(x) = 2x2
−6x+1
This equation can be factorised into
f(x) = G(x−q1)(x−q2)
and then the roots q1,q2 of the quadratic can be found by setting f(x) to 0 and solving for x (in
this case the roots are x = 0.38 and x = 2.62). G is the gain, which simply scales the function.
We can plot the polynomial as shown in Figure 10.16a. The roots of this expression (the
points at which f(x) = 0) are marked on the graph. Although the roots only mark two points, they
do in fact describe the general curve. Figure 10.16b shows the effect of varying the gain - the
curve can become steeper or shallower, or even inverted if the gain is set to a negative number, but
in all cases the general shape of the curve is the same. Because setting x to one of the roots results
in f(x) = 0, we call these roots zeros. Now consider the same polynomial used as a denominator
expression, where by convention we keep the gain as a numerator term:
f(x) =
1
2x2 −6x+1
=
G
′
(x− p1)(x− p2)](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-323-320.jpg)
![302 Chapter 10. Signals and Filters
complex (0.5+0.5j and 0.5−0.5j) and it can be seen that the poles do not now lie on the x-axis
(which would only happen if they were completely real).
As the poles and zeros completely characterise the filter (apart from the gain), we can dis-
pense with the need to draw 3-dimensional graphs and simply plot where the poles and zeros lie
on the z-plane. This is shown in Figure 10.18c. Figures 10.18b and 10.18d show the 3-d plot and
pole-zero plot for a different 2nd order filter.
This analysis can be extended to any digital LTI filter. As the general form transfer function,
H(z) is defined as the ratio of two polynomials
H(z) =
b0 +b1z−1 +...+bMz−M
1−a1z−1 −...−aNz−N
this can always be expressed in terms of its factors:
H(z) = g
(1−q1z−1)(1−q2z−1)...(1−qMz−1)
(1− p1z−1)(1− p2z−1)...(1− pNz−1)
from which pole-zero analysis can be performed. Furthermore, a factored filter can be split into
smaller filters:
H(z) = g
1−q1z−1
1− p1z−1
| {z }
.
1−q2z−1
1− p2z−1
| {z }
...
1−qMz−1
1− pMz−1
| {z }
= H1(z)H2(z)...HN(z) (10.66)
which is often a useful step when constructing or analysing complicated filter.
10.5.2 Frequency Interpretation of z-domain transfer function
Recall that we showed the relationship between the z-transform
X(z) =
∞
∑
n=−∞
x[n]z−n
and the discrete-time Fourier transform by setting z = ejω̂:
X(ejω̂
) =
∞
∑
−∞
x[n]e− jωn
Given any z-domain expression, we can generate the equivalent frequency domain expression by
simply replacing z with ejω. In section 10.2.6, we explained that the spectrum X(ejω̂) of a digital
signal is periodic with period 2π for normalised frequency. This can be thought of as a function
which sweeps out a circle of radius 1 in the z-domain. The frequency domain is in effect the
particular ‘slice” on the z-domain that this circle describes.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-326-320.jpg)
![Section 10.5. Digital filter analysis and design 303
-1.5
-1
-0.5
0
0.5
1
1.5
imaginary
-1.5
-1
-0.5
0
0.5
1
1.5
real
-1
0
1
2
3
4
|H(z)|
(a) z-domain
0
2
4
6
8
10
-4 -3 -2 -1 0 1 2 3 4
amplitude
normalise frequency
(b) frequency domain
Figure 10.19 Relationship between the z-domain and the frequency domain. The unit circle drawn
on the z-domain diagram can be unfolded and straightened to produce the frequency response shown
in figure (b). [ MORE: circle has yet to be drawn]
Let us return to the 3-dimensional representation of the magnitude of H(z), for an example
2nd order filter, as shown in Figure 10.19a. This is as before, but the unit circle is now superim-
posed. The value at the unit circle rises and falls with the shape of the plot, and when we “unfold”
this, we get a 2-dimensional graph of magnitude against frequency as shown in Figure 10.19b. The
phase spectrum can be generated in the same way, that is, by taking the values from the unit circle](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-327-320.jpg)
![304 Chapter 10. Signals and Filters
from the z-domain phase function. As previously mentioned, we can dispense with 3-d graphs
because the pole-zero diagram fully characterises the transfer function. The frequency domain
can easily be obtained from the z-domain by use of the poles and zeros.
10.5.3 Filter characteristics
We will now use the above results on pole zero analysis and frequency domain interpretation to
examine the properties of first order and second order filters, and show how these can be extended
to the general case of all digital LTI filters.
Consider again a first order IIR filter:
h[n] = b0x[n]−a1y[n−1]
H(z) =
b1
1−a1z−1
H(ejω
) =
b1
1−a1e− jω
Figure 10.20, shows the time domain, z-domain pole zero plot and frequency domain magnitude
spectrum for this filter with b0 = 1 and a1 varying over 0.8,0.7,0.6,0.4. Figure 10.20d shows a
3-d z-domain magnitude plot for a single value, a1 = 0.7.
It is clear that the value of the coefficient a1 completely describes the shape of the three
functions. For a single term polynomial, the root of 1−a1z−1 is simply a1. As the filter coefficient
is real, so will the root, and hence the pole will always line on the real axis in the pole-zero plot.
This type of filter is commonly called a resonator because it accurately models the types
of resonances commonly found in nature. The peak in the the frequency response is known as
a resonance because signals at or near that frequency will be amplified by the filter. For our
purposes, it provides a good model of a single vocal tract formant. Resonances are normally
described by three properties: amplitude - the height of the peak, frequency - where the peak lies
in the spectrum, and bandwidth - a measure of the width or sharpness of the peak.
The radius of the pole, r controls both the amplitude and the bandwidth of the resonance.
As the value of a1 decreases, its pole moves towards along the real axis towards the origin, and
hence the frequency response has a lower amplitude at ω̂ = 0. From the difference equations, we
know that a first order IIR filter with a1 1 is simply a decaying exponential. Small values of a1
correspond to slow decay and narrow bandwidth, large values (still 1) a steeper decay and wider
bandwidth. This is exactly what we would expect from the Fourier transform scaling property
The standard measure of bandwidth is the 3db down point which is the width of the reso-
nance at 1/
√
2 down from the peak. For larger values of r = |a1|, two useful approximations of
the bandwidth (in normalised frequency) are:
B̂ ≈ −2ln(r)
B̂ ≈ 2(1−r)
√
r
0.5 r 1.0 (10.67)](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-328-320.jpg)
![Section 10.5. Digital filter analysis and design 307
0
1
2
3
4
5
6
7
-4 -3 -2 -1 0 1 2 3 4
amplitude
normalise frequency
(a) pole-zero plot
-1
-0.5
0
0.5
1
-1 -0.5 0 0.5 1
amplitude
time
/tmp/tmpV_ny2Q using 1:2
/tmp/tmpYSkHQr using 1:2
/tmp/tmpmWEdeH using 1:2
/tmp/tmp5L1og7 using 1:2
/tmp/tmpwCVli3 using 1:2
(b) frequency response
Figure 10.22 Plots of second order IIR filter for a series of poles. All have the same angle, but
differ in distance from the unit circle. The poles in the pole-zero diagram are conjugate symmetric,
meaning they are symmetric in the real-axis. Because of this the frequency response will have peaks
at the same positive and negative values
a single pole in the positive range. If we have two poles, we naturally now have a second order
filter:
h[n] = b0x[n]−a1y[n−1]−a2y[n−2]
H(z) =
b1
1−a1z−1 −a2z−2
H(z) =
1
(1− p1z1)(1− p2z1)
H(ejω
) =
b1
1−a1e− jω −a2e−2jω
The full frequency response can be seen more clearly by expressing the poles in polar form:
p1 = rejθ
p2 = p∗
1 = re− jθ](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-331-320.jpg)
![Section 10.5. Digital filter analysis and design 311
0
20
40
60
80
100
120
0 1000 2000 3000 4000 5000 6000 7000
amplitude
frequency (Hz)
(a) frequency response
-1
-0.5
0
0.5
1
-1 -0.5 0 0.5 1
imaginary
real
/tmp/tmpKzOhgH using 1:2
/tmp/tmp8uLEzm using 1:2
(b) pole-zero plot
Figure 10.25 Three formant, 6 pole model of /ih/ vowel
We construct a second order IIR filter for each formant, Fn:
Hn(z) =
1
(1− pnz−1)(1− p∗
nz−1)
where pn is the pole and p∗
n is complex conjugate of the pole for the formant n. We create the
complete transfer function H/ih/ for the /ih/ vowel by simply multiplying each formant’s transfer
function:
H/ih/(z) = H1(z)H3(z)H3(z)
=
1
(1− p1z−1)(1− p∗
1z−1)(1− p2z−1)(1− p∗
2z−1)(1− p3z−1)(1− p∗
3z−1)
Figure 10.25 shows the pole-zero plot and frequency response of the transfer function. From the
plot, we can see that the formant frequencies do indeed appear at the approximately the desired
locations.
By definition, H gives the transfer function and frequency response for a unit impulse. In
reality of course, the vocal tract input for vowels is the quasi-periodic glottal waveform. For
demonstration purposes, we will examine the effect of the /ih/ filter on a square wave, which we
will use as a (very) approximate glottal source. We can generate the output waveforms y[n] by
using the difference equation, and find the frequency response of this vowel from H(ejω). The
input and output in the time domain and frequency domain are shown in figure 10.26. If the
transfer function does indeed accurately describe the frequency behaviour of the filter, we should
expect the spectra of y[n], calculated by DFT to match H(ejω)X(ejω). We can see from figure
10.26 that indeed it does.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-335-320.jpg)
![314 Chapter 10. Signals and Filters
Transforms
• The Fourier transform converts a continuous waveform into a continuous spectrum:
X(jω) =
Z ∞
−∞
x(t)e− jω
dt
• The discrete-time Fourier transform converts a digital waveform into a continuous spectrum:
X(ejω̂
) =
∞
∑
−∞
x[n]e− jω̂n
• The discrete Fourier transform converts a digital waveform into a discrete spectrum:
X[k] =
N−1
∑
n=0
x[n]e− j 2π
N kn
• The z-transform converts a digital waveform into a continuous z-domain representation:
X(z) =
∞
∑
n=−∞
x[n]z−n
• All transforms have the properties of linearity, scaling, time delay, frequency shift and con-
volution.
Digital filters
• Digital, time-invariant linear filters operate on an input signal x[n] and produce an output
signal y[n].
• A filter is fully defined by two sets of real coefficients {al} and {bk}.
• The operation of the filter in the time domain is given by the difference equation.
y[n] =
M
∑
k=0
bkx[n−k]+
N
∑
l=1
aly[n−l] (10.71)
• The time domain characteristics of a filter are fully captured by the response h[n] to a unit
impulse δ[n].
• The operation of the filter in the z-domain is given by the transfer function.
H(z) =
b0 +b1z−1 +...+bMz−M
1−a1z−1 −...−aNz−N
• The transfer function H(z) is the z-transform of the impulse response h[n].](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-338-320.jpg)
![11
ACOUSTIC MODELS OF
SPEECH PRODUCTION
The speech production process was qualitatively described in Chapter 7. There we showed that
speech is produced by a source, such as the glottis, which is subsequently modified by the vocal
tract acting as a filter. In this chapter, we turn our attention to developing a more formal quantita-
tive model of speech production, using the techniques of signals and filters described in Chapter
10.
11.1 ACOUSTIC THEORY OF SPEECH PRODUCTION
Such models often come under the heading of the Acoustic Theory of Speech Production, which
refers both to the general field of research in mathematical speech production models, and to the
book of that title by Fant [158]. While considerable previous work in this field had been done prior
to its publication, this book was the first to bring together various strands of work and describe
the whole process in a unified manner. Furthermore, Fant backed his study up with extensive
empirical studies with x-rays and mechanical models to test and verify the speech production
models being proposed. Since then, many refinements have been made to the model, as researchers
have investigated trying to improve the accuracy and practicalities of these models. Here we focus
the single most widely accepted model, but conclude the chapter with a discussion on variations
on this.
As with any modelling process, we have to reach a compromise between a model which
accurately describes the phenomena in question, and one which is simple, effective and suited
to practical needs. If we tried to capture every aspect of the vocal organs directly, we would
have to account for every muscle movement, articulator shape and tissue absorption characteristic
directly, and then determine how each of these affected how speech production. This would be a
huge undertaking, and even if an accurate description of all these factors could be determined, the
result may be too complicated for useful analysis. We therefore have to make quite a number of
simplifying assumptions in order to obtain a usable model.
316](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-340-320.jpg)
![318 Chapter 11. Acoustic Models of Speech Production
mass
Figure 11.2 Simple mass/spring oscillator. If the mass is displaced from its original, at rest posi-
tion, it will indefinitely oscillate horizontally.
For example, in the case of vowels, speech is produced by the glottal source waveform
travelling through the pharynx, and as the nasal cavity is shut off, the waveform progresses through
the oral cavity and is radiated into the open air via the lips. Hence as filters connected in series are
simply multiplied in the z-domain, we can write the system equation for vowels as:
Y(z) = U(z)P(z)O(z)R(z) (11.1)
Where U(z) is the glottal source, with P(z), O(z) and R(z) representing the transfer functions of the
pharynx, the oral cavity and the lips respectively. As P(z), O(z) linearly combine, it is normal to
define a single vocal tract transfer function V(z) = P(z)O(z), such that Equation 11.1 is written
Y(z) = U(z)V(z)R(z) (11.2)
which can be represented in the (discrete) time domain as:
y[n] = u[n]⊗v[n]⊗r[n] (11.3)
The configuration of the vocal tract governs which vowel sound is produced, and by studying
this we can again an understanding of how the physical properties of air movement relate to the
transfer function, and sounds produced. In a similar fashion, transfer functions can be determined
for the various types of consonants. To determine the form these transfer functions take, we have
to investigate the physics of sound, and this is dealt with next.
11.2 THE PHYSICS OF SOUND
11.2.1 Resonant systems
We have informally observed that the vocal tract filter acts as a resonator; that is, it amplifies
certain frequencies and attenuates others. How does this behaviour arise?
The resonant nature of systems is often demonstrated by considering the motion of a mass
and spring as shown in Figure 11.2. Imagine this system is set in motion by displacing the mass to](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-342-320.jpg)
![Section 11.2. The physics of sound 321
fixed
end
Figure 11.5 If the held end of the rope is given a jerk, this creates an impulse, which will travel
along the rope towards the wall.
11.2.2 Travelling waves
The behaviour of the vocal tract is in many ways similar to the behaviour of the mass/spring/damped
system just described. The input, “forcing function” is the glottal source, which is used to drive the
system (the vocal tract) whose physical properties determine resonant characteristics, which then
modify the input signal so as to amplify some frequencies and attenuate others. One crucial differ-
ence however, is that the mass/spring/damped system is what is known as a lumped parameter
system in that it comprises easily identifiable, discrete components such as the mass and spring.
By contrast, the vocal tract is a distributed system, in which the properties of inertia, capacitance
and resistance are evenly distributed throughout the system. We can formulate the properties of
a distributed system by modelling it as a large number of small connected discrete systems; that
is. Using calculus, we can determine the behaviour of a distributed system by considering the
behaviour of the system as the size of the elements tends to zero. This steps involved in this are
fully described in Rabiner and Schafer [368]; for our purposes though, we will simply take the
results of this analysis and apply it to the problem of the sound propagation in the vocal tract.
Before considering this directly, we will appeal again to an equivalent physical system, as
this will let us examine an aspect of resonant behaviour that is central to our study, namely that
of travelling waves. Consider first the easy to visualise example of waves travelling in a rope,
in which one end is in a person’s hand and the other is tied to a wall. This is shown in Figures
11.5 and 11.6. If the person introduces some energy in the rope by giving the rope a jerk (i.e. an
impulse), a wave will form at that end and travel down the rope. When the wave reaches the tied
end at the wall, it will be reflected, and will travel backwards towards the person. When it reaches
the held end, the wave will be reflected back again towards the wall and so on, until all the energy
has eventually been lost from the system.
The person can create a forcing function (i.e. source) input to the system by moving their
hand in a repetitive fashion. In doing so, a series of pulses will form and travel forwards down
the rope towards the wall, and these pulses will create a travelling wave, the frequency of which
is dictated by the hand movement. When this forwards travelling wave is reflected, a backwards
travelling wave will be created. The forward wave and backward wave will interfere - if the two
have a peak at a given point, a large peak will occur, and if one has a negative peak while the
other has a positive peak then these will cancel out leaving the rope at that point in its neutral](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-345-320.jpg)
![326 Chapter 11. Acoustic Models of Speech Production
lungs
lips
glottis
Figure 11.9 Vocal tract model composed of sequence of joined uniform tubes.
each other out; when the forward wave meets an equal size volume velocity of the backward wave,
the overall effect is to have no motion.
Equation 11.5 allows us to state an expression for pressure in terms of the forward and
backward volume velocities at a point in space and time in a similar way
p(x,t) =
ρc
A
u+
(t −x)+u−
(t +x)
(11.7)
Here ρc/A is the characteristic impedance Z0 of the tube. Note that in this case, the terms are
additive; an increase in pressure from the forward wave meeting an increase in pressure from the
backward wave results in an even higher combined total pressure. If the area of the tube remains
constant, the wave simply propagates through the tube. If however the area changes, then the
impedance changes, and this causes reflection. The reflections set up standing waves and these
cause resonances. In this way, the impedance patterns of the tube govern the resonant properties
of the model2
11.3 VOWEL TUBE MODEL
In vowel production, the glottal source creates a wave which passes through the pharyngeal cavity,
then the oral cavity and finally radiates through the lips. The nasal cavity is blocked off, and so
the pharyngeal cavity and mouth can be combined and modelled as a single tube. This tube varies
in cross sectional area along its length, and it is the ability of a speaker to vary the configuration
of this tube that gives rise to the different vowel sounds. Modelling a continuously varying tube is
complicated, but we can approximate this by considering a series of short uniform tubes connected
in series, shown in Figure 11.9. If the number of tubes is sufficiently large, a continuous tube can
be modelled with arbitrary accuracy using this method (we return to the issue of just how many
tubes are needed in Section 11.3.6).
Recall that we described the overall process of vowel production as one where a signal from
the glottis u[n] travels through and is modified by the vocal tract and lips to produce the speech
waveform y[n]. In the z-domain, this is represented by Equation 11.2:
Y(z) = U(z)V(z)R(z)
2 Note that the above expressions can be derived from a first principles solution of the wave equations and the proper-
ties of air. This is beyond the scope of our study, but several good accounts are available [368] [364].](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-350-320.jpg)
![Section 11.3. Vowel Tube Model 327
where U(z) is the z-transform of the glottal source, V(z) is the vocal tract filter and R(z) is the
radiation effect from the lips. Initially, we wish to find V(z) as the is the part of the expression
that characterises the phonetic nature of the sound. From V(z) we can find the spectral envelope,
formants and other linguistically useful information. To find V(z), we study the evolution of the
volume velocity signal from the glottis to the lips. We will denote the volume velocity at the lips
UL(z), and to avoid confusion, with denote the volume velocity glottal source as UG(z) rather than
simply as U(z). Hence our expression is
UL(z) = V(z)UG(z) (11.8)
After this, we can determine the final speech signal Y(z) by applying the radiation characteristic
R(z) to UL(z). Rather than attempt to find V(z) from the tube properties directly, we will use the
tube model to find UL(z) in terms of UG(z), and divide these to find V(z).
11.3.1 Discrete time and distance
In practice we wish to use our analysis on digital signals, so our first task is to convert the expres-
sions for volume velocity and pressure in a tube into a digital form, so that we can perform the
analysis independent of sampling rate.
Recall from Section 10.2.1 that we can convert time t to a discrete representation time vari-
able n by
n = tFs
where Fs is the sampling rate. As is standard in discrete signal processing, we will use normalised
frequency (F̂) and normalised angular frequency (ω̂), which are calculated from frequency (F)
and angular frequency (ω) by
F̂ =
ω̂
2π
=
F
Fs
In addition, it is convenient to normalise our measure of distance from the start of the tube, x, with
respect to the sampling frequency. It also helps simplify the analysis if we normalise this with
respect to the speed of sound. As with time and frequency normalisation, this allows us to derive
general expressions for the system independent of the sampling frequency used and the speed of
sound in the medium. Hence normalised distance d is given by
d =
xFs
c
(11.9)
Using this result we can state the volume velocity and pressure functions, given in Equations
11.6, 11.7 and 11.9, in discrete normalised time and distance as
u[d,n] = u+
[n−d]−u−
[n+d] (11.10a)
p[d,n] =
ρc
A
h
u+
[n−d]−u−
[n+d]
i
(11.10b)](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-351-320.jpg)
![328 Chapter 11. Acoustic Models of Speech Production
U [n − D ]
k
+
k
−
U [n]
−
U [n]
k
+
U [n]
k+1
−
U [n − D ]
k+1
+
U [n − D ]
k+1
−
U [n]
k+1
+
−
+
k
−
U [n + D ]
Figure 11.10 Behaviour of forwards and backwards travelling waves at a tube junction.
In the next section we will develop a general expression for the reflection and transmission
of a wave at the junction of two tubes and in in the subsequent sections, show how this result can
be used to determine the behaviour of the whole system.
11.3.2 Junction of two tubes
If we consider a single tube k in our model, from Equations 11.6, 11.7 and 11.9 we can state the
volume velocity and pressure functions in discrete normalised time and distance as
uk[d,n] = u+
k [n−d]−u−
k [n+d] (11.11a)
pk[d,n] =
ρc
Ak
h
u+
k [n−d]−u−
k [n+d]
i
(11.11b)
Now we consider the junction between a tube k with normalised length Dk and its neighbouring
tube k + 1, shown in Figure 11.10. As this tube is also uniform, it can also be described by
Equations 11.11a and 11.11b.
Joining two tubes creates a boundary condition at the junction, and it is this junction which
determines how the sound propagates as a whole. In considering behaviour at a junction, we can
use an important principle of physics which states that pressure and volume velocity cannot change
instantaneously anywhere. So despite the sudden change in cross sectional area, the volume veloc-
ity and pressure cannot abruptly change, and it follows from this that at the point of the junction
the pressure and volume velocity must be equal. Hence
uk[Dk,n] = uk+1[0,n]
pk[Dk,n] = pk+1[0,n]
Using these boundary conditions we can relate the forwards and backwards volume velocity equa-
tions of tube k and tube k+1 at the junction. That is, at the end (d = Dk) of tube k and the start of
tube k +1 (d = 0)
u+
k [n−Dk]−u−
k [n+Dk] = u+
k+1[n]−u−
k+1[n] (11.12)](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-352-320.jpg)
![Section 11.3. Vowel Tube Model 329
U [n]
U [n]
U [n − D ]
A
A
transmission
coefficient
reflection
coefficient
k
+
+
k+1
−
k+1
k k+1
Figure 11.11 Tube junction showing how wave u+ is made from reflection of U- in tube k+1, and
transmission of u+ from tube k
and likewise with the pressure:
ρc
Ak
u+
k [n−Dk]+u−
k [n+Dk]
=
ρc
Ak+1
u+
k+1[n]+u−
k+1[n]
(11.13a)
1
Ak
u+
k [n−Dk]+u−
k [n+Dk]
=
1
Ak+1
u+
k+1[n]+u−
k+1[n]
(11.13b)
we can substitute these equations into one another and eliminate one of the volume velocity terms.
For example, from Equation 11.12 we know that u−
k [n + Dk] = u+
k+1[n] − u−
k+1[n] + u+
k [n − Dk],
which we can substitute into Equation 11.13b to give an expression for u+
k+1[n]
u+
k+1[n] =
h 2Ak+1
Ak +Ak+1
i
u+
k [n−Dk]+
hAk+1 −Ak
Ak +Ak+1
i
u−
k+1[n] (11.14)
This equation is interesting in that it shows the make up of the forward travelling wave in tube
k +1. As we would expect, this is partly made up of the forward travelling wave from tube k and
partly from some of the backward travelling wave being reflected at the junction. The coefficient
of u+
k in Equation 11.14 is the amount of u+
k that is transmitted into the next tube, k + 1, and is
termed the transmission coefficient. The coefficient of u−
k+1 is the amount of u−
k+1 that is reflected
back into tube k + 1 at the junction and is termed the reflection coefficient. As they are simply
related, only one is needed to describe the junction and by convention reflection coefficients are
used. The reflection coefficient is often denoted rk (meaning the reflection between tubes k and
k +1) and from 11.14 is given by:
rk =
Ak+1 −Ak
Ak +Ak+1
(11.15)
Of particular importance is that fact that the boundary condition is completely described by the
reflection coefficient, and this is only governed by the areas of the tubes; the speed of sound and](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-353-320.jpg)
![330 Chapter 11. Acoustic Models of Speech Production
density terms have been eliminated. We shall see that reflection coefficients are a simple and useful
way to define the characteristics between two tubes and we will henceforth use them instead of
tube areas. Using reflection coefficients, we can find expressions that relate the forward wave or
the backward in one tube with the forward and negative waves in the other tube:
u+
k [n−Dk] =
1
1+rk
u+
k+1[n]+
rk
1+rk
u−
k+1[n] (11.16a)
u−
k [n+Dk] =
−rk
1+rk
u+
k+1[n]+
1
1+rk
u−
k+1[n] (11.16b)
11.3.3 Special cases of junction
The above result can be used to help set boundary conditions for the tube at the lips and the glottis.
To see this, let us consider three special cases resulting from changing areas in the tubes.
Firstly, if two connected tubes have the same cross sectional area we see from 11.15 that
the reflection coefficient is 0 and the transmission coefficient is 1. That is, there is no reflection
and all the wave is transmitted past the boundary. This is entirely what we would expect from
our uniform tube model. Next, let us consider the situation where the tube is completely closed at
one end. A closed termination can be modelled by having the second tube have an infinitely small
area (that is, the size of the junction between the tubes disappears to nothing and hence creates
a solid wall). The effect of this can be found from Equation 11.15. As the area of the second
tube, Ak+1 → 0, the reflection coefficient rk → −1. Now consider a completely open tube. This
can again be modelled with the tube junction model, but this time the area of the second tube is
infinitely big. Hence, from Equation 11.15, as Ak+1 → ∞ the reflection coefficient rk → 1. Hence
both a closed termination and a open termination completely reflect the wave.
There is a crucial difference between the two situations however and this can be seen from
the sign of the reflection coefficient. In the closed termination case, the tube ends in a solid wall
and the reflection coefficient is -1. From Equation 11.16b we can see that in such a case the
volume velocity will be 0. If we find the equivalent pressure expressions for Equation 11.16b, we
see that the pressure terms add, and hence the pressure at this point is (ρc)/A1. Intuitively, we can
explain this as follows. The solid wall prevents any particle movement and so the volume velocity
at this point must be 0. The pressure however is at its maximum, as the wall stopping the motion
is in effect an infinite impedance, meaning no matter how much pressure is applied, no movement
will occur. In the open end situation, when the reflection coefficient is 1, we see from Equation
11.16b that the volume velocity is at a maximum and now the pressure is 0. Intuitively, this can
be explained by the fact that now the impedance is 0, and hence no pressure is needed to move the
air particles. Hence the pressure p(L,t) is 0, and the volume velocity is at its maximum.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-354-320.jpg)
![336 Chapter 11. Acoustic Models of Speech Production
The principle behind finding the N tube transfer function is the same as for the two tube
model - start at the lips and apply Equations 11.18a and 11.18b for each tube in turn. This is tedious
to perform in the manner used for the two tube system, but we can use matrix multiplication to
speed up the process.
Equations 11.17b and 11.17b can be expressed in matrix form as:
Uk = QkUk+1
where
Uk =
U+
k (z)
U−
k (z)
#
and because D = 1/2, then
Qk =
z1/2
1+rk
−rkz1/2
1+rk
−rkz−1/2
1+rk
z−1/2
1+rk
#
We can use the special reflection coefficients at the lips and glottis to model the terminating bound-
ary conditions as before. In the matrix formulation, this gives
UN+1 =
UL(z)
0
#
=
1
0
#
UL(z) (11.24)
for the lips and
UG =
h 1
1+rG
1rG
1+rG
i
U1(z) (11.25)
for the glottis. We can use these boundary conditions to formulate an expression for the transfer
function for an arbitrary number of tubes
1
V(z)
=
UG
UL
=
h 1
1+rG
1rG
1+rG
i N
∏
k=1
Qk
1
0
#
which can be solved by iteratively calculating the terms. The result of this will always take the
form:
V(z) =
1+r0
2
z−N/2
N
∏
k=1
(1+rk)
1−a1z−1 −a2z−2...−aNz−N
(11.26)
which as we saw before, is simply a delay in the numerator, followed by an all-pole expression in
the denominator. This is an important result, because it shows that a vocal tract without internal
losses can be accurately modelled by an all-pole filter.3
3 Rabiner and Schafer [368] show that the above can be calculated more easily, and without direct matrix multiplication
if we assume that the glottal impedance is infinite.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-360-320.jpg)
![Section 11.4. Source and radiation models 339
sound. At the start of this cycle the vocal folds are closed. Pressure is exerted from the lungs which
causes a build up in pressure beneath the folds until eventually the pressure forces open the folds.
When the folds open, air moves through them and as it does do, the pressure beneath the folds
reduces. Eventually the pressure is low enough so that the tension closes the vocal folds, after
which the cycle repeats. The rate at which this cycle occurs determines the fundamental frequency
of the source, and ranges from about 80Hz to 250Hz for a typical male, and 120Hz to 400Hz for
a typical female or child. A plot of volume velocity against time for the glottal source is shown in
Figure 11.15. Here we see the three main parts of the cycle: the open-phase where the glottis is
opening due to the pressure from the lungs, the return phase where the glottis is closing due to
the vocal tract tension, and the closed phase.
Modelling the glottis is a tricky problem in that building a simple model seems relatively
easy, but developing a highly accurate model that mimics glottal behaviour in all circumstances is
extremely difficult and has not in any way been solved. We will return to the difficulties in accurate
glottal modelling in Section 13.3.5, but for now we will demonstrate a simple model which at least
gives us a flavour of how this works.
An extensive study into glottal modelling is given in Flanagan [164], which describes various
mass/spring/damped systems. These models can be somewhat difficult to model in discrete time
systems, so instead we adopt models which simply generate a time domain function which has the
properties described above. One such model [376] [368] is given by:
u[n] =
1
2 1−cos(πn
N1
0 ≤ n ≤ N1
cos(π(n−N1)
2N2
N1 ≤ n ≤ N2
0 otherwise
(11.30)
where N1 marks the end of the open phase and N2 marks the end of the return phase/start of the
closed phase. Another quite widely used model is the Lijencrants-Fant model [159] (sometimes
known as the LF model). This describes the glottal waveform derivative and is given by
u[n] =
0 0 ≤ n To
E0eα(n−T0)sin[Ωo(n−T0)] T0 ≤ n Te
−E1[eβ(n−Te) −eβ(Tc−Te)] Te ≤ n Tc
(11.31)
where T0 is the instant of glottal opening, Te is the position of the negative minimum, Tc is the
instant of glottal closure, α, β and Ω control the shape of the function, and E0 and E1 determine
the relative heights of the positive and negative part of the curve. The LF function and its integral,
which is the normal glottal flow function, is shown in figure 11.16. Quatieri [364] gives a third
model, in which g(n) is created by two time reversed exponentially decaying sequences:
u[n] = (β−n
u[−n])∗(β−n
u[−n]) (11.32)
which have the nice property of having a simple z-transform](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-363-320.jpg)
![340 Chapter 11. Acoustic Models of Speech Production
0 50 100 150 200 250 300
−15000
−10000
−5000
0
5000
Figure 11.16 Plot Lijencrants-Fant model for glottal flow and derivative glottal flow for a single
period.
U(z) =
1
(1−βz)2
(11.33)
The most important control element in the glottal source is of course the rate at which the
cycles occur. In Equation 11.30, this is determined by the positions of N1 and N2; in Equation
11.32 by T0, Te and T0; and in Equation 11.31 this is determined by the impulse function u[n].
In the last case, the glottal volume velocity function can be thought of as a low pass filtering of
an impulse stream. From these expressions and empirical measurements, it is known that this low
pass filter creates a roll off of about -12dB per octave. It is this, combined with the radiation effect,
that gives all speech spectra their characteristic spectral slope.
While Equation 11.32 gives a reasonable approximation of the glottal volume velocity signal,
it does not model the secondary effects of jitter, shimmer and ripple as mentioned in Section 17.
More realism can be added to the glottal signal by the addition of zeros into the transfer function,
giving the expression:
U(z) =
M
∏
k=1
(1−ukz−1)
(1−βz)2
(11.34)
We have defined U(z) as a volume velocity signal, mainly for purposes of developing the
vocal tract transfer function. While in reality, the radiation R(z) occurs after the operation of V(z),
we aren’t restricted to this interpretation mathematically. As we shall see, it is often useful to
combine U(z) and R(z) into a single expression. The effect of this is to have a system where
the radiation characteristic is applied to the glottal flow waveform before it enters the vocal tract.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-364-320.jpg)
![Section 11.5. Model refinements 341
This is equivalent to measuring the pressure waveform at the glottis, and if we adopt the radiation
characteristic of Equation 11.29, the effect of this is to differentiate the glottal flow waveform. We
call this signal the glottal flow derivative, denoted ṡ[n] and Ṡ(z) and and this is shown along with
the normal glottal flow in Figure 11.16.
The interesting feature of the glottal flow derivative is that it shows that the primary form
of excitation into the vocal tract is a large negative impulse. This is useful in that the output of
the vocal tract immediately after this should approximate well to the true impulse response of the
system.
11.5 MODEL REFINEMENTS
11.5.1 Modelling the nasal Cavity
When the velum is lowered during the production of nasals and nasalised vowels, sound enters
via the velar gap, propagates through the nasal cavity and radiates through the nose. Hence for a
more complete model, we have to add a component for the nasal cavity. This in itself is relatively
straightforward to model; for a start, it is a static articulator, so doesn’t have anywhere near the
complexity of shapes that occur in the oral cavity. By much the same techniques we employed
above, we can construct an all-pole transfer function for the nasal cavity.
The complication arises because of the behaviour of the oral cavity. For the case of vowels,
the sound wave moved from the glottis through the oral cavity and radiated from the lips. In
the case of nasalised vowels, the sound propagates from the glottis and through the pharynx as
before, but at the velum some sound enters the nasal cavity while the remainder propagates through
the mouth. The sound entering through the velum is filtered by the nasal cavity and radiates
through the nose. The sound continuing through the oral cavity radiates through the lips in the
manner described above. One complication to this is that the some sound in the oral cavity is
reflected backwards and will enter the velar gap, and likewise some sound in the nasal cavity will
be reflected backwards and will enter the oral cavity. If we however ignore this interaction for
now, the system for nasalised vowels can be modelled as shown in Figure 11.17 with a simple
splitting operation at the velum. Because we now have two filters operating in parallel, it is not
possible to derive a single transfer function for the whole system. Rather we construct a transfer
function for the oral cavity, a different one for the nasal cavity and then simply add the outputs of
both to find the final signal.
The case of nasal consonants is more complicated. As with nasalised vowels, the sound wave
starts at the glottis moves through the pharynx and partly through the mouth. Some sound now
travels through the velum, nasal cavity and radiates through the nose. The remainder continues
travelling through the oral cavity, but instead of being radiated at the lips, now meets a closure at
the lips. Hence a full reflection occurs, and so the sound travels backwards through the mouth, and
some continues down towards the glottis and some enters the nasal cavity and radiates from the](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-365-320.jpg)
![344 Chapter 11. Acoustic Models of Speech Production
11.5.3 Models with Vocal Tract Losses
The tube model described in Section 11.3 is termed “lossless” because no energy losses occur
during the propagation of sound through the tube itself. Any energy losses which do occur happen
through the lips and glottis, where the signal travels outwards but is not reflected back. Several
studies have been made into modelling the losses which of course do occur in the tubes themselves.
In terms of the discussion given in Section 11.2.3, this means that the impedance now is just
composed of an inductive and capacitance component, but also has a resistive component.
Rabiner and Schafer [368] give an extensive review of this issue and summarise that there
are three main ways in which losses can occur. First, the passage of the wave will cause the walls
of the vocal tract (e.g. the cheeks) to vibrate. This can be modelled by refinements to the transfer
function. The effect is to raise the formant centre frequencies slightly, dampen the formants and
especially the lower formants; we should expect this as it the mass of the vocal tract walls will
prevent motion at the higher frequencies. The effects of friction and thermal conduction can also
be modelled by adding resistive terms to the transfer function. The overall effect of these is the
opposite of above; the formant centre frequencies are somewhat lowered, and the effects of this
are more pronounced at higher frequencies.
Apart from the overall contribution to damping, these effects tend to cancel each other out,
and so the overall contribution of these resistive terms is quite small. As the degree of loss from
the lips and glottis is considerably higher, these loss terms are often ignored altogether as they add
considerable complication to the model with little gain in modelling accuracy. It should be noted
that it is the particular form of the IIR expression that we derived that gives rise to a lossless tube;
other IIR filter expressions can easily model losses.
11.5.4 Source and radiation effects
Our model of the glottis and lips was extremely simple in that we used special reflection coeffi-
cients rG and rL to generate reflections and losses (for values other that |1|). As we might expect,
in reality the situation is much more complicated. Firstly, there is no such thing as an “ideal
source” where we can simply add volume velocity into a system without any further effect. The
source in fact is always to a greater or lesser degree coupled to the rest of the system, and has to
be modelled if more accuracy is desired. In general this is modelled by having a real impedance
connecting the source to the rest of the system. Again, assuming that sound can just propagate
from the lips and “disappear” with no further effect is quite simplistic. A more accurate model can
be created by modelling the radiation as a complex impedance connected as part of the system.
In this chapter we have concentrated on sound propagation through a tube, and saw that even
this, the simplest of all acoustic systems can be quite complicated. Another entire branch of the
field of acoustics is devoted to sound propagation through three dimensional spaces, and this can
be drawn on to produce a better model of radiation through the lips. As with the other refinements
mentioned in this section, modelling both the source and lip losses with more realistic expressions](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-368-320.jpg)
![346 Chapter 11. Acoustic Models of Speech Production
signal (equivalent to an infinite sampling rate) we see that we require an infinite number of
zero length tubes, ie a continuous tube. For all practical purposes, the discrete nature of the
tube has no effect on accuracy.
All pole modelling Only vowel and approximant sounds can be modelled with complete accuracy
by all-pole transfer functions. We will see in Chapter 12 that the decision on whether to
include zeros in the model really depends on the application to which the model is put,
and mainly concerns tradeoffs between accuracy and computational tractability. Zeros in
transfer functions can in many cases be modelled by the addition of extra poles. The poles
can provide a basic model of anti-resonances, but can not model zero effects exactly. The
use of poles to model zeros is often justified because the ear is most sensitive to the peak
regions in the spectrum (naturally modelled by poles) and less sensitive to the anti-resonance
regions. Hence using just poles can often generate the required spectral envelope. One
problem however is that as poles are used for purposes other than their natural one (to model
resonances) they become harder to interpret physically, and have knock on effects in say
determining the number of tubes required, as explained above.
Lossless tube The real losses in the vocal tract are slight, but as the difference between a pole
being on the unit circle (no loss whatsoever) and slightly within the circle (say at a radius
of 0.95) is quite noticeable, then any loss factors will affect the overall system. That said,
so long as any ignored losses in the vocal tract are compensated for by increased loss at the
lips and glottis, the effect should not be too great. If only the lips and glottis are used for
controlling losses, the effect is that a single damping factor is applied to all the formants, and
hence it is not possible to generate formants with independent bandwidths.
Radiation characteristics The use of a reflection coefficient to model the radiation from the lips
is really quite crude. In reality, radiation is very complicated and effected by the exact shape
of the lips and head. In our simple example we effectively modelled the region outside
the lips as an additional tube - for the complete reflection case this is an infinitely wide
tube, that is, the vocal tract opens into a wall that extends infinitely far in all directions.
For the cases of less that total reflection, an additional tube of finite area was used. As a
real head is actually concave with respect to the direction of radiation, rather than convex
in the additional tube case, we see that this in fact is a very crude model. Rabiner and
Schafer, and Flanagan develop more sophisticated expressions for lip radiation [368] [164].
An important radiation effect that we have not touched upon is that the radiation determines
the directional characteristics of the sound propagation. One commonly observed effect is
that high frequencies are more narrowly directed, so that if a speaker is facing away from the
listener, the high frequencies are not heard - the speech sounds “muffled”.
Source filter separation This is one area where our model starts to break down significantly.
Firstly we saw that modelling the glottis as a completely isolated source is quite unrealistic.
In addition however, there are many physiological effects that make a complete source/filter
separation unrealistic. For example, it is clear that when speaking with a high fundamental](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-370-320.jpg)
![Section 11.6. Discussion 347
frequency, the entire set of muscles in the throat and mouth are much tenser than in low
pitched speech.
Glottal source Our model of the source itself is perhaps the crudest and least accurate part of
our model. We know for example that the shape of the volume velocity waveform changes
with frequency - it is not the case that a fixed low pass filter will do. The model described
above also says nothing about vocal effort. It is clear that shouting is not just the same
as speaking with a louder source input - the entire characteristic of the glottal source has
changed. Furthermore, the dynamics of glottal signals are very complicated. In particular,
when real glottal signals are studied in the regions where voicing starts and stops, we see
many irregular pitch periods, isolated periods of very short duration, and glottal signals that
seem very noisy. The glottal model described in Section 11.4 models none of these affects. A
significant amount of research has been conducted into building more sophisticated models,
and progress has been made. However, many of these more sophisticated models are difficult
to implement and control and hence the simple model explained above is often used. We
shall return to this issue in Section 13.3.5, where we shall see that difficulties in modelling
source characteristics well are in fact the most difficult problem in speech synthesis, to the
extent that they have been one of the main driving factors in determining the path of speech
synthesis in recent years.
11.6.1 Further reading
Many books deal extensively with modelling the vocal tract acoustically, including Flanagan [164],
Rabiner and Schafer [368] and of course Frant [158].
11.6.2 Summary
Model of the Vocal Organs
• The vocal apparatus can be modelled as a set of discrete, interconnected components.
• Each component functions as either a source component, which generates a basic sound, or
a filter component, which modifies this sound.
• The filter components are joined in various configurations (Figure 11.17) to form tubes.
• For vowels, nasals and approximant, the source component is the glottis. The glottis is
located at the end of the vocal tract tube.
• Radiation occurs at the lips and nose, and this can also be modelled by a filter.
• The filter components are the pharynx, oral cavity and nasal cavity.
• All of the above are modelled as lossless tubes.
• All of the above act as linear time invariant filters.
• Each component tube can be modelled, without loss of accuracy, as a series of connected
shorter uniform tubes. The number of uniform tubes required is a function of the required](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-371-320.jpg)
![12
ANALYSIS OF SPEECH
SIGNALS
In this chapter we turn to the topic of speech analysis, which tackles the problem of deriving rep-
resentations from recordings of real speech signals. This book is of course concerned with speech
synthesis - and at first sight it may seem that the techniques for generating speech “bottom-up” as
described in Chapters 10 and 11 may be sufficient for our purpose. As we shall see however, many
techniques in speech synthesis actually rely on an analysis phase which captures key properties
of real speech and then uses these to generate new speech signals. In addition, the various tech-
niques here enable useful characterisation of real speech phenomena for purposes of visualisation
or statistical analysis. Speech analysis then is the process of converting a speech signal to an alter-
native representation that in some way better represents the information which we are interested
in. We need to perform analysis because waveforms do not usually directly gives us the type of
information we are interested in.
Nearly all speech analysis is concerned with a three key problems. First, we wish to remove
the influence of phase; second, we wish to perform source/filter separation, so that we can study
the spectral envelope of sounds independent of the source that they are spoken with. Finally we
often wish to transform these spectral envelopes and source signals into other representations, that
are coded more efficiently, have certain robustness properties, or which more clearly show the
linguistic information we require.
All speech signals in the real world are continuous signals which describe the pattern of air
pressure variation over time. These signals can be recorded with a variety of analogue means,
but for computer analysis, we require our signals to be digitised such that the continuous signal
is converted to a discrete signal. For our needs, we need not go into the details of analogue
sound capture and recording as it is most likely that any reader will either come across signals
that are already digitised, or that they will be able to “record” any acoustic signal directly into a
digitised form by using a standard computer sound card, without the need for any interim analogue
recording step1.
1 Many standard books explain the analogue to digital conversion process, for instance see Rabiner and Schafer [368].
350](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-374-320.jpg)
![Section 12.1. Short term speech analysis 351
12.1 SHORT TERM SPEECH ANALYSIS
A common starting point in speech analysis is to find the magnitude spectrum from a speech signal.
We want this to be in a discrete form that is easy to calculate and store store in a computer, so we
use the discrete Fourier transform (DFT) as our principle algorithm. In practice the fast Fourier
transform (FFT) implementation is nearly always used due to its speed (see Section 10.2.4).
12.1.1 Windowing
When we speak, the glottis and vocal tract are constantly changing. This is problematic for most of
the techniques we introduced in Chapter 10 as these were designed to work on stationary signals.
We can get around this problem by assuming that the speech signal is in fact stationary if consid-
ered over a sufficiently short period of time. Therefore we model a complete speech waveform as
a series of short term frames of speech, each of which we consider as a stationary time-invariant
system.
A frame of speech x[n] is obtained from the full waveform s[n] by multiplication by a window
w[n] in the time domain:
x[n] = w[n]s[n] (12.1)
Three common windows are the rectangular window, hanning window and hamming window:
rectangular w[n] =
(
1 0 ≤ n ≤ L−1
0 otherwise
(12.2)
hanning w[n] =
(
0.5−0.5cos(2πn
L ) 0 ≤ n ≤ L−1
0 otherwise
(12.3)
hamming w[n] =
(
0.54−0.46cos(2πn
L ) 0 ≤ n ≤ L−1
0 otherwise
(12.4)
Figure 12.1 shows a sinusoid after windowing by the three windows just described. The rectangu-
lar window abruptly cuts off the signal at the window edges, while the hanning tapers the signal so
that the discontinuities are absent. The hamming window is just a raised version of the hanning,
and so has a similar shape accept for a slight discontinuity at the window edge.
While an essential part of speech analysis, the windowing process unfortunately introduces
some unwanted effects in the frequency domain. These are easily demonstrated by calculating the
DFT of the sinusoid. The “true” frequency domain representation for this, calculated in continuous
time with a Fourier transform, is a single delta peak at the frequency of the sinusoid. Figures 12.2
show the DFT and log DFT for the three window types. The DFT of the rectangular window](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-375-320.jpg)
![360 Chapter 12. Analysis of Speech Signals
0 50 100 150 200 250 300
−12
−10
−8
−6
−4
−2
0
2
(a) Filter bank analysis with evenly spaced bins
0 50 100 150 200 250 300
−12
−10
−8
−6
−4
−2
0
2
(b) Filter bank analysis with bins spaced accord-
ing to the mel-scale
Figure 12.10 Filter bank analysis on magnitude spectra
line with the human sensitivity to sounds at different frequencies.
The mel-scale was the product of experiments with sinusoids in which subjects were re-
quired to divide frequency ranges into sections. From this, a new scale was defined in which
one mel equalled one thousandth of the pitch of a 1KHz tone [415]. The mapping from linear
frequency to this scale is given by:
m = 2595log10(1+ f/700) (12.5)
Further studies based on loudness (i.e. the percept of amplitude) found that a more accurate
representation was one that had a linear response for lower frequencies, becoming logarithmic for
higher frequencies. One popular scale based on this is the Bark-scale [522]:
B = 13arctan(0.76f/1000)+3.5arctan((f/7500)2
) (12.6)
A third popular scale is the equivalent rectangular bandwidth or ERB scale, which is
measured from the ability to detect sinusoids in the presence of noise. From this auditory percep-
tion is described in terms of an ”equivalent rectangular bandwidth” (ERB) as a function of centre
frequency [316].
B = 6.23×10−6
f2
+9.33910−2
f +28.52 (12.7)
12.2 FILTER BANK ANALYSIS
We will now turn to the important problem of source-filter separation. In general, we wish to do
this because the two components of the speech signal have quite different and independent linguis-
tic functions. The source controls the pitch, which is the acoustic correlate of intonation, while the
filter controls the spectral envelope and formant positions, which determine which phones are be-
ing produced. There are three popular techniques for performing source-filter separation. First we](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-384-320.jpg)
![Section 12.3. The Cepstrum 361
will examine filter bank analysis in this section, before turning to cepstral analysis and linear
prediction in the next sections.
If we visually examine a short length DFT such as that in Figure 12.4, or in fact any spec-
trogram, we can see the general shape of the spectral envelope. That is, if we “squint” and ignore
the harmonics, in many cases the shape of spectral envelope is discernible independent of the par-
ticular positions of the harmonics. When doing this we are effectively averaging out the influence
of individual harmonics and creating a blurring effect and this is the principle behind filter bank
analysis.
Filter bank analysis is performed by first creating a series of bins each centred on a particular
frequency. Time domain filter-bank analysis can be performed by creating a band pass filter
(either FIR or IIR) which allows frequencies at or near the bin frequency to pass, but attenuates
all other frequencies to zero. After the filtering, the amount of energy is calculated and assigned
to that bin. Once this has been done for every bin, we have a representation of how energy varies
according to frequency. The effect of the filtering operation is to blur the effect of individual
harmonics, such that the final representation is largely that of the vocal tract and not the source.
Alternatively, we can perform filter bank analysis on spectra. Here, we take a magnitude spectra
and multiply it by a windowing function centred on a bin frequency. This has the effect of setting
all energy outside the window to zero, while allowing us to measure the energy inside the window.
This is show in Figure 12.10a. Often the bins are not spaced at equal frequencies, but defined
logarithmically or according to one of the frequency scales of Section 12.1.5. This is shown in
Figure 12.10b.
Filter bank analysis is a simple and robust technique for finding the spectral envelope, but
it is only a partial solution to source filter separation. First, the amount of “blurring” required
to eliminate the harmonics may be too severe and may eliminate some of the detail we require;
alternatively, and especially with high pitched speech, we can find that peaks in the envelope fall
between two harmonics and therefore are ignored. Secondly, filter bank analysis only gives us
the spectral envelope derived from the vocal tract transfer function; it does not give us the source.
For many applications (e.g. ASR) this is not a problem, but in TTS we generally require both
components of the signal, and so other techniques are normally used.
12.3 THE CEPSTRUM
12.3.1 Cepstrum definition
The cepstrum is (most commonly) defined as the inverse DFT of the log magnitude of the DFT
of a signal:
c[n] = F −1
log| F { x[n] } | (12.8)](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-385-320.jpg)
![362 Chapter 12. Analysis of Speech Signals
0
20
40
60
80
100
120
0 1000 2000 3000 4000 5000 6000 7000 8000
amplitude
normalise frequency
(a) Magnitude spectrum and envelope
0
2
4
6
8
10
12
14
0 1000 2000 3000 4000 5000 6000 7000 8000
amplitude
normalise frequency
(b) log Magnitude spectrum and envelope
-300
-200
-100
0
100
200
300
400
500
600
700
0 50 100 150 200 250
amplitude
samples
(c) Cepstrum
Figure 12.11 Steps involved in calculating the cepstrum. For demonstration purposes, an estimate
of the spectral envelope has been overlaid on the two DFT spectra. The “periodicity” of the harmonics
in the log spectrum can clearly be seen. this gives rise to the spike at point 120 in the cepstrum. The
low values (30) in the cepstrum describe the envelope. It should be clear that these low values are
separated from the spike - because of this it is a simple task to separate the source and filter.
where F is the DFT and F −1 is the inverse DFT. For a windowed frame of speech y[n] the
cepstrum is therefore
c[n] =
N−1
∑
n=0
log
N−1
∑
n=0
x[n]e− j 2π
N kn
ej 2π
N kn
(12.9)
Figure 12.11 shows the process of calculating the DFT, log, and inverse DFT on a single frame of
speech. We will now look at how this operation performs source/filter separation.
12.3.2 Treating the magnitude spectrum as a signal
Consider the magnitude spectrum of a periodic signal, such as that shown in Figure 12.11. As we
have proved and seen in practice, this spectrum will contain harmonics at evenly spaced intervals.
Because of the windowing effects, these harmonics will not be delta function spikes, but will be
somewhat more “rounded”. For most signals, the amplitude of the harmonics tails off quite quickly](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-386-320.jpg)
![Section 12.3. The Cepstrum 363
as frequency increases. We can reduce this effect by compressing the spectrum with respect to
amplitude; this can easily be achieved by calculating the log spectrum. It should be clear in the
many examples of log spectra in this chapter that the relative amplitudes of the harmonics are
much closer in the log spectra than the absolute spectra. Figure 12.5b shows this effect.
Now, forget for a moment what the log spectrum actually represents, and consider it a wave-
form instead. In we were told Figure 12.11b was a waveform, we would informally describe this
as some sort of quasi-periodic signal (after it, it has a repeating pattern of harmonics), with a time
evolving “amplitude effect” which makes some “periods” higher than others. It is clear that the
rate at which the “periods” change is much greater than the evolution of the “amplitude effect”.
If we wished to separate these, we would naturally turn to our standard tool for analysing the
frequency content of waveforms, namely the Fourier transform. If we calculate the DFT of our
“waveform” we get a new representation, which we will designate the cepstrum. This is shown in
Figure 12.11c. As we would expect, the periodicity of the “waveform” is shown by a large spike
at about position 120. Because we haven’t performed any windowing operations which cause
blurring, we would expect this to be close to a delta function.
Again, as we would expect, because the “periods” were not sinusoidal, this has harmonics
at multiples of the main spike. The “amplitude effect” will be present in the cepstrum also, but as
this is varying much more slowly than the “periods”, it will be in the lower range of the cepstrum.
As the “amplitude effect” is the spectral envelope and the spikes represent the harmonics (ie the
pitch), we see that these operations have produced a representation of the original signal in which
the two components lie at different positions. A simple filter can now separate them.
12.3.3 Cepstral analysis as deconvolution
Recall from Section 11.1 that Equation 11.2 described the speech production process as
y[n] = u[n]⊗v[n]⊗r[n]
where where u[n] is the glottal or other source, v[n] is the vocal tract filter and r[n] is the radiation.
In the time domain, these expressions are convolved and so are not easily separable. If we don’t
know the form of the vocal tract filter, we can’t perform any explicit inverse filtering operation to
separate these components.
We can however perform separation as follows. Let us subsume the radiation filter r[n] into
v[n] to give a single filter v′[n]. We can then write the time domain expression of Equation 11.2 as
y[n] = u[n]⊗v′
[n] (12.10)
Taking the Fourier transform of Equation 12.10 gives
Y(ejω
) = U(ejω
)V′
(ejω
)](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-387-320.jpg)
![364 Chapter 12. Analysis of Speech Signals
-6
-4
-2
0
2
4
6
8
0 1000 2000 3000 4000 5000 6000 7000 8000
log
magnitude
frequency (Hz)
Figure 12.12 Series of spectra calculated from cepstra. The 512 point cepstrum has all its values
after a cut off point K set to 0 and then a 512 point DFT is applied to generate a spectrum. The
figure shows this spectrum for 4 values of K, 10, 20, 30, 50. (In the figure each has been graphically
separated for clarity, with K = 10 as the bottom spectrum and K = 50 as the top. In reality all lie in the
same amplitude range.) As the number of coefficients increases, so does the level of spectral detail,
until effects of harmonics start appearing. The choice of optimal cut-off point depends on whether
spectral precision or elimination of harmonics is the top priority.
If we now take the log of the magnitude of this we get
log(|Y(ejω
)|) = log(|U(ejω
|)V′
(|ejω
)|)
= log(|U(ejω
)|)+log(|V′
(ejω
)|)
in which the source spectrum and filter spectrum are now just added together. We can now return
to the time domain by taking an inverse DFT, giving us
c[n] = cu[n]+cv[n] (12.11)
that is, a time domain representation created by addition of source and filter. Unlike Equation
12.10, the source and filter are now added and not convoluted in the time domain.
12.3.4 Cepstral analysis discussion
If we take the DFT of a signal and then take the inverse DFT of that, we of course get back to
where we started. The cepstrum calculation differs in two important respects. First we are only
using the magnitude of the spectrum - in effect we are throwing away the phase information. The
inverse DFT of a magnitude spectrum is already very different from the inverse DFT of a normal
(complex) spectrum. The log operation scales the harmonics which emphasises the “periodicity”
of the harmonics. It also ensures that the cepstrum is the sum of the source and filter components
and not their convolution.
The cepstrum is useful as it splits the signal into an envelope, given by the first few coeffi-
cients, and a source, given by the spike. Subsequent analysis usually throws away one of these](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-388-320.jpg)
![Section 12.4. Linear prediction analysis 365
parts: if we are interested in the vocal tract we use only the low coefficients, if we are interested in
the pitch and behaviour of the glottis we study the peak. We can demonstrate this by calculating a
spectrum in which the spike is eliminated - this is done by simply setting a cut off point K above
which all the cepstral coefficients are set to zero. From this, we can create a log magnitude spec-
trum by applying a DFT on the modified cepstrum. Figure 12.12 shows this for 4 different values
of K. The figure shows that a reasonably accurate spectral envelope can be generated from about
30 coefficients, but the number chosen really depends on the degree of precision required in the
spectrum.
For many cases where the spectral envelope is required, it is advantageous to keep the low
part of the cepstrum as it is, and not convert it back to a frequency domain spectrum. The low
cepstral coefficients form a very compact representation of the envelope, and have the highly de-
sirable statistical modelling property of being independent so that only their means and variances,
and not their covariances, need be stored to provide an accurate statistical distribution. It is for this
reason that cepstral are the feature representation of choice in most speech recognition systems.
Despite this, cepstra are quite difficult to visually interpret, so we often find that while systems use
cepstra internally, we still use spectra when we wish to visually inspect a signal. The higher part
of the cepstrum contains the pitch information, and is often used as the basis for pitch detection
algorithms, explained in section 12.7.1
12.4 LINEAR PREDICTION ANALYSIS
Linear Prediction (LP) is another technique for source filter separation, in which we use the
techniques of LTI filters to perform an explicit separation of source and filter. In LP we adopt a
simple system for speech production, where we have an input source x[n], which passed through a
linear time invariant filter h[n], to give the output speech signal y[n]. In the time domain this is:
y[n] = x[n]⊗h[n] (12.12)
The convolution makes it difficult to separate the source and filter in the time domain, but this is
easily achieved if we transform our system to the z-domain
H(z) =
X(z)
Y(z)
(12.13)
In Section 12.12, we showed that the lossless tube model was a reasonable approximation for the
vocal tract during the production of a vowel. If we assume for now that H(z) can therefore be
represented by an all-pole filter, we can write
Y(z) = X(z)H(z) (12.14)
=
1
1−
P
∑
k=1
akz−k
.X(z) (12.15)](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-389-320.jpg)
![366 Chapter 12. Analysis of Speech Signals
where {ak} are the filter coefficients. By taking the inverse z-transform2 we see that in the time
domain this becomes:
y[n] =
P
∑
k=1
aky[n−k]+x[n] (12.16)
The importance of this equation is that it states that the value of the signal y at time n can be
determined by a linear function of the previous values of y, apart from the addition of the current
value of the input x[n]. Because of this, the technique is known as linear prediction. In essence,
this states that an output signal is closely approximated by a linear sum of the previous samples:
ỹ[n] =
P
∑
k=1
aky[n−k] (12.17)
and this is identical to the difference Equation 12.16 except for the input term. In linear pred-
ication, the input x[n] is usually referred to as the error, and is often written e[n] to show this.
This may seem a little harsh, but the concept comes from the fact that y[n] is the actual signal and
˜
y[n] is its approximation by the model, the only difference being the contribution from x[n]. As
this accounts for the difference between the predicted and actual signal it is called the error. We
now turn to the issue of how to automatically find the predictor coefficients of H(z) for a frame of
speech.
12.4.1 Finding the coefficients: the covariance method
The goal of linear prediction is to find the set of coefficients {ak} which generate the smallest error
e[n], over the course of the frame of speech. We will now examine two closely related techniques
for finding the ak coefficients based on error minimisation. We can sate equation 12.16 in terms of
the error for a single sample:
e[n] = y[n]−
P
∑
k=1
aky[n−k] (12.18)
2
We can most easily prove this in reverse; if
y[n] =
P
∑
k=1
aky[n−k]+x[n]
then taking the z-transform of both sides results in
Y(z) =
n=∞
∑
n=−∞
P
∑
k=1
aky[n−k]
z−n
+E(z)
=
P
∑
k=1
ak
n=∞
∑
n=−∞
y[n−k]z−n
+E(z)
Using the dummy substitution m = n−k gives
=
P
∑
k=1
ak
m=∞
∑
m=−∞
y[m]z−m
z−k
+E(z)
=
P
∑
k=1
akY(z)z−k
+E(z)
rearranging the Y(z) terms gives the all-pole transfer func-
tion of Equation 12.15.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-390-320.jpg)
![Section 12.4. Linear prediction analysis 367
We wish to find the values {ak} which minimise the squared error, summed over the frame of
speech. Using E to denote the sum squared error we can state:
E =
N−1
∑
n=0
e2
[n]
=
N−1
∑
n=0
y[n]−
P
∑
k=1
aky[n−k]
!2
We can find the minimum of E by differentiating with respect to each coefficient ak, and setting
to zero:
δE
δaj
= 0 =
N−1
∑
n=0
2
y[n]−
P
∑
k=1
aky[n−k]
y[n− j]
!
for j = 1,2,3...P
= −2
N−1
∑
n=0
y[n]y[n− j]+2
N−1
∑
n=0
P
∑
k=1
aky[n−k]y[n− j]
which gives
N−1
∑
n=0
y[n− j]y[n] =
P
∑
k=1
ak
N−1
∑
n=0
y[n− j]y[n−k] for j = 1,2,3...P (12.19)
If we define φ(k, j) as
φ(j,k) =
N−1
∑
n=0
y[n− j]y[n−k] (12.20)
Equations 12.19 can be written more succinctly as:
φ(j,0) =
P
∑
k=1
φ(j,k)ak (12.21)
the full form of which is
φ(1,0) = a1φ(1,1)+a2φ(1,2)+...+aPφ(1,P)
φ(2,0) = a1φ(2,1)+a2φ(2,2)+...+aPφ(2,P)
φ(3,0) = a1φ(3,1)+a2φ(3,2)+...+aPφ(3,P)
.
.
.
φ(P,0) = a1φ(P,1)+a2φ(P,2)+...+aPφ(P,P)
We can easily calculate all the quantities φ(1,0),φ(2,0), and so it should be clear that this is a set
of P equations with P unknown terms, namely the ak coefficients. This can be written in matrix
form as](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-391-320.jpg)
![368 Chapter 12. Analysis of Speech Signals
φ(1,0)
φ(2,0)
φ(3,0)
.
.
.
φ(P,0)
=
a1
a2
a3
.
.
.
aP
φ(1,1) φ(1,2) ··· φ(1,P)
φ(2,1) φ(2,2) ··· φ(2,P)
φ(3,1) φ(3,2) ··· φ(3,P)
.
.
.
φ(P,1) φ(P,2) ··· φ(P,P)
(12.22)
and expressed as
φa = ψ
We can find the value of a (that is, the coefficients ak) by solving this equation. This can be
done simply, by solving for a1 in terms of the other coefficients and substituting again, but this is
extremely inefficient. Because the square matrix φ is symmetric a much faster technique known
as Cholskey decomposition can be used [302] [295].
This method is called the covariance method (for somewhat obscure reasons). It works by
calculating the error in the region of samples from n = 0 to n = N −1, but in fact uses samples in
he region n = −P to n = N −1 to do so. This is because the method requires a full calculation of
the error at time n = 0, and to do so, it must use some previous samples to do so. The calculation
over many frames can be found by either simply using the main signal itself, and moving the
start position n to the designated frame start each time; or by creating a separate windowed frame
of speech, of length N + P samples, and performing the calculation starting P samples into the
frame. The choice is mainly determined by implementation issues (e.g. a system may have already
performed the windowing operation before we start). No special windowing functions are required
with this method: either we can perform the calculations directly on the main waveform, or a
rectangular window can be used if windowing must be performed.
12.4.2 Autocorrelation Method
A slightly different method know as the autocorrelation method can also be used, and as this
has many practical advantages over the covariance method, it is often the method most used in
practice.
Recall from Section 10.3.9 that the autocorrelation function of a signal is defined as:
R[n] =
∞
∑
m=−∞
y[m]y[n−m] (12.23)
This is similar in form to Equation 12.20, and if we can calculate the autocorrelation, we can
make use of important properties of this signal. To do so, in effect we have to calculate the
squared error from −∞ to ∞. This differs from the covariance method, where we just consider the
speech samples from a specific range. To perform the calculation from −∞ to ∞, we window the
waveform using a hanning, hamming or other window, which has the effect of setting all values
outside the range 0 = n N to 0.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-392-320.jpg)
![Section 12.4. Linear prediction analysis 369
While we are calculating the error from −∞ to ∞, this will always be 0 before the windowed
area, as all the samples are 0 and hence multiplying them by the filter coefficients will just produce
0. This will also be the case significantly after the window. Importantly however, for P samples
after the window, the error will not be 0 as it will still be influenced by the last few samples at the
end of the region. Hence the calculation of the error from −∞ to ∞ can be rewritten as
φ(j,k) =
∞
∑
−∞
y[n− j]y[n−k]
=
N−1+P
∑
n=0
y[n− j]y[n−k]
which can be rewritten as
φ(j,k) =
N−1−( j−k)
∑
n=0
y[n]y[n+ j −k] (12.24)
As this is a function of one independent variable j−k, rather than the two of 12.24, we can rewrite
it as
φ(j,k) =
N−1−k
∑
n=0
y[n]y[n+k] (12.25)
As the limits of this expression equate to −∞ to ∞, this is the same as the autocorrelation function
12.23, and we can write:
R(j −k) = φ(j,k)
Now we can rewrite the matrix form of 12.21 in terms of the autocorrelation function:
φ(j,0) =
P
∑
k=1
φ(j,k)ak (12.26)
R(j) =
P
∑
k=1
R(j −k)ak (12.27)
This significance of all of this is that the autocorrelation function (of any signal) is even, that is
R(j −k) = R(j +k) = R(| j −k|), and this means that when we now write the matrix in Equation
12.22 in terms of R:
R(1)
R(2)
R(3)
.
.
.
R(P)
=
a1
a2
a3
.
.
.
aP
R(0) R(1) R(2) ··· R(P−1)
R(1) R(0) R(1) ··· R(P−2)
R(2) R(1) R(0) ··· R(P−3)
.
.
.
R(P−1) R(P−2) R(P−3) ··· R(0)
(12.28)](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-393-320.jpg)
![Section 12.5. Spectral envelope and vocal tract representations 375
ai
i = ai−1
j −k2ai−1
i− j (12.32)
Let us now examine what happens when we substitute Equation 12.31 into equation 12.32.
For demonstration we shall choose the case for a third order filter, in which case we have three
expressions:
a3
3 = k3
a3
2 = a2
2 −k3a2
1
a3
1 = a2
1 −k3a2
2
The expansion of A3(z) is
A3
(z) = 1−a3
1z−1
−a3
2z−2
−a3
3z−3
(12.33)
and if we substitute equations for 12.31 into this we get
A3(z) = 1−[a2
1 −k3a2
2]z−1 −[a2
2 −k3a2
1]z−2 −k3z−3
= 1−a2
1z−1 −a2
2z−2 +k3a2
2z−1 +k3a2
1z−2 −k3z−3
The first terms, 1−a2
1z−1 −a2
2z−2, are of course just the expansion of the second order filter A2(z)
A3
(z) = A2
(z)−k3[a2
2z−1
+a2
1z−2
+z−3
] (12.34)
The other terms are similar to the polynomial expansion, except that the coefficients are in the
reverse order to normal. If we multiply each term by z3 we can express these terms as a polynomial
in z rather than z−1 which gives:
A3
(z) = A2
(z)−k3z−3
[a2
2z2
+a2
1z+1]
A3
(z) = A2
(z)−k3z−3
A2
(z−1
) (12.35)
In a similar fashion, the above relation can shown for any order of filter.
The significance of all this can be seen if we compare Equation 12.35 with Equation 11.27.
Recall that Equation 11.27 was used to find the transfer function for the tube model from the
reflection coefficients, in the special case where losses only occurred at the lips. The equations are
in fact identical if we set rk = −ki. This means that as a by product of Levinson-Durbin recursion,
we can in fact easily determine the reflection coefficients which would give rise to the tube model
having the same transfer function to that of the LP model. This is a particularly nice result; not
only have we shown that the tube model is all-pole and that the LP model is all pole; we have
now in fact shown that the two models are in fact equivalent and hence we can find the various
parameters of the tube directly from LP analysis.
Reflection coefficients are a popular choice of LP representation because they are easily
found, they are robust to quantisation error and can in fact be used directly without having to](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-399-320.jpg)
![376 Chapter 12. Analysis of Speech Signals
be converted into predictor coefficients (see Rabiner and Schafer [368] for how this is done).
Furthermore, and perhaps most significantly for our purposes, they have a physical interpretation,
which means that they are amenable to interpolation. For example, if we had two different vocal
tract configurations and wished to generate a set of vocal tract configurations between these, we
could do by interpolating the reflection coefficients, in the knowledge that by doing so we are
creating a set of reasonable vocal tract shapes. This process could not be done with the predictor
coefficients themselves, as this would very likely lead to unnatural or unstable filters.
12.5.4 Log area ratios
Recall from our tube model, that the reflection coefficients were originally defined in terms of
the ratio of the areas between two adjacent tubes. Equation 11.15 stated that the amount of the
forward travelling wave that was reflected back into the tube was given as:
rk =
Ak+1 −Ak
Ak+1 +Ak
Rearranging this gives:
Ak
Ak+1
=
1−rk
1+rk
(12.36)
Hence it is possible to find the ratio of two adjoining tubes from the reflection coefficients. It is
not however possible to find the absolute values of the areas of the tubes (this is independent, just
as the gain is independent of the filter coefficients). A common way of representing 12.36 is in
terms of log area ratios, which are defined as:
gk = ln
Ak
Ak+1
=
1−rk
1+rk
(12.37)
log area ratios come in useful in a variety of ways, for example as a means of smoothing across
joins in a manner that reflects physical reality.
12.5.5 Line spectrum frequencies
Recall from the tube model of Section 11.3 that losses in the model are only caused by wave
propagation from the ends of the tube; no losses occur within the tube itself. Recall also from
section 11.5.3 that the resistance factors which cause the losses have virtually no effect on the
location of the resonant peaks. In using the Levinson-Durbin solution to the LP equations, we
have simplified the loss situation further by assuming that losses only occur from one end of the
tube. Line spectrum frequencies (LSFs) are found from an adaption of the tube model where we
eliminate the losses from the glottis. This can be thought of as a process of creating a perfectly
lossless tube and calculating its transfer function. Because the resistive factors have a minor impact
of the locations of the formants in the frequency domain, this modelling assumption helps us find
good estimates for the formant positions.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-400-320.jpg)
![380 Chapter 12. Analysis of Speech Signals
Figure 12.16 Example spectrogram with formants superimposed, as calculated by a formant track-
ing algorithm.
the most important representations in traditional phonetics, but in addition they are used in formant
synthesis, explained later in Section 13.2, and they have also been used as cost functions in unit
selection, as explained in Section 16.
In traditional algorithms, which operated by a mixture of signal processing and rules, a three
stage approach is generally followed. First the speech is pre-processed, for example to downsam-
ple the speech to the frequency range where the formants occur; second candidate formants are
found, and finally a heuristic step is applied. that determines the formants from the most likely
candidates. As we have just seen, in many cases the poles which result from LP analysis can be
considered to give rise to formants, and so it is natural that many formant tracking algorithms use
LP poles as their initial candidates. An alternative is to find peaks in either the DFT or LP spec-
trum and assign these as candidates. Yet another technique is that of multiple centroid analysis,
based on the properties which define the overall shape of the spectrum [116]. Often the algorithms
fix the number of formants to be found (usually three), and the post processing stage uses dynamic
time warping to find the best formant paths (formant tracks) across the course of the utterance.
More recently, statistical processing has been brought to bear on this problem. In essence,
formant tracking is no different from any statistical problem in which we have evidence extracted
from features, and constraints determined from a knowledge of the problem. Given such an ap-
proach, it is possible to put this in statistical terms in which the candidates form a likelihood,
and the known dynamics of the formants (e.g. their number and how much they can move from
frame to frame) can form a prior. With this, a Viterbi search can be performed which finds the
most likely formants for a sentence in a single step. Algorithms of this kind make use of Hidden
Markov Models [4], nonlinear prediction algorithms [129], and graphical models [297]. These
statistical algorithms are significantly more accurate and robust than their heuristic counterparts.
In the early days of speech research, formants held the centre stage as the representation
for speech. Today however, it is extremely rare to find formants used in speech recognition,
synthesis or other systems. This is partly because robust formant trackers have always proved
hard to develop, and so researchers moved towards cepstra and LP representations that could be
robustly found from speech, even if their form didn’t explicitly show the type of information that](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-404-320.jpg)
![Section 12.6. Source representations 381
was thought most important. Furthermore, formants are really only a powerful representation for
vowel sounds; they are not particularly good discriminators of most consonants. Hence even if
a perfect formant tracker was available, other types of representations would be needed for some
consonant sounds.
12.6 SOURCE REPRESENTATIONS
We now turn to the topic of finding and analysing the source signal from the speech waveform. Of
course linear prediction attempts to do just this, but just as the LP transfer function doesn’t always
equate to the vocal tract transfer function, neither does x[n] equate to the source signal. This is
basically because in LP analysis H(z) over-extends itself due to the minimisation criteria, and so
subsumes some of the characteristics of the true glottal source. With a little effort however, we
can in fact extract a good representation of the source by a number of techniques which we will
now examine. As with all our other modelling problems, we have a range of trade-offs between
modelling complexity and ease of use, and we will examine a few of these now.
12.6.1 Residual signals
In the LP formulation of Equation 12.16
y[n] =
P
∑
k=1
aky[n−k]+x[n]
we termed x[n] the error signal, because it is the difference between the predicted and actual
signal. Hence for a single frame of speech, if we perform LP analysis and note the values of x[n]
we will have a complete error signal for that frame. Alternatively, we can find this by a process of
inverse filtering. From simple rearrangement of Equation 12.16, we can express x[n] as follows:
x[n] = y[n]−
P
∑
k=1
aky[n−k]
=
P
∑
k=0
aky[n−k] where a0 = 1
Here we have expressed x[n] purely in terms of the coefficients {ak} and the current and previous
values of y[n]: if we compare this to equation 10.54, we see that this is just a standard FIR filter.
Hence if we have the speech signal and the coefficients for a single frame, we can find x[n] by a
standard FIR filter operation. Apart from a gain normalisation, this will be identical to the error
signal x[n]. The inverse filtering technique is useful because often we are required to find the error
signal after the LP analysis has been performed, or for cases (explained below) where we wish to](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-405-320.jpg)
![382 Chapter 12. Analysis of Speech Signals
find the residual for a different interval than that used for analysis. If the residual signal is then fed
through the normal LP (IIR) filter, the original speech is reconstructed. This process is shown in
Figure 12.17.
If we apply the inverse filtering technique to each frame in the entire speech waveform, we
create a sequence of residual signals, one for each frame. A single residual waveform for the
whole utterance can be made from these residual frames; but as the details of this are somewhat
subtle, a description of this is left until section 12.6.5.
How do we now find the glottal source signal from the residual x[n]? Recall that in the
z-domain, the main filter expression is
Y(z) = U(z)V(z)R(z)
where Y(z) is the speech output, U(z) is the source, V(Z) is the vocal tract and R(z) the radiation.
Recall from section 11.4, that we can model this source in two main ways. Firstly, we can model
it explicitly as a time domain function, which has a number of parameters determining the length
of the opening phase, the length of the closed phase and so on. The Lijencrants-Fant (LF) [159]
model of this is shown in Figure 11.16. Alternatively, we can model it as a sequence of impulses
which are passed through a glottal filter G(z). From experimentation and explicit modelling, we
know that this glottal filter will contain poles and zeros, but will be of considerably lower order
than the vocal tract filter. Here we will examine the filter model, as a proper understanding of this
helps in the determination of the equivalent parameters for the time domain model. Using I(z) to
represent the impulse sequence, we can therefore express U(z) as:
U(z) = G(Z)I(z) (12.42)
U(z) =
M
∏
k=0
bG
k z−k
1−
N
∏
l=1
aG
l z−l
I(z) (12.43)
where {bG
k } and {aG
l } represent the coefficients of the glottal filter G(z). Recall from section 11.4
that the radiation can be approximated as a differentiator, which can be expressed as a single zero
FIR filter. Because the filter coefficients of the G(z)V(z)R(z) will combine, the result will be that
the numerator will have the zeros of the glottal source and radiation, while on the denominator,
the poles of the glottal filter will combine with the poles of the vocal tract. The problem is that
the result of LP analysis will give us a single expression H(z) for the glottal and vocal tract poles,
without being able to show which poles belong to which filter. Furthermore, the fact that we are
attempting to model a system with poles and zeros with an all-pole analysis further complicates
the matter. To perform more explicit source/filter separation, we need to find a way of cancelling
the effect of the zeros and separating the poles into their component filters.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-406-320.jpg)
![Section 12.6. Source representations 385
being completely reflective and having an infinite impedance; by contrast, during the open phase
there are losses at the glottis, but worse there will definitely be interaction with the sub-glottal
system (that is, the assumption made in section 11.3.3 where we assume that no signal is reflected
back from the lungs will not hold).
So given that G(z) has negligible effect during the closed phase, all we need now do is remove
the effect of R(z) to find V(z). From experimentation and knowledge of radiation properties, we
can approximate R(z) as a differentiator of the signal. This can be modelled by a first order FIR
filter of the form:
y[n] = x[n]−αx[n−1]
Hence the effect of this can be removed by multiplying the signal with the inverse equivalent of
this, which is an IIR filter, acting as an integrator:
x[n] = y[n]+αx[n−1]
For reasons of stability, it is best to have α slightly less than the “true” integrating/differentiating
value of 1; values in the range 0.95 to 0.99 are often used. This process is known as pre-emphasis.
Now we have a reasonably accurate expression for the vocal tract transfer function, we can
inverse filter the speech to find the glottal flow signal. The trick here is to inverse filter the signal
over a whole pitch period (or longer) rather than just the interval over which the LP coefficients
were calculated. By doing this, we inverse filter over the region where the glottis is open, and from
this we can find the form of the input, glottal signal. This process is shown in Figure 12.19, along
with the shapes of the glottal flow and glottal flow derivative from the Lijencrants-Fant model.
The figure clearly shows that the glottal flow derivative is highly similar to the residual signal,
especially with regard to the large negative spike at the instant of glottal closure. The figure also
shows that the assumptions made about the impulse excitation in the basic LP model are also quite
justified; the excitation in the closed phase comes from the large negative going spike in the glottal
derivative signal can be approximated by an impulse quite easily.
In many cases, especially in female or other high pitched speech, the length of the closed
phase can be very small, perhaps only 20 or 30 samples. In the autocorrelation method, the initial
samples of the residual are dominated by the errors caused by calculating the residual from the
zero signal before the window. The high (and erratic) error in the residual can be seen in the first
few samples of the residual in Figure 12.17b. For short analysis windows this can lead to a residual
dominated by these terms, and for this reason, covariance analysis is most commonly adopted for
closed phase analysis.
It should be noted that in many real-life cases, the pre-emphasis term is not applied, and
hence the calculated residual is in fact a representation of the glottal flow derivative. This is the
case for most speech synthesis purposes, and henceforth we will assume that all residuals are
calculated in this way, unless otherwise stated.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-409-320.jpg)
![388 Chapter 12. Analysis of Speech Signals
doesn’t actually save any space, the trick comes in replacing the residual by a much simpler (and
more economical) source signal.
In what we shall henceforth call classical linear prediction the source is modelled by two
systems; one for voiced sounds and one for obstruent sounds. In Figure 12.19, we saw that the in-
tegrated residual signal was a reasonably close approximation to the glottal flow derivative signal.
The primary feature in such a signal is the negative spike caused by the glottal closure. As this is
reasonably close to an impulse, the classical LP model simply uses a single impulse as the excita-
tion for voiced sounds. To do this properly, one impulse is generated for every pitch period, and so
in analysis, an estimation of the pitch period has to be found and stored/transmitted. All obstruent
sounds have a frication source, generated by white noise in our model from section 11.5.2. In the
classical LP system, all we need store/transmit is the fact that we have an obstruent sound, upon
which a noise source is used during reconstruction. For voiced obstruents, both the impulse and
noise source are used.
The impulse/noise model provides a very efficient means of encoding the residual, but does
so at some cost, the effects of which can clearly be heard in the reconstructed signal as a buzz or
metallic sound. Apart from the space saving, having an impulse excited system has the additional
benefit in that we are free to generate the impulses when we like; in other words, we are not
limited to generating the impulses at the same spacing as in the analysis. This provides a means
of changing the pitch of a signal and in fact forms the basis of the synthesis technique described
in Section 13.3.
12.6.5 Parameterisation of glottal flow signal
Performing the above analysis allows us to find a reasonable estimate of the glottal flow or glottal
flow derivative. In fact, we can go one stage further and parameterise this by further analysis. One
technique described by Quatieri [364] involves fitting the Lijencrants-Fant model to the residual by
means of least squares minimisation. In philosophy, this is of course similar to linear prediction
analysis itself, where we attempt to find the set of predictor coefficients which give the lowest
error. In practice, the LF model is non-linear and so the LP technique is not available to us.
Instead, numerical techniques such as the Gauss-Newton method are used to iterate to a set of
optimum model parameters. Dutoit [148] describes a different technique where the glottal source
parameters are actually found at the same time as the predictor coefficients.
While we can find the optimal parameters for any given frame for a model such as LF, we
never actually achieve a perfect fit. In fact, we can determine a sample by sample difference
between the model and the residual. This of course, is another “residual” or “error” signal, which
now represents the modelling error after the LP model and the LF model have been taken into
account. One interesting study by Plumpe et al [356] describes how this signal is in fact highly
individual to speakers and can be used as the basis of speaker recognition or verification.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-412-320.jpg)
![Section 12.7. Pitch and epoch detection 389
12.7 PITCH AND EPOCH DETECTION
12.7.1 Pitch detection
The process of finding the fundamental frequency (F0) or pitch from a signal is known as pitch
detection or pitch tracking. Recall from Chapter 7 that fundamental frequency is defined as the
frequency at which the glottis vibrates during voiced speech, whereas pitch is the perceived rate
of periodicity in the speech signal by the listener. In general, we ignore this distinction, mainly
because in most cases any difference between these two values is less than the known accuracy of
any detection algorithm. One point that is important though is that F0 tracks are in fact used for
many quite different purposes. In some of the signal processing operations described below, an
accurate frame by frame value is required; however in intonational analysis, we are usually only
interested in the general shape of the track, and so small perturbations are often smoothed for such
needs.
As with formant tracking, traditional pitch detection algorithms PDAs often have three
stages; a pre-processing stage, where for instance the speech is downsampled or low-pass filtered,
a bottom up candidate finding step, and a decision making step whereby the best candidates are
picked.
A wide variety of techniques have been proposed for pitch tracking, including:
• Using the cepstrum. As cepstral analysis generates a representation where the pitch infor-
mation appears as a spike, it is possible to find this spike and determine the pitch from that.
• Autocorrelation function. In the time domain, successive pitch periods are usually similar to
one another. As the autocorrelation functions measure how similar a signal is to itself when
shifted, it will have a high value when a a pitch period is most closely overlapped with its
neighbour. We can hence find the period by finding these points where the autocorrelation
function is at a maximum. In a similar fashion, the normalised cross correlation function can
be used.
• Frequency domain. As the harmonics in the DFT are spaced at multiples of the period,
these can be used to find the period. One strength in this technique is that there are multiple
harmonics and hence more than one measurement can be made. A drawback is that they can
often be hard to measure accurately, especially at higher frequencies.
• Residual analysis. As the residual is mostly free from vocal tract influence, it represents the
source, and from this the period can be seen. In practice this technique is rarely used as the
artefacts in LP analysis often make the determination of the period quite difficult.
The classic work on pitch detection is that of Hess [202] who gives a thorough overview
of all the PDA techniques just mentioned. With PDAs, the basic techniques only go so far and
an important component of the algorithm is the use of many heuristic or empirically determined
factors. Bagshaw [28] conducted an objective test on 6 popular techniques and claims that the time
domain super resolution pitch determination algorithm [309] is the best. Other well known pitch-](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-413-320.jpg)
![Section 12.7. Pitch and epoch detection 391
trackers include Talkin’s get f0 and RAPT algorithms [431] (these were not included in Bagshaw’s
test, but are generally acknowledged to be very accurate). As with formant tracking, traditional
PDAs suffer somewhat from their heuristic rather than probabilistic formulation. Recently, more
interest in probabilistic techniques has arisen [141], and for instance Li et al [283] cite accuracy
results for a probabilistic technique which is significantly more accurate than get f0.
12.7.2 Epoch detection: finding the instant of glottal closure
We have just seen that closed phase linear prediction requires that we analyse each pitch period
separately. This type of speech analysis is called pitch synchronous analysis and can only be
performed if we are in fact able to find and isolate individual periods of speech. We do this by
means of an pitch marking orepoch detection algorithm (EDA).
The idea behind an EDA is to locate a single instant in each pitch period that serves as an
“anchor” for further analysis. These positions are often known as pitch marks, pitch epochs or
simply epochs. In general, they can refer to any reference point, but are often described in terms
of salient positions on the glottal flow signal, such as the peak of the flow for each pitch period.
For many types of analysis (such as TD-PSOLA described in Section 14.2) it doesn’t really matter
where the anchor point is chosen, so long as it is consistent from period to period. Often a time-
domain reference point such as the peak of the highest excursion in each period is used, or the
trough of the lowest excursion. That said, many analysis techniques focus on one particular point
known as the instant of glottal closure (IGC). This is manifested as a large negative going spike in
the glottal flow derivative, as shown in Figure 12.21. If we view this as the input to the vocal tract
(which we do be including the radiation term), we see that it closely resembles a large negative
impulse. Such an impulse excites the filter in such a way that we can examine the filter properties
without interference from the ongoing effects of the glottis. As we saw above, this is the basis on
which close-phase LP analysis is performed.
Locating the instant of glottal closure and epoch detection in general is widely acknowledged
as a thorny problem. In the past, most pitch synchronous applications could use a rough estimate
of the ICG, as the technique was not particularly sensitive to determining these instants exactly.
Indeed, the closed phase LP technique described above does not need a particularly accurate esti-
mate - the window used by the technique does not have to start exactly at the ICG. With the more
widespread adoption of pitch synchronous speech synthesis techniques (described in Chapter 14)
it becomes more important to locate the exact ICG, or some other anchor point, more exactly.
Dutoit [148] gives a review of some of the difficulties involved, and describes a process common
in the early 1990s whereby researchers would actually mark epochs by hand. This was seen as
being an acceptable burden when speech databases were small and when the only alternative was
to use other much poorer sounding synthesis techniques. Dutoit describes some of the difficulties
involved in this, most specifically the fact that because of various phase effects, it is not always
the largest negative peak in the waveform which corresponds to the IGC. Furthermore, it is known
that phase distortion can make time domain marking of waveforms virtually impossible for either](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-415-320.jpg)
![392 Chapter 12. Analysis of Speech Signals
0 2000 4000 6000 8000 10000 12000 14000
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
2.2
Figure 12.21 Comparison of (from top to bottom) waveform, laryngograph signal, integrated resid-
ual and residual.
machine or human.
These difficulties led many to a radical alternative, which was to measure the glottal be-
haviour by means other than from the speech waveform. A common techniques was to use a
device known as an electroglottograph or a laryngograph. This consists of two metal discs
which are placed on each side of the protruding part of the larynx (the “adam’s apple”). A very
high frequency alternating current is passed through the system, such that the current passes from
one disc, through the larynx, to the other disc, and from this the electrical impedance and flow of
current across the larynx can be measured. If the glottis is open, the impedance is high and the
current is low; if the glottis is shut the impedance is low and the current is high. Laryngograph sig-
nals (known as Lx signals) do not however generate the glottal flow signal itself as the relationship
between the electrical impedance and flow is not simply related. The Lx signal is however quite
simple, and from it, it is usually possible to find the instants when the glottis shuts (the IGC) and
opens. Figures 12.21 and 12.22 show a speech waveform, an Lx signal and a LP derived estimate
of the glottal flow derivative. The Lx signal clearly shows each pitch period, and is particularly
clear at showing exactly when the transition between voiced and unvoiced sounds occurs, which
is one of the most difficult areas to perform accurate epoch detection. Epoch detection by means
of Lx signals has been quite widely used in TTS in the last 15 years, but it is rare to find explicit
mention of this. Acero [3] does however provide a good account of the process of Lx use for TTS.
There are two main problems with Lx use. Firstly, the technique doesn’t work well for some](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-416-320.jpg)
![Section 12.7. Pitch and epoch detection 393
6000 6200 6400 6600 6800 7000
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
2.2
instant of
glottal closure
instant of
glottal opening
Figure 12.22 Fine grained comparison of (from top to bottom) waveform, laryngograph signal,
integrated residual and residual. The vertical lines show the instants of glottal closure and glottal
opening. Here we can see that while the general periodicity of the form of the LX and residual glottal
flow are similar, the functions are not simply related in shape. Note that when the glottis is open, the
LX signal is at a minimum because the current flow is at a minimum. The integrated residual signal
(c) approximates the glottal volume velocity. Ideally the closed phase in this signal should be flatter,
but the many modelling errors combine to give this part of the waveform a gradual upward slope. A
small sharpening in the gradient of this can be seen at the instant of glottal opening.
speakers due to physiological reasons. Secondly, it is simply awkward to have to record all the
speech data with both a microphone and the Lx device. It is not particularly comfortable (although
it is completely safe) and some speakers, particularly highly paid voice talents, are not happy with
its use. Thankfully, a number of fully automatic techniques now do exist that can extract the
IGCs from the speech waveform directly. Many of these work on the residual, and it is known
that the negative pulse of the residual can give quite an accurate estimation of the IGC if care is
taken with the analysis and the signal does not suffer from phase distortion. Recently several other
techniques have been developed for automatic extraction from speech. Brookes [68] describes a
technique which uses the glottal energy flow, and claims that this is significantly more accurate
than residual analysis. Vedhius describes a general technique that uses dynamic programming
to constrain the epoch search, and shows that this technique can be extended to other points of
interest besides the IGC [469].](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-417-320.jpg)
![Section 12.8. Discussion 395
quite different.
Currently, mel-scale cepstral coefficients, and perceptual linear prediction coefficients trans-
formed into cepstral coefficients, are popular choices for the above reasons. Specifically they are
chosen because they are robust to noise, can be modelled with diagonal covariance, and with the
aid of the perceptual scaling are more discriminative than would otherwise be. From a speech
synthesis point of view, these points are worth making, not because the same requirements exist
for synthesis, but rather to make the reader aware that the reason MFCCs and PLPs are so often
used in ASR systems is for the above reasons, and not because they are intrinsically better in any
general purpose sort of way. This also helps explain why there are so many speech representations
in the first place; each has strengths in certain areas, and will be used as the application demands.
In fact, as we shall see in Chapter 16, the application requirements which make, say, MFCCs so
suitable for speech recognition are almost entirely absent for our purposes. We shall leave a dis-
cussion as to what representations really are suited for speech synthesis purposes until Chapter
16.
12.8.1 Further reading
There are several good books on speech analysis, most of which adopt an application neutral
attitude. The classic work is Rabiner and Schafer [368] and while written several decades ago,
this book remains an excellent reference for all the fundamentals of speech analysis. It assumes
the reader is familiar with basic signal processing, and so can be a little difficult for those with
no prior experience. Given an understanding of the topics covered in Chapter 10 of this book, the
reader should have little difficulty. An excellent more recent book is Quatieri [364]. The subject
matter doesn’t differ much from that of Rabiner and Schafer, but the book covers many of the
techniques developed since. Other good references include Furui [168], Huang et al [224], Kleijn
and Paliwal [256], Deller and Proakis [128] and Gold and Morgan [176].
12.8.2 Summary
• Speech analysis is in general concerned with three main problems:
1. eliminate phase; the magnitude DFT does this
2. separate source and filter; this can be done with cepstral analysis or linear prediction
3. transform the representation into a space which has more desirable properties; log mag-
nitude spectra follow the ear’s dynamic range; mel-scaled cepstra scales according to
the frequency sensitivity to the ear. Log area ratios are amenable to simple interpolation
and line spectral frequencies show the formant patterns robustly.
• The continuously varying signal is divided by a process called windowing into a sequence
of stationary smaller signals called frames.
• Windowing affects the signal and so a perfect representation of the original waveform is not
possible.](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-419-320.jpg)
![398 Chapter 13. Synthesis Techniques Based on Vocal Tract Models
[F0 ] = 121,123,127,128,126 (5 frame example)
[F0 ] = 121,123,127,128,126,121,118 (7 frame example)
and so a example of a complete specification item would look like
s
s
sj =
PHONEME /ah/
DURATION 70
F0 125,123,119,117,115,115,114
Having different numbers of F0 values per item can be unwieldy, so someti](https://image.slidesharecdn.com/anintroductiontotext-to-speechsynthesis-230806191602-01ffc67c/85/An-Introduction-To-Text-To-Speech-Synthesis-422-320.jpg)
An Introduction To Text-To-Speech Synthesis
- 1. Text-to-Speech Synthesis Paul Taylor University of Cambridge
- 3. Summary of Contents 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 2 Communication and Language . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 3 The Text-to-Speech Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 4 Text Segmentation and Organisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 5 Text Decoding. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 6 Prosody Prediction from Text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 7 Phonetics and Phonology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 8 Pronunciation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193 9 Synthesis of Prosody . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227 10 Signals and Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265 11 Acoustic Models of Speech Production . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316 12 Analysis of Speech Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349 13 Synthesis Techniques based on Vocal Tract Models . . . . . . . . . . . . . . . . . . . . . . . . . . 396 14 Synthesis by Concatenation and Signal Processing Modification . . . . . . . . . . . . . . 422 15 Hidden Markov Model Synthesis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .446 16 Unit Selection Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 484 17 Further Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528 18 Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 545 iii
- 5. Summary of Contents v Preface. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xix 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.1 What are text-to-speech systems for? . . . . . . . . . . . . . . . . . . . . . . . 2 1.2 What should the goals of text-to-speech system development be? . . . . . . . . 3 1.3 The Engineering Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.4 Overview of the book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 1.4.1 Viewpoints within the book . . . . . . . . . . . . . . . . . . . . . . . 5 1.4.2 Readers’ backgrounds . . . . . . . . . . . . . . . . . . . . . . . . . . 6 1.4.3 Background and specialist sections . . . . . . . . . . . . . . . . . . . 7 2 Communication and Language . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 2.1 Types of communication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 2.1.1 Affective communication . . . . . . . . . . . . . . . . . . . . . . . . . 8 2.1.2 Iconic communication . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2.1.3 Symbolic communication . . . . . . . . . . . . . . . . . . . . . . . . 10 2.1.4 Combinations of symbols . . . . . . . . . . . . . . . . . . . . . . . . 11 2.1.5 Meaning, form and signal . . . . . . . . . . . . . . . . . . . . . . . . 12 2.2 Human Communication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 2.2.1 Verbal communication . . . . . . . . . . . . . . . . . . . . . . . . . . 14 2.2.2 Linguistic levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 2.2.3 Affective Prosody . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.2.4 Augmentative Prosody . . . . . . . . . . . . . . . . . . . . . . . . . . 18 2.3 Communication processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 2.3.1 Communication factors . . . . . . . . . . . . . . . . . . . . . . . . . . 19 2.3.2 Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 2.3.3 Encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2.3.4 Decoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 2.3.5 Understanding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 3 The Text-to-Speech Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 3.1 Speech and Writing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 3.1.1 Physical nature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 3.1.2 Spoken form and written form . . . . . . . . . . . . . . . . . . . . . . 28 3.1.3 Use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 3.1.4 Prosodic and verbal content . . . . . . . . . . . . . . . . . . . . . . . 31 3.1.5 Component balance . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 3.1.6 Non-linguistic content . . . . . . . . . . . . . . . . . . . . . . . . . . 32 3.1.7 Semiotic systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 3.1.8 Writing Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 3.2 Reading aloud . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
- 6. vi Summary of Contents 3.2.1 Reading silently and reading aloud . . . . . . . . . . . . . . . . . . . . 35 3.2.2 Prosody in reading aloud . . . . . . . . . . . . . . . . . . . . . . . . . 36 3.2.3 Verbal content and style in reading aloud . . . . . . . . . . . . . . . . 37 3.3 Text-to-speech system organisation . . . . . . . . . . . . . . . . . . . . . . . . 38 3.3.1 The Common Form model . . . . . . . . . . . . . . . . . . . . . . . . 38 3.3.2 Other models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 3.3.3 Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 3.4 Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 3.4.1 A Simple text-to-speech system . . . . . . . . . . . . . . . . . . . . . 41 3.4.2 Concept to speech . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 3.4.3 Canned Speech and Limited Domain Synthesis . . . . . . . . . . . . . 43 3.5 Key problems in Text-to-speech . . . . . . . . . . . . . . . . . . . . . . . . . . 44 3.5.1 Text classification with respect to semiotic systems . . . . . . . . . . . 44 3.5.2 Decoding natural language text . . . . . . . . . . . . . . . . . . . . . 46 3.5.3 Naturalness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 3.5.4 Intelligibility: encoding the message in signal . . . . . . . . . . . . . . 48 3.5.5 Auxiliary generation for prosody . . . . . . . . . . . . . . . . . . . . . 49 3.5.6 Adapting the system to the situation . . . . . . . . . . . . . . . . . . . 50 3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 4 Text Segmentation and Organisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 4.1 Overview of the problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 4.2 Words and Sentences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 4.2.1 What is a word? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 4.2.2 Defining words in text-to-speech . . . . . . . . . . . . . . . . . . . . . 55 4.2.3 Scope and morphology . . . . . . . . . . . . . . . . . . . . . . . . . . 59 4.2.4 Contractions and Clitics . . . . . . . . . . . . . . . . . . . . . . . . . 60 4.2.5 Slang forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 4.2.6 Hyphenated forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 4.2.7 What is a sentence? . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 4.2.8 The lexicon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 4.3 Text Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 4.3.1 Tokenisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 4.3.2 Tokenisation and Punctuation . . . . . . . . . . . . . . . . . . . . . . 65 4.3.3 Tokenisation Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . 66 4.3.4 Sentence Splitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 4.4 Processing Documents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 4.4.1 Markup Languages . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 4.4.2 Interpreting characters . . . . . . . . . . . . . . . . . . . . . . . . . . 70 4.5 Text-to-Speech Architectures . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 4.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
- 7. Summary of Contents vii 4.6.1 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 4.6.2 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 5 Text Decoding: Finding the words from the text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 5.1 Overview of Text Decoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 5.2 Text Classification Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . 80 5.2.1 Features and algorithms . . . . . . . . . . . . . . . . . . . . . . . . . 80 5.2.2 Tagging and word sense disambiguation . . . . . . . . . . . . . . . . . 83 5.2.3 Ad-hoc approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84 5.2.4 Deterministic rule approaches . . . . . . . . . . . . . . . . . . . . . . 84 5.2.5 Decision lists . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 5.2.6 Naive Bayes Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . 87 5.2.7 Decision trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 5.2.8 Part-of-speech Tagging . . . . . . . . . . . . . . . . . . . . . . . . . . 89 5.3 Non-Natural Language Text . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 5.3.1 Semiotic Classification . . . . . . . . . . . . . . . . . . . . . . . . . . 93 5.3.2 Semiotic Decoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 5.3.3 Verbalisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 5.4 Natural Language Text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 5.4.1 Acronyms and letter sequences . . . . . . . . . . . . . . . . . . . . . 100 5.4.2 Homograph disambiguation . . . . . . . . . . . . . . . . . . . . . . . 101 5.4.3 Non-homographs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 5.5 Natural Language Parsing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 5.5.1 Context Free Grammars . . . . . . . . . . . . . . . . . . . . . . . . . 103 5.5.2 Statistical Parsing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 5.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 5.6.1 Further reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 5.6.2 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 6 Prosody Prediction from Text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 6.1 Prosodic Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 6.2 Phrasing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 6.2.1 Phrasing Phenomena . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 6.2.2 Models of Phrasing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 6.3 Prominence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 6.3.1 Syntactic prominence patterns . . . . . . . . . . . . . . . . . . . . . . 117 6.3.2 Discourse prominence patterns . . . . . . . . . . . . . . . . . . . . . . 119 6.3.3 Prominence systems, data and labelling . . . . . . . . . . . . . . . . . 120 6.4 Intonation and tune . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 6.5 Prosodic Meaning and Function . . . . . . . . . . . . . . . . . . . . . . . . . . 123 6.5.1 Affective Prosody . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 6.5.2 Suprasegmental . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
- 8. viii Summary of Contents 6.5.3 Augmentative Prosody . . . . . . . . . . . . . . . . . . . . . . . . . . 125 6.5.4 Symbolic communication and prosodic style . . . . . . . . . . . . . . 127 6.6 Determining Prosody from the Text . . . . . . . . . . . . . . . . . . . . . . . . 128 6.6.1 Prosody and human reading . . . . . . . . . . . . . . . . . . . . . . . 128 6.6.2 Controlling the degree of augmentative prosody . . . . . . . . . . . . . 129 6.6.3 Prosody and synthesis techniques . . . . . . . . . . . . . . . . . . . . 129 6.7 Phrasing prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 6.7.1 Experimental formulation . . . . . . . . . . . . . . . . . . . . . . . . 130 6.7.2 Deterministic approaches . . . . . . . . . . . . . . . . . . . . . . . . . 131 6.7.3 Classifier approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . 133 6.7.4 HMM approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 6.7.5 Hybrid approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 6.8 Prominence Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 6.8.1 Compound noun phrases . . . . . . . . . . . . . . . . . . . . . . . . . 137 6.8.2 Function word prominence . . . . . . . . . . . . . . . . . . . . . . . . 139 6.8.3 Data driven approaches . . . . . . . . . . . . . . . . . . . . . . . . . . 139 6.9 Intonational Tune Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 6.10 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 6.10.1 Labelling schemes and labelling accuracy . . . . . . . . . . . . . . . . 140 6.10.2 Linguistic theories and prosody . . . . . . . . . . . . . . . . . . . . . 142 6.10.3 Synthesising suprasegmental and true prosody . . . . . . . . . . . . . 143 6.10.4 Prosody in real dialogues . . . . . . . . . . . . . . . . . . . . . . . . . 144 6.10.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 6.10.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 7 Phonetics and Phonology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 7.1 Articulatory phonetics and speech production . . . . . . . . . . . . . . . . . . 147 7.1.1 The vocal organs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 7.1.2 Sound sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 7.1.3 Sound output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 7.1.4 The vocal tract filter . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 7.1.5 Vowels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 7.1.6 Consonants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 7.1.7 Examining speech production . . . . . . . . . . . . . . . . . . . . . . 157 7.2 Acoustics phonetics and speech perception . . . . . . . . . . . . . . . . . . . . 158 7.2.1 Acoustic representations . . . . . . . . . . . . . . . . . . . . . . . . . 159 7.2.2 Acoustic characteristics . . . . . . . . . . . . . . . . . . . . . . . . . 161 7.3 The communicative use of speech . . . . . . . . . . . . . . . . . . . . . . . . . 162 7.3.1 Communicating discrete information with a continuous channel . . . . 162 7.3.2 Phonemes, phones and allophones . . . . . . . . . . . . . . . . . . . . 164 7.3.3 Allophonic variation and phonetic context . . . . . . . . . . . . . . . . 168
- 9. Summary of Contents ix 7.3.4 Coarticulation, targets and transients . . . . . . . . . . . . . . . . . . . 169 7.3.5 The continuous nature of speech . . . . . . . . . . . . . . . . . . . . . 170 7.3.6 Transcription . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171 7.3.7 The distinctiveness of speech in communication . . . . . . . . . . . . 173 7.4 Phonology: the linguistic organisation of speech . . . . . . . . . . . . . . . . . 173 7.4.1 Phonotactics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174 7.4.2 Word formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180 7.4.3 Distinctive Features and Phonological Theories . . . . . . . . . . . . . 182 7.4.4 Syllables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185 7.4.5 Lexical Stress . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187 7.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190 7.5.1 Further reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190 7.5.2 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191 8 Pronunciation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193 8.1 Pronunciation representations . . . . . . . . . . . . . . . . . . . . . . . . . . . 193 8.1.1 Why bother? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193 8.1.2 Phonemic and phonetic input . . . . . . . . . . . . . . . . . . . . . . . 194 8.1.3 Difficulties in deriving phonetic input . . . . . . . . . . . . . . . . . . 195 8.1.4 A Structured approach to pronunciation . . . . . . . . . . . . . . . . . 196 8.1.5 Abstract phonological representations . . . . . . . . . . . . . . . . . . 197 8.2 Formulating a phonological representation system . . . . . . . . . . . . . . . . 198 8.2.1 Simple consonants and vowels . . . . . . . . . . . . . . . . . . . . . . 198 8.2.2 Difficult consonants . . . . . . . . . . . . . . . . . . . . . . . . . . . 200 8.2.3 Diphthongs and affricates . . . . . . . . . . . . . . . . . . . . . . . . 201 8.2.4 Approximant-vowel combinations . . . . . . . . . . . . . . . . . . . . 202 8.2.5 Defining the full inventory . . . . . . . . . . . . . . . . . . . . . . . . 203 8.2.6 Phoneme names . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205 8.2.7 Syllabic issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207 8.3 The Lexicon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208 8.3.1 Lexicon and Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209 8.3.2 Lexicon formats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211 8.3.3 The offline lexicon . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214 8.3.4 The system lexicon . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215 8.3.5 Lexicon quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216 8.3.6 Determining the pronunciation of unknown words . . . . . . . . . . . 217 8.4 Grapheme-to-Phoneme Conversion . . . . . . . . . . . . . . . . . . . . . . . . 219 8.4.1 Rule based techniques . . . . . . . . . . . . . . . . . . . . . . . . . . 219 8.4.2 Grapheme to phoneme alignment . . . . . . . . . . . . . . . . . . . . 220 8.4.3 Neural networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220 8.4.4 Pronunciation by analogy . . . . . . . . . . . . . . . . . . . . . . . . 221
- 10. x Summary of Contents 8.4.5 Other data driven techniques . . . . . . . . . . . . . . . . . . . . . . . 222 8.4.6 Statistical Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . 222 8.5 Further Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223 8.5.1 Morphology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223 8.5.2 Language origin and names . . . . . . . . . . . . . . . . . . . . . . . 224 8.5.3 Post-lexical processing . . . . . . . . . . . . . . . . . . . . . . . . . . 224 8.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225 9 Synthesis of Prosody . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227 9.1 Intonation Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227 9.1.1 F0 and pitch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228 9.1.2 Intonational form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228 9.1.3 Models of F0 contours . . . . . . . . . . . . . . . . . . . . . . . . . . 230 9.1.4 Micro-prosody . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231 9.2 Intonational Behaviour . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231 9.2.1 Intonational tune . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232 9.2.2 Downdrift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233 9.2.3 Pitch Range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235 9.2.4 Pitch Accents and Boundary tones . . . . . . . . . . . . . . . . . . . . 237 9.3 Intonation Theories and Models . . . . . . . . . . . . . . . . . . . . . . . . . . 239 9.3.1 Traditional models and the British school . . . . . . . . . . . . . . . . 239 9.3.2 The Dutch school . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239 9.3.3 Autosegmental-Metrical and ToBI models . . . . . . . . . . . . . . . . 240 9.3.4 The INTSINT Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 241 9.3.5 The Fujisaki model and Superimpositional Models . . . . . . . . . . . 242 9.3.6 The Tilt model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244 9.3.7 Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246 9.4 Intonation Synthesis with AM models . . . . . . . . . . . . . . . . . . . . . . 248 9.4.1 Prediction of AM labels from text . . . . . . . . . . . . . . . . . . . . 248 9.4.2 Deterministic synthesis methods . . . . . . . . . . . . . . . . . . . . . 249 9.4.3 Data Driven synthesis methods . . . . . . . . . . . . . . . . . . . . . . 250 9.4.4 Analysis with Autosegmental models . . . . . . . . . . . . . . . . . . 250 9.5 Intonation Synthesis with Deterministic Acoustic Models . . . . . . . . . . . . 251 9.5.1 Synthesis with superimpositional models . . . . . . . . . . . . . . . . 251 9.5.2 Synthesis with the Tilt model . . . . . . . . . . . . . . . . . . . . . . 252 9.5.3 Analysis with Fujisaki and Tilt models . . . . . . . . . . . . . . . . . 252 9.6 Data Driven Intonation Models . . . . . . . . . . . . . . . . . . . . . . . . . . 252 9.6.1 Unit selection style approaches . . . . . . . . . . . . . . . . . . . . . 253 9.6.2 Dynamic System Models . . . . . . . . . . . . . . . . . . . . . . . . . 254 9.6.3 Hidden Markov models . . . . . . . . . . . . . . . . . . . . . . . . . 255 9.6.4 Functional models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
- 11. Summary of Contents xi 9.7 Timing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257 9.7.1 Formulation of the timing problem . . . . . . . . . . . . . . . . . . . . 257 9.7.2 The nature of timing . . . . . . . . . . . . . . . . . . . . . . . . . . . 258 9.7.3 Klatt rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259 9.7.4 Sums of products model . . . . . . . . . . . . . . . . . . . . . . . . . 260 9.7.5 The Campbell model . . . . . . . . . . . . . . . . . . . . . . . . . . . 260 9.7.6 Other regression techniques . . . . . . . . . . . . . . . . . . . . . . . 261 9.8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261 9.8.1 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262 9.8.2 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263 10 Signals and Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265 10.1 Analogue signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265 10.1.1 Simple periodic signals: sinusoids . . . . . . . . . . . . . . . . . . . . 266 10.1.2 General periodic signals . . . . . . . . . . . . . . . . . . . . . . . . . 268 10.1.3 Sinusoids as complex exponentials . . . . . . . . . . . . . . . . . . . . 270 10.1.4 Fourier Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272 10.1.5 Frequency domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275 10.1.6 The Fourier transform . . . . . . . . . . . . . . . . . . . . . . . . . . 278 10.2 Digital signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283 10.2.1 Digital waveforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283 10.2.2 Digital representations . . . . . . . . . . . . . . . . . . . . . . . . . . 284 10.2.3 The discrete-time Fourier transform . . . . . . . . . . . . . . . . . . . 284 10.2.4 The discrete Fourier transform . . . . . . . . . . . . . . . . . . . . . . 285 10.2.5 The z-Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286 10.2.6 The frequency domain for digital signals . . . . . . . . . . . . . . . . 288 10.3 Properties of Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288 10.3.1 Linearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288 10.3.2 Time and Frequency Duality . . . . . . . . . . . . . . . . . . . . . . . 289 10.3.3 Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289 10.3.4 Impulse Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289 10.3.5 Time delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290 10.3.6 Frequency shift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291 10.3.7 Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291 10.3.8 Analytical and Numerical Analysis . . . . . . . . . . . . . . . . . . . 292 10.3.9 Stochastic Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292 10.4 Digital Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292 10.4.1 Difference Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 293 10.4.2 The impulse response . . . . . . . . . . . . . . . . . . . . . . . . . . 294 10.4.3 Filter convolution sum . . . . . . . . . . . . . . . . . . . . . . . . . . 296 10.4.4 Filter transfer function . . . . . . . . . . . . . . . . . . . . . . . . . . 297
- 12. xii Summary of Contents 10.4.5 The transfer function and the impulse response . . . . . . . . . . . . . 298 10.5 Digital filter analysis and design . . . . . . . . . . . . . . . . . . . . . . . . . . 299 10.5.1 Polynomial analysis: poles and zeros . . . . . . . . . . . . . . . . . . 299 10.5.2 Frequency Interpretation of z-domain transfer function . . . . . . . . . 302 10.5.3 Filter characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . 304 10.5.4 Putting it all together . . . . . . . . . . . . . . . . . . . . . . . . . . . 310 10.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313 11 Acoustic Models of Speech Production . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316 11.1 Acoustic Theory of Speech Production . . . . . . . . . . . . . . . . . . . . . . 316 11.1.1 Components in the model . . . . . . . . . . . . . . . . . . . . . . . . 317 11.2 The physics of sound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318 11.2.1 Resonant systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318 11.2.2 Travelling waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321 11.2.3 Acoustic waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323 11.2.4 Acoustic reflection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325 11.3 Vowel Tube Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326 11.3.1 Discrete time and distance . . . . . . . . . . . . . . . . . . . . . . . . 327 11.3.2 Junction of two tubes . . . . . . . . . . . . . . . . . . . . . . . . . . . 328 11.3.3 Special cases of junction . . . . . . . . . . . . . . . . . . . . . . . . . 330 11.3.4 Two tube vocal tract model . . . . . . . . . . . . . . . . . . . . . . . . 331 11.3.5 Single tube model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333 11.3.6 Multi-tube vocal tract model . . . . . . . . . . . . . . . . . . . . . . . 335 11.3.7 The all pole resonator model . . . . . . . . . . . . . . . . . . . . . . . 337 11.4 Source and radiation models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338 11.4.1 Radiation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338 11.4.2 Glottal source . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338 11.5 Model refinements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341 11.5.1 Modelling the nasal Cavity . . . . . . . . . . . . . . . . . . . . . . . . 341 11.5.2 Source positions in the oral cavity . . . . . . . . . . . . . . . . . . . . 343 11.5.3 Models with Vocal Tract Losses . . . . . . . . . . . . . . . . . . . . . 344 11.5.4 Source and radiation effects . . . . . . . . . . . . . . . . . . . . . . . 344 11.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345 11.6.1 Further reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347 11.6.2 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347 12 Analysis of Speech Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349 12.1 Short term speech analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350 12.1.1 Windowing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350 12.1.2 Short term spectral representations . . . . . . . . . . . . . . . . . . . . 351 12.1.3 Frame lengths and shifts . . . . . . . . . . . . . . . . . . . . . . . . . 353 12.1.4 The spectrogram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
- 13. Summary of Contents xiii 12.1.5 Auditory scales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358 12.2 Filter bank analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359 12.3 The Cepstrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360 12.3.1 Cepstrum definition . . . . . . . . . . . . . . . . . . . . . . . . . . . 360 12.3.2 Treating the magnitude spectrum as a signal . . . . . . . . . . . . . . . 361 12.3.3 Cepstral analysis as deconvolution . . . . . . . . . . . . . . . . . . . . 362 12.3.4 Cepstral analysis discussion . . . . . . . . . . . . . . . . . . . . . . . 363 12.4 Linear prediction analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364 12.4.1 Finding the coefficients: the covariance method . . . . . . . . . . . . . 365 12.4.2 Autocorrelation Method . . . . . . . . . . . . . . . . . . . . . . . . . 367 12.4.3 Levinson Durbin Recursion . . . . . . . . . . . . . . . . . . . . . . . 369 12.5 Spectral envelope and vocal tract representations . . . . . . . . . . . . . . . . . 370 12.5.1 Linear prediction spectra . . . . . . . . . . . . . . . . . . . . . . . . . 370 12.5.2 Transfer function poles . . . . . . . . . . . . . . . . . . . . . . . . . . 372 12.5.3 Reflection coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . 372 12.5.4 Log area ratios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375 12.5.5 Line spectrum frequencies . . . . . . . . . . . . . . . . . . . . . . . . 375 12.5.6 Linear prediction cepstrum . . . . . . . . . . . . . . . . . . . . . . . . 377 12.5.7 Mel-scaled cepstrum . . . . . . . . . . . . . . . . . . . . . . . . . . . 378 12.5.8 Perceptual linear prediction . . . . . . . . . . . . . . . . . . . . . . . 378 12.5.9 Formant tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378 12.6 Source representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380 12.6.1 Residual signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380 12.6.2 Closed-phase analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 383 12.6.3 Open-phase analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 385 12.6.4 Impulse/noise models . . . . . . . . . . . . . . . . . . . . . . . . . . 386 12.6.5 Parameterisation of glottal flow signal . . . . . . . . . . . . . . . . . . 387 12.7 Pitch and epoch detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388 12.7.1 Pitch detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388 12.7.2 Epoch detection: finding the instant of glottal closure . . . . . . . . . . 390 12.8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393 12.8.1 Further reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394 12.8.2 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394 13 Synthesis Techniques Based on Vocal Tract Models . . . . . . . . . . . . . . . . . . . . . . . . . . 396 13.1 Synthesis specification: the input to the synthesiser . . . . . . . . . . . . . . . . 396 13.2 Formant Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397 13.2.1 Sound sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 398 13.2.2 Synthesising a single formant . . . . . . . . . . . . . . . . . . . . . . 399 13.2.3 Resonators in series and parallel . . . . . . . . . . . . . . . . . . . . . 400 13.2.4 Synthesising consonants . . . . . . . . . . . . . . . . . . . . . . . . . 402
- 14. xiv Summary of Contents 13.2.5 Complete synthesiser . . . . . . . . . . . . . . . . . . . . . . . . . . . 403 13.2.6 The phonetic input to the synthesiser . . . . . . . . . . . . . . . . . . 405 13.2.7 Formant synthesis quality . . . . . . . . . . . . . . . . . . . . . . . . 407 13.3 Classical Linear Prediction Synthesis . . . . . . . . . . . . . . . . . . . . . . . 408 13.3.1 Comparison with formant synthesis . . . . . . . . . . . . . . . . . . . 409 13.3.2 Impulse/noise source model . . . . . . . . . . . . . . . . . . . . . . . 410 13.3.3 Linear prediction diphone concatenative synthesis . . . . . . . . . . . 411 13.3.4 Complete synthesiser . . . . . . . . . . . . . . . . . . . . . . . . . . . 413 13.3.5 Problems with the source . . . . . . . . . . . . . . . . . . . . . . . . . 414 13.4 Articulatory synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415 13.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417 13.5.1 Further reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419 13.5.2 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 420 14 Synthesis by Concatenation and Signal Processing Modification . . . . . . . . . . . . . . 422 14.1 Speech units in second generation systems . . . . . . . . . . . . . . . . . . . . 423 14.1.1 Creating a diphone inventory . . . . . . . . . . . . . . . . . . . . . . . 424 14.1.2 Obtaining diphones from speech . . . . . . . . . . . . . . . . . . . . . 425 14.2 Pitch synchronous overlap and add (PSOLA) . . . . . . . . . . . . . . . . . . . 426 14.2.1 Time domain PSOLA . . . . . . . . . . . . . . . . . . . . . . . . . . 426 14.2.2 Epoch manipulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 427 14.2.3 How does PSOLA work? . . . . . . . . . . . . . . . . . . . . . . . . . 430 14.3 Residual excited linear prediction . . . . . . . . . . . . . . . . . . . . . . . . . 433 14.3.1 Residual manipulation . . . . . . . . . . . . . . . . . . . . . . . . . . 434 14.3.2 Linear Prediction PSOLA . . . . . . . . . . . . . . . . . . . . . . . . 434 14.4 Sinusoidal models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435 14.4.1 Pure sinusoidal models . . . . . . . . . . . . . . . . . . . . . . . . . . 435 14.4.2 Harmonic/Noise Models . . . . . . . . . . . . . . . . . . . . . . . . . 437 14.5 MBROLA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440 14.6 Synthesis from Cepstral Coefficients . . . . . . . . . . . . . . . . . . . . . . . 440 14.7 Concatenation Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442 14.8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444 14.8.1 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444 14.8.2 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444 15 Hidden Markov Model Synthesis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .446 15.1 The HMM formalism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447 15.1.1 Observation probabilities . . . . . . . . . . . . . . . . . . . . . . . . . 447 15.1.2 Delta coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450 15.1.3 Acoustic representations and covariance . . . . . . . . . . . . . . . . . 450 15.1.4 States and transitions . . . . . . . . . . . . . . . . . . . . . . . . . . . 452 15.1.5 Recognising with HMMs . . . . . . . . . . . . . . . . . . . . . . . . . 452
- 15. Summary of Contents xv 15.1.6 Language models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455 15.1.7 The Viterbi algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 456 15.1.8 Training HMMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459 15.1.9 Context-sensitive modelling . . . . . . . . . . . . . . . . . . . . . . . 463 15.1.10 Are HMMs a good model of speech? . . . . . . . . . . . . . . . . . . 467 15.2 Synthesis from hidden Markov models . . . . . . . . . . . . . . . . . . . . . . 468 15.2.1 Finding the most likely observations given the state sequence . . . . . 469 15.2.2 Finding the most likely observations and state sequence . . . . . . . . 471 15.2.3 Acoustic representations . . . . . . . . . . . . . . . . . . . . . . . . . 474 15.2.4 Context sensitive synthesis models . . . . . . . . . . . . . . . . . . . . 475 15.2.5 Duration modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 476 15.2.6 HMM synthesis systems . . . . . . . . . . . . . . . . . . . . . . . . . 476 15.3 Labelling databases with HMMs . . . . . . . . . . . . . . . . . . . . . . . . . 477 15.3.1 Determining the word sequence . . . . . . . . . . . . . . . . . . . . . 477 15.3.2 Determining the phone sequence . . . . . . . . . . . . . . . . . . . . . 478 15.3.3 Determining the phone boundaries . . . . . . . . . . . . . . . . . . . . 478 15.3.4 Measuring the quality of the alignments . . . . . . . . . . . . . . . . . 480 15.4 Other data driven synthesis techniques . . . . . . . . . . . . . . . . . . . . . . 481 15.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481 15.5.1 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481 15.5.2 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482 16 Unit Selection Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 484 16.1 From Concatenative Synthesis to Unit Selection . . . . . . . . . . . . . . . . . 484 16.1.1 Extending concatenative synthesis . . . . . . . . . . . . . . . . . . . . 485 16.1.2 The Hunt and Black Algorithm . . . . . . . . . . . . . . . . . . . . . 488 16.2 Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489 16.2.1 Base Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489 16.2.2 Linguistic and Acoustic features . . . . . . . . . . . . . . . . . . . . . 491 16.2.3 Choice of features . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492 16.2.4 Types of features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493 16.3 The Independent Feature Target Function Formulation . . . . . . . . . . . . . . 494 16.3.1 The purpose of the target function . . . . . . . . . . . . . . . . . . . . 494 16.3.2 Defining a perceptual space . . . . . . . . . . . . . . . . . . . . . . . 496 16.3.3 Perceptual spaces defined by independent features . . . . . . . . . . . 496 16.3.4 Setting the target weights using acoustic distances . . . . . . . . . . . 498 16.3.5 Limitations of the independent feature formulation . . . . . . . . . . . 502 16.4 The Acoustic Space Target Function Formulation . . . . . . . . . . . . . . . . 503 16.4.1 Decision tree clustering . . . . . . . . . . . . . . . . . . . . . . . . . 504 16.4.2 General partial-synthesis functions . . . . . . . . . . . . . . . . . . . . 506 16.5 Join functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 508
- 16. xvi Summary of Contents 16.5.1 Basic issues in joining units . . . . . . . . . . . . . . . . . . . . . . . 508 16.5.2 Phone class join costs . . . . . . . . . . . . . . . . . . . . . . . . . . 509 16.5.3 Acoustic distance join costs . . . . . . . . . . . . . . . . . . . . . . . 510 16.5.4 Combining categorical and and acoustic join costs . . . . . . . . . . . 511 16.5.5 Probabilistic and sequence join costs . . . . . . . . . . . . . . . . . . 512 16.5.6 Join classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514 16.6 Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515 16.6.1 Base Types and Search . . . . . . . . . . . . . . . . . . . . . . . . . . 516 16.6.2 Pruning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519 16.6.3 Pre-selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 520 16.6.4 Beam Pruning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 520 16.6.5 Multi-pass search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 520 16.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 521 16.7.1 Unit selection and signal processing . . . . . . . . . . . . . . . . . . . 522 16.7.2 Features, costs and perception . . . . . . . . . . . . . . . . . . . . . . 523 16.7.3 Example unit selection systems . . . . . . . . . . . . . . . . . . . . . 524 16.7.4 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526 16.7.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526 17 Further Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528 17.1 Databases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528 17.1.1 Unit Selection Databases . . . . . . . . . . . . . . . . . . . . . . . . . 528 17.1.2 Text materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 529 17.1.3 Prosody databases . . . . . . . . . . . . . . . . . . . . . . . . . . . . 530 17.1.4 Labelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 530 17.1.5 What exactly is hand labelling? . . . . . . . . . . . . . . . . . . . . . 531 17.1.6 Automatic labelling . . . . . . . . . . . . . . . . . . . . . . . . . . . 532 17.1.7 Avoiding explicit labels . . . . . . . . . . . . . . . . . . . . . . . . . 532 17.2 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533 17.2.1 System Testing: Intelligibility and Naturalness . . . . . . . . . . . . . 534 17.2.2 Word recognition tests . . . . . . . . . . . . . . . . . . . . . . . . . . 534 17.2.3 Naturalness tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 535 17.2.4 Test data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536 17.2.5 Unit or Component testing . . . . . . . . . . . . . . . . . . . . . . . . 536 17.2.6 Competitive evaluations . . . . . . . . . . . . . . . . . . . . . . . . . 538 17.3 Audiovisual Speech Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 538 17.3.1 Speech Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 539 17.4 Synthesis of Emotional and Expressive Speech . . . . . . . . . . . . . . . . . . 540 17.4.1 Describing Emotion . . . . . . . . . . . . . . . . . . . . . . . . . . . 540 17.4.2 Synthesizing emotion with prosody control . . . . . . . . . . . . . . . 541 17.4.3 Synthesizing emotion with voice transformation . . . . . . . . . . . . 542
- 17. Summary of Contents xvii 17.4.4 Unit selection and HMM techniques . . . . . . . . . . . . . . . . . . . 542 17.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 543 18 Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 545 18.1 Speech Technology and Linguistics . . . . . . . . . . . . . . . . . . . . . . . . 545 18.2 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 548 18.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 550 Appendix 552 A Probability. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .552 A.1 Discrete Probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 552 A.1.1 Discrete Random Variables . . . . . . . . . . . . . . . . . . . . . . . . 552 A.1.2 Probability Mass Function . . . . . . . . . . . . . . . . . . . . . . . . 553 A.1.3 Expected Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553 A.1.4 Moments of a PMF . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554 A.2 Pairs of discrete random variables . . . . . . . . . . . . . . . . . . . . . . . . . 554 A.2.1 Marginal Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . 555 A.2.2 Independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 555 A.2.3 Expected Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 556 A.2.4 Moments of a joint distribution . . . . . . . . . . . . . . . . . . . . . 556 A.2.5 Higher-Order Moments and covariance . . . . . . . . . . . . . . . . . 556 A.2.6 Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557 A.2.7 Conditional Probability . . . . . . . . . . . . . . . . . . . . . . . . . . 557 A.2.8 Bayes’ Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 558 A.2.9 Sum of Random Variables . . . . . . . . . . . . . . . . . . . . . . . . 558 A.2.10 The chain rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 559 A.2.11 Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 559 A.3 Continuous Random Variables . . . . . . . . . . . . . . . . . . . . . . . . . . 560 A.3.1 Continuous Random Variables . . . . . . . . . . . . . . . . . . . . . . 560 A.3.2 Expected Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 562 A.3.3 Gaussian Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . 562 A.3.4 Uniform Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . 562 A.3.5 Cumulative Density Functions . . . . . . . . . . . . . . . . . . . . . . 563 A.4 Pairs of Continuous Random Variables . . . . . . . . . . . . . . . . . . . . . . 563 A.4.1 Independence vs Uncorrelated . . . . . . . . . . . . . . . . . . . . . . 564 A.4.2 Sum of Two Random Variables . . . . . . . . . . . . . . . . . . . . . 565 A.4.3 Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 565 A.4.4 Kullback-Leibler Distance . . . . . . . . . . . . . . . . . . . . . . . . 565 B Phone Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567
- 18. This book is dedicated to the technical team at Rhetorical Systsms.
- 19. Foreward Speech processing technology has been a mainstream area of research for more than 50 years. The ultimate goal of speech research is to build systems that mimic (or potentially surpass) human capabilities in understanding, generating and coding speech for a range of human-to-human and human-to-machine interactions. In the area of speech coding a great deal of success has been achieved in creating systems that significantly reduce the overall bit rate of the speech signal (from order of 100 kilobits per second, to rates on the order of 8 kilobits per second or less), while maintaining speech intelligibility and quality at levels appropriate for the intended applications. The heart of the modern cellular industry is the 8 kilobit per second speech coder, embedded in VLSI logic on the more than 2 billion cellphones in use worldwide at the end of 2007. In the area of speech recognition and understanding by machines, steady progress has en- abled systems to become part of everyday life in the form of call centers for the airlines, financial, medical and banking industries, help desks for large businesses, form and report generation for the legal and medical communities, and dictation machines that enable individuals to enter text into machines without having to explicitly type the text. Such speech recognition systems were made available to the general public as long as 15 years ago (in 1992 AT&T introduced the Voice Recognition Call Processing system which automated operator-assisted calls, handling more than 1.2 billion requests each year with error rates below 0.5penetrated almost every major industry since that time. Simple speech understanding systems have also been introduced into the mar- ketplace and have had varying degrees of success for help desks (e.g., the How May I Help You system introduced by AT&T for customer care applications) and for stock trading applications (IBM system), among others. It has been the area of speech generation that has been the hardest speech technology area to obtain any viable degree of success. For more than 50 years researchers have struggled with the problem of trying to mimic the physical processes of speech generation via articulatory models of the human vocal tract, or via terminal analog synthesis models of the time-varying spectral and temporal properties of speech. In spite of the best efforts of some outstanding speech researchers, the quality of synthetic speech generated by machine was unnatural most of the time and has been unacceptable for human use in most real world applications. In the late 1970s the idea of gener- ating speech by concatenating basic speech units (in most cases diphone units which represented pieces of pairs of phonemes) was investigated and shown to be practical once researchers learned how to reliably excise diphones from human speech. After more than a decade of investigation as to how to optimally concatenate diphones, the resulting synthetic speech was often highly intelli- gible (a big improvement over earlier systems) but regrettably remained highly unnatural. Hence concatenative speech synthesis systems remained lab curiosities but were not employed in real world applications such as reading email, user interactions in dialogue systems, etc. The really big breakthrough in speech synthesis came in the late 1980s when Yoshinori Sagisaka at ATR in Japan made the leap from single diphone tokens as the basic unit set for speech synthesis, to multiple di- xix
- 20. xx Preface phone tokens, extracted from carefully designed and read speech databases. Sagisaka realized that in the limiting case, where you had thousands of tokens of each possible diphone of the English language, you could literally concatenate the “correct” sequence of diphones and produce natural sounding human speech. The new problem that arose was deciding exactly which of the thousands of diphones should be used at each diphone position in the speech being generated. History has shown that, like most large scale computing problems, there are solutions that make the search for the optimum sequence of diphones (from a large, virtually infinite database) possible in reasonable time and memory. The rest is now history as a new generation of speech researchers investigated virtually every aspect of the so-called “unit selection” method of concatenative speech synthesis, showing that high quality (both intelligibility and naturalness) synthetic speech could be obtained from such systems for virtually any task application. Once the problem of generating natural sounding speech from a sequence of diphones was solved (in the sense that a practical demonstration of the feasibility of such high quality synthesis was made with unit selection synthesis systems), the remaining long-standing problem was the conversion from ordinary printed text to the proper sequence of diphones, along with associated prosodic information about sound duration, loudness, emphasis, pitch, pauses, and other so-called suprasegmental aspects of speech. The problem of converting from text to a complete linguistic description of associated sound was one that has been studied almost as long as synthesis itself; much progress had been made in almost every aspect of the linguistic description of speech as in the acoustic generation of high quality sounds. It is the success of unit selection speech synthesis systems that has motivated the research of Paul Taylor, the author of this book on text-to-speech synthesis systems. Paul Taylor has been in the thick of the research in speech synthesis systems for more than 15 years, having worked at ATR in Japan on the CHATR synthesizer (the system that actually demonstrated near perfect speech quality on some subset of the sentences that were input), at the Centre for Speech Technology research at the University of Edinburgh on the Festival synthesis system, and as Chief Technical Officer of Rhetorical Systems, also in Edinburgh. Based on decades of research and the extraordinary progress over the past decade, Taylor has put together a book which attempts to tie it all together and to document and explain the processes involved in a complete text-to-speech synthesis system. The first nine chapters of the book ad- dress the problem of converting printed text to a sequence of sound units (which characterize the acoustic properties of the resulting synthetic sentence), and an accompanying description of the associated prosody which is most appropriate for the sentence being spoken. The remaining eight chapters (not including the conclusion) provide a review of the associated signal processing tech- niques for representing speech units and for seamlessly tying them together to form intelligible and natural speech sounds. This is followed by a discussion of the three generations of synthesis methods, namely articulatory and terminal analog synthesis methods, simple concatenative meth- ods using a single representation for each diphone unit, and the unit selection method based on multiple representations for each diphone unit. There is a single chapter devoted to a promising new synthesis approach, namely a statistical method based on the popular Hidden Markov Model
- 21. Preface xxi (HMM) formulation used in speech recognition systems. According to the author, “Speech synthesis has progressed remarkably in recent years, and it is no longer the case that state-of-the-art systems sound overtly mechanical and robotic”. Although this statement is true, there remains a great deal more to be accomplished before speech synthesis systems are indistinguishable from a human speaker. Perhaps the most glaring need is expressive synthesis that not only imparts the message corresponding to the printed text input, but imparts the emotion associated with the way a human might speak the same sentence in the context of a dialogue with another human being. We are still a long way from such emotional or expressive speech synthesis systems. This book is a wonderful addition to the literature in speech processing and will be a must- read for anyone wanting to understand the blossoming field of text-to-speech synthesis. Lawrence Rabiner August 2007
- 23. Preface I’d like to say that Text-to-Speech Synthesis was years in the planning but nothing could be further from the truth. In mid 2004, as Rhetorical Systems was nearing its end, I suddenly found myself with spare time on my hands for the first time since..., well, ever to be honest. I had no clear idea of what to do next and thought it might “serve the community well” to jot down a few thoughts on TTS. The initial idea was a slim “hardcore” technical book explaining the state of the art, but as writing continued I realised that more and more background material was needed. Eventually the initial idea was dropped and the more comprehensive volume that you now see was written. Some early notes were made in Edinburgh, but the book mainly started during a summer I spent in Boulder Colorado. I was freelancing at that point but am grateful to Ron Cole and his group for accomodating me in such a friendly fashion at that point. A good deal of work was completed outside the office and I am enternally grateful to Laura Michalis for putting me up, putting up with me and generally being there and supportive during that phase. While the book was in effect written entirely by myself, I should pay particular thanks to Steve Young and Mark Gales as I used the HTK book and lecture notes directly in Chapter 15 and the appendix. Many have helped by reading earlier drafts, especially Bob Ladd, Keiichi Tokuda, Rochard Sproat, Dan Jurafsky, Ian Hodson, Ant Tomlinson, Matthew Aylett and Rob Clark. From October 2004 to August 2006 I was a visitor at the Cambridge University Engineering depertment, and it was there that the bulk of the book was written. I am very grateful to Steve Young in particular for taking me in and making me feel part of what was a very well established group. Without Steve’s help Cambridge wouldn’t have happended for me and the book probably would have not been written. Within the department the other faculty staff made me feel very welcome. Phil Woodland and Mark Gales always showed an interest and were of great help on many issues. Finally, Bill Byrne arrived at Cambridge at the same time as me, and proved a great comrade in our quest to understand how on all earth Cambridge University actually worked. Since, he has become a good friend and was a great encouragment in the writing of this book. All these and the other members of the Machine Intelligence lab have become good friends as well as respected colleagues. To my long suffering room mate Gabe Brostow I simply say thanks for being a good friend. Because of the strength of our friendship I know he won’t mind me having the last word in a long running argument and saying that Ireland really is a better place than Texas. It is customary in acknowlegdments such as these to thank one’s wife and family; unfortu- natley while I may have some skill in TTS, this has (strangely) not transfered into skill in holding down a relationship. None the less, no man is an island and I’d like to thank Laura, Kirstin, and Kayleigh for being there while I wrote the book. Final and special thanks must go to Maria Founda, who has a generosity of spirit second to none. For most of my career I have worked in Edinburgh, first at the University’s Centre for Speech Technology Research (CSTR) and laterly at Rhetorical Systems. The book reflects the climate in both those organisations which, to my mind, had a healthy balance between eclectic knowledge xxiii
- 24. xxiv Preface of engineering, computer science and linguistics on one hand and robust system building on the other. There are far too many individuals there to thank personally, but I should mention Bob Ladd, Alan Wrench and Steve Isard as being particular influential, especially in the early days. The later days in the university were mainly spent on Festival and the various sub-projects it included. My long term partner in crime Alan Black deserves particular thanks for all the long conversations about virtually everything that helped form that system and gave me a better understanding on so many issues in computational linguistics and computer science in general. The other main author in Festival, Richard Caley, made an enormous and often unrecognised contribution to Festival. Tragically Richard died in 2005 before his time. I would of course like to thanks everyone else in CSTR and the general Edinburgh University speech and language community for making it such a great place to work. Following CSTR, I co-founded Rhetorical Systems in 2000 and for 4 years had a great time. I think the team we assembled there was world class and for quite a while we had (as far as our tests showed) the highest quality TTS system at that time. I thought the technical environment there was first class and helped finalise my attitudes towards a no-nonsense approach to TTS system. All the techical team there were outstanding but I would like in particular to thank the R& D team who most clearly impacted my thoughts and knowled on TTS. They were Kathrine Hammervold, Wojciech Skut, David McKelvie, Matthew Aylett, Justin Fackrell, Peter Rutten and David Talkin. In recognition of the great years at Rhetorical and to thank all those who had to have me as a boss, this book is dedicated to the entire Rhetorical technical team. Paul Taylor Cambridge November 19, 2007
- 25. 1 INTRODUCTION This is a book about getting computers to read out loud. It is therefore about three things: the process of reading, the process of speaking, and the issues involved in getting computers (as opposed to humans) to do this. This field of study is known both as speech synthesis, that is the “synthetic” (computer) generation of speech, and text-to-speech or TTS; the process of converting written text into speech. As such it compliments other language technologies such as speech recognition, which aims to convert speech into text, machine translation which converts writing or speech in one language into writing or speech in another. I am assuming that most readers have heard some synthetic speech in their life. We experi- ence this in a number of situations; some telephone information systems have automated speech response, speech synthesis is often used as an aid to the disabled, and Professor Stephen Hawking has probably contributed more than anyone else to the direct exposure of (one particular type of) synthetic speech. The idea of artificially generated speech has of course been around for a long time - hardly any science fiction film is complete without a talking computer of some sort. In fact science fiction has had an interesting effect on the field and our impressions of it. Sometimes (less technically aware) people believe that perfect speech synthesis exists because they “heard it on Star Trek”1. Often makers of science fiction films fake the synthesis by using an actor, although usually some processing is added to the voice to make it sound “computerised”. Some actually use real speech synthesis systems, but interestingly these are usually not state of the art systems, as these sound too natural, and may mislead the viewer2 One of the films that actually makes a genuine attempt to predict how synthetic voices will sound is the computer HAL in the film 2001: A Space Odyssey [265]. The fact that this computer spoke with a calm and near humanlike voice gave rise to the sense of genuine intelligence in the machine. While many parts of this film were wide of the mark (especially, the ability of HAL to understand, rather than just recognise human speech), the makers of the film just about got it right in predicting how good computer voices would be in the year in question. 1 Younger readers please substitute the in-vogue science-fiction series of the day 2 In much the same way, when someone types the wrong password on a computer, the screen starts flashing and saying “access denied”. Some even go so far as to have a siren sounding. Those of us who use computers know this never happens, but in a sense we go along with the exaggeration as it adds to the drama. 1
- 26. 2 Chapter 1. Introduction Speech synthesis has progressed remarkably in recent years, and it is no longer the case that state-of-the-art systems sound overtly mechanical and robotic. That said, it is normally fairly easy to tell that it is a computer talking rather than a human, and so substantial progress is still to be made. When assessing a computer’s ability to speak, one fluctuates between two judgments. On the one hand, it is tempting to paraphrase Dr Johnson’s famous remark [61] “Sir, a talking computer is like a dog’s walking on his hind legs. It is not done well; but you are surprised to find it done at all.” And indeed, even as an experienced text-to-speech researcher who has listened to more synthetic speech than could be healthy in one life, I find that sometimes I am genuinely surprised and thrilled in a naive way that here we have a talking computer: “like wow! it talks!”. On the other hand it is also possible to have the impression that computers are quite dreadful at the job of speaking; they make frequent mistakes, drone on, and just sound plain wrong in many cases. These impressions are all part of the mysteries and complexities of speech. 1.1 WHAT ARE TEXT-TO-SPEECH SYSTEMS FOR? Text-to-speech systems have an enormous range of applications. Their first real use was in reading systems for the blind, where a system would read some text from a book and convert it into speech. These early systems of course sounded very mechanical, but their adoption by blind people was hardly surprising as the other options of reading braille or having a real person do the reading were often not possible. Today, quite sophisticated systems exist that facilitate human computer interaction for the blind, in which the TTS can help the user navigate around a windows system. The mainstream adoption of TTS has been severely limited by its quality. Apart from users who have little choice (as in the case with blind people), people’s reaction to old style TTS is not particularly positive. While people may be somewhat impressed and quite happy to listen to a few sentences, in general the novelty of this soon wears off. In recent years, the considerable advances in quality have changed the situation such that TTS systems are more common in a number of applications. Probably the main use of TTS today is in call-centre automation, where a user calls to pay an electricity bill or book some travel and conducts the entire transaction through an automatic dialogue system Beyond this, TTS systems have been used for reading news stories, weather reports, travel directions and a wide variety of other applications. While this book concentrates on the practical, engineering aspects of text-to-speech, it is worth commenting that research in this field has contributed an enormous amount to our general understanding of language. Often this has been in the form of “negative” evidence, meaning that when a theory thought to be true was implemented in a TTS system it was shown to be false; in fact as we shall see, many linguistic theories have fallen when rigorously tested in speech systems. More positively, TTS systems have made a good testing grounds for many models and theories, and TTS systems are certainly interesting in their own terms, without reference to application or use.
- 27. Section 1.2. What should the goals of text-to-speech system development be? 3 1.2 WHAT SHOULD THE GOALS OF TEXT-TO-SPEECH SYSTEM DEVELOPMENT BE? One can legitimately ask, regardless of what application we want a talking computer for, is it really necessary that the quality needs to be high and that the voice needs to sound like a human? Wouldn’t a mechanical sounding voice suffice? Experience has shown that people are in fact very sensitive, not just to the words that are spoken, but to the way they are spoken. After only a short while, most people find highly mechanical voices irritating and discomforting to listen to. Furthermore tests have shown that user satisfaction increases dramatically the more “natural” sounding the voice is. Experience (and particularly commercial experience) shows that users clearly want natural sounding (that is human-like) systems. Hence our goals in building a computer system capable of speaking are to first build a system that clearly gets across the message, and secondly does this using a human-like voice. Within the research community, these goals are referred to as intelligibility and naturalness. A further goal is that the system should be able to take any written input; that is, if we build an English text-to-speech system, it should be capable of reading any English sentence given to it. With this in mind, it is worth making a few distinctions about computer speech in general. It is of course possible to simply record some speech, store it on a computer and play it back. We do this all the time; our answer machine replays a message we have recorded, the radio plays interviews that were previously recorded and so on. This is of course simply a process of playing back what was originally recorded. The idea behind text-to-speech is to “play back” messages that weren’t originally recorded. One step away from simple playback is to record a number of common words or phrases and recombine them, and this technique is frequently used in telephone dialogue services. Sometimes the result is acceptable, sometimes not, as often the artificially joined speech sounded stilted and jumpy. This allows a certain degree of flexibility, but falls short of open ended flexibility. Text-to-speech on the other hand, has the goal of being able to speak anything, regardless of whether the desired message was originally spoken or not. As we shall see in Chapter 13, there are a number of techniques for actually generating the speech. These generally fall into two camps, which we can call bottom-up and concatenative. In the bottom-up approach, we generate a speech signal “from scratch”, using our knowledge of how the speech production system works. We artificially create a basic signal and then modify it, much the same way that the larynx produces a basic signal which is then modified by the mouth in real human speech. In the concatenative approach, there is no bottom-up signal creation per se; rather we record some real speech, cut this up into small pieces, and then recombine these to form “new” speech. Sometimes one hears the comment that concatenative techniques aren’t “real” speech synthesis in that we aren’t generating the signal from scratch. This point may or may not be relevant, but it turns out that at present concatenative techniques far out perform other techniques, and for this reason concatenative techniques currently dominate.
- 28. 4 Chapter 1. Introduction 1.3 THE ENGINEERING APPROACH In this book, we take what is known as an engineering approach to the text-to-speech problem. The term “engineering” is often used to mean that systems are simply bolted together, with no underlying theory or methodology. Engineering is of course much more than this, and it should be clear that great feats of engineering, such as the Brooklyn bridge were not simply the result of some engineers waking up one morning and banging some bolts together. So by “engineering”, we mean that we are tackling this problem in the best traditions of other engineering; these include, working with the materials available and building a practical system that doesn’t for instance take days to produce a single sentence. Furthermore, we don’t use the term engineering to mean that this field is only relevant or accessible to those with (traditional) engineering backgrounds or education. As we explain below, TTS is a field relevant to people from many different backgrounds. One point of note is that we can contrast the engineering approach with the scientific ap- proach. Our task is to build the best possible text-to-speech system, and in doing so, we will use any model, mathematics, data, theory or tool that serves our purpose. Our main job is to build an artefact and we will use any means possible to do so. All artefact creation can be called engineer- ing, but good engineering involves more: often we wish to make good use of our resources (we don’t want to use a hammer to crack a nut); we also in general want to base our system on solid principles. This is for several reasons. First, using solid (say mathematical) principles assures us we are on well tested ground; we can trust these principles and don’t have to experimentally verify every step we do. Second, we are of course not building the last ever text-to-speech system; our system is one step in a continual development; by basing our system on solid principles we hope to help others to improve and build on our work. Finally, using solid principles has the advantage of helping us diagnose the system, for instance to help us find why some components do perhaps better than expected, and allow the principles of which these components are based to be used for other problems. Speech synthesis has also been approached from a more scientific aspect. Researchers who pursue this approach are not interested in building systems for their own sake, but rather as models which will shine light on human speech and language abilities. As such, the goals are different, and for example, it is important in this approach to use techniques which are at least plausible possibilities for how humans would handle this task. A good example of the difference is in the concatenative waveform techniques which we will use predominantly; recording large numbers of audio waveforms, chopping them up and gluing them back together can produce very high quality speech. It is of course absurd to think that this is how humans do it. We bring this point up because speech synthesis is often used (or was certainly used in the past) as a testing ground for many theories of speech and language. As a leading proponent of the scientific viewpoint states, so long as the two approaches are not confused, no harm should arise (Huckvale [225]).
- 29. Section 1.4. Overview of the book 5 1.4 OVERVIEW OF THE BOOK I must confess to generally hating sections entitled “how to read this book” and so on. I feel that if I bought it, I should be able to read it any way I damn well please! Nevertheless, I feel some guidelines may be useful. This book is what one might call an extended text book. A normal text book has the job of explaining a field or subject to outsiders and this book certainly has that goal. I qualify this by using the term “extended” for two reasons. Firstly the book contains some original work, and is not simply a summary, collection or retelling of existing ideas. Secondly, the book aims to take the reader right up to the current state of the art. In reality this can never be fully achieved, but the book is genuinely intended to be an “all you ever need to know”. More modestly, it can be thought of as “all that I know and can explain”. In other words: this is it: I certainly couldn’t write a second book which dealt with more advanced topics. Despite these original sections, the book is certainly not a monograph. This point is worth reinforcing: because of my personal involvement in many aspects of TTS research over the last 15 years, and specifically because of my involvement in the development of many well-known TTS systems, including Chatr [53], Festival [55], and rVoice, many friends and colleagues have asked me whether this is a book about those systems or the techniques behind those systems. Let me clearly state that this is not the case; Text-to-Speech Synthesis is not a system book that describes one particular system; rather I aim for a general account that describes current techniques without reference to any particular single system or theory. 1.4.1 Viewpoints within the book That said, this book aims to provide a single, coherent picture of text-to-speech, rather than simply a list of available techniques. And while not being a book centred on any one system, it is certainly heavily influenced by the general philosophy that I have been using (and evolving) over the past years, and I think it is proper at this stage to say something about what this philosophy is and how it may differ from other views. In the broadest sense, I adopt what is probably the current mainstream view in TTS, that this is an engineering problem, that should be approached with the aim of producing the best possible system, rather than with the aim of investigating any particular linguistic or other theory. Within the engineering view, I again have taken a more specialist view in posing the text-to- speech problem as one where we have a single integrated text analysis component followed by a single integrated speech synthesis component. I have called this the common form model (this and other models are explained in Chapter 3.) While the common form model differs significantly from the usual “pipelined” models, most work that has been carried out in one framework can be used in the other without too much difficulty. In addition to this, there are many parts which can be considered as original (at least to my knowledge) and in this sense, the book diverges from being a pure text book at these points.
- 30. 6 Chapter 1. Introduction Specifically, these parts are 1. the common form model itself 2. the formulation of text analysis as a decoding problem, 3. the idea that text analysis should be seen as a semiotic classification and verbalisation prob- lem, 4. the model of reading aloud, 5. the general unit selection framework. 6. the view that prosody is composed of the functionally separate systems of affective, aug- mentative and suprasegmental prosody. With regard to this last topic, I should point out that my views on prosody diverge consid- erably from the mainstream. My view is that mainstream linguistics, and as a consequence much of speech technology has simply got this area of language badly wrong. There is a vast, confus- ing and usually contradictory literature on prosody, and it has bothered me for years why several contradictory competing theories (of say intonation) exist, why no-one has been able to make use of prosody in speech recognition and understanding systems, and why all prosodic models that I have tested fall far short of the results their creators say we should expect. This has led me to propose a completely new model of prosody, which is explained in Chapters 3 and 6. 1.4.2 Readers’ backgrounds This book is intended for both an academic and commercial audience. Text-to-speech or speech synthesis does not fall neatly into any one traditional academic discipline, and so the level and amount of background knowledge will vary greatly depending on a particular reader’s background. Most TTS researchers I know come from an electrical engineering, computer science or linguistics background. I have aimed the book at being directly accessible to readers with these backgrounds, but the book should in general be accessible to those from other fields. I assume that all readers are computer literate and have some experience in programming. To this extent, concepts such as algorithm, variable, loop and so on are assumed. Some areas of TTS are mathematical, and here I have assumed that the entry level is that of an advanced high school or first year university course in maths. While some of the mathematical concepts are quite advanced, these are explained in full starting with the entry level knowledge. For those readers with little mathematical knowledge (or inclination!) don’t worry; many areas of TTS do not require much maths. Even for those areas which do, I believe a significant understanding can still be achieved by reading about the general principles, studying the graphs, and above all, trying the algorithms in practice. Digital filters can seem like a complicated and abstract subject to many; but I have seen few people fail to grasp its basics when give the opportunity to play around with filters in a GUI package. My commercial experience made it clear that it was difficult to find software developers with any knowledge of TTS. It was seen as too specialist a topic and even for those who were interested
- 31. Section 1.4. Overview of the book 7 in the field, there was no satisfactory introduction. I hope this book will help solve this problem, and I have aimed it at being accessible to software engineers (regardless of academic background) who wish to learn more about this area. While this book does not give a step by step recipe for building a TTS system, it does go significantly beyond the theory, and tries to impart a feel for the subject and tries to pass on some of the “folklore” that is necessary for successful development. I believe the book covers enough ground such that a good software engineer should not have too much difficulty with implementation. 1.4.3 Background and specialist sections The book contains a significant amount of background material. This is included for two reasons. Firstly, as just explained, I wanted to make the book seem complete to any reader outside the field of speech technology. I believe it is to the readers’ benefit to have introductory sections on phonology or signal processing in a single book, than have to resort to the alternative of pointing the reader to other works. There is a second reason however, which is that I believe that the traditional approach to explaining TTS is too disjointed. Of course TTS draws upon many other disciplines, but the dif- ferences between these to me are often overstated. Too often, it is believed that only “an engineer” (that is someone who has a degree in engineering) can understand the signal processing, only “a linguist” (again, degree in linguistics) can understand the phonetics and so on. I believe that this view is very unhelpful; it is ridiculous to believe that someone with the ability to master signal processing isn’t able to understand phonetics and vice versa. I have attempted to bridge these gaps by providing a significant amount of background material, but in doing so have tried to make this firstly genuinely accessible and secondly focused on the area of text to speech. I have therefore covered topics found in introductory texts in engineering and linguistics, but tried to do so in a novel way that makes the subject matter more accessible to readers with different backgrounds. It is difficult to judge potential readers’ exposure to the fundamentals of probability as this is now taught quite extensively. For this reason, I have assumed a knowledge of this in the body of the book, and have included an reference section on this topic in the appendix. The book is written in English and mostly uses English examples. I decided to write the book and focus on one language rather than make constant references to the specific techniques or variations that would be required for every language of interest to us. Many newcomers (and indeed many in the field who don’t subscribe to the data-driven view) believe that the differences between languages are quite substantial, and that what works for English, is unlikely to work for French, Finnish, or Chinese. While languages obviously do differ, in today’s modern synthesisers these differences can nearly all be modelled by training and using appropriate data; the same core engine suffices in all cases. Hence concentrating on English does not mean we are building a system that will only work on one language.
- 32. 2 COMMUNICATION AND LANGUAGE Before delving into the details of how to perform text-to-speech conversion, we will first examine some of the fundamentals of communication in general. This chapter looks at the various ways in which people communicate and how communication varies depending on the situation and the means which are used. From this we can develop a general model which will then help us specify the text-to-speech problem more exactly in the following chapter. 2.1 TYPES OF COMMUNICATION We experience the world though our senses and we can think of this as a process of gaining information. We share this ability with most other animals; if an animal hears running water it can infer that there is a stream nearby; if it sees a ripe fruit it can infer there is food available. This ability to extract information from the world via the senses is a great advantage in the survival of any species. Animals can however cause information to be created: many animals make noises, such as barks or roars, or gestures such as flapping or head nodding, which are intended to be interpreted by other animals. We call the process of deliberate creation of information with the intention that it be interpreted as communication. The prerequisites for communication are an ability to create information in one being, an ability to transmit this information, and an ability to perceive the created information by another being. All three of these pre-requisites strongly influence the nature of communication; for exam- ple, animals that live in darkness or are blind would be unlikely to use a visual system. But despite these restrictions, it is clear that they are still many possible ways to make use of the possibilities of creation, medium and perception to communicate. We will now examine the three fundamental communication techniques that form the basis for human communication. 2.1.1 Affective communication The most basic and common type of communication is affective communication, where we ex- press a primary emotional state with external means. A good example of this is the expression of pain, where we might let out a yell or cry upon hurting ourselves. A defining characteristic of 8
- 33. Section 2.1. Types of communication 9 Figure 2.1 (British) road sign, indicating a slippery road and high chance of skidding. this type of communication is that the intensity of the external form is clearly a function of the the intensity of feeling; the more intense the pain the louder the yell. Other primary mental states such as happiness, anger and sadness can be expressed in this way. This type of communication is one that is common to most higher animals. While the ability to express these affective states is common among animals, the precise means by which these are expressed is not universal or always obvious. A high pitched yell, squeal or cry often means pain, but it by no means obvious that a dog’s wagging tail or a cat’s purring are expressions of happiness. 2.1.2 Iconic communication Though fundamental and powerful, affective communication is severely limited in the range of things it can be used to express. Happiness can readily be conveyed, but other simple mental states such as hunger or tiredness are significantly more difficult to convey. To express more complex messages, we can make use of a second communication mechanism known as iconic communication. An iconic system is one where the created form of the communication somehow resembles what the intended meaning. We use the term “form” here in a technical sense that al- lows us to discuss the common properties of communication systems: in acoustic communication, “form” can be thought of as a type of sound; in visual communication form might be types of hand signals, facial expressions and so on. For example, it is common to communicate tiredness iconically by the “sleeping gesture”, where someone closes her eyes, puts her hands together and places her head sideways on her hands. The person isn’t really asleep - she using a gesture that (crudely) mimics sleep to indicate tiredness. Another good example of iconic communication is the road sign shown in Figure 2.1.1. In this case, the form is the diagram, and the meaning is slippery road, and the fact that the form visually resembles what can happen when a road is slippy means that this communication is iconic. Note that just with the sleep example, the form isn’t a particularly accurate picture of a car, road, skid or so on; the idea is to communicate the essence
- 34. 10 Chapter 2. Communication and Language of the meaning and little else. Anecdotally, we sometimes think of pre-human communication also working like this, and in such a system, the presence of a saber tooth tiger might be communicated by imitating its growl, or by acting like a tiger and so on. Such systems have a certain advantage of transparency, in that if the intended recipient does not know what you mean, a certain amount of mimicry or acting may get the point across. In essence, this is just like the road sign example, in that the form and meaning have a certain direct connection. When travelling in a country where we don’t speak the language, we often resort to miming some action with the hope that the meaning will be conveyed. Iconic systems though have several drawbacks. One is that while it may be easy to imitate a tiger, it is less easy to imitate more abstract notions such as “nervous” or “idea”. More importantly though, iconic systems can suffer from a lack of precision: when a first caveman imitates the tiger, a second caveman may not get this reference exactly - he might be sitting there thinking, “well, it could be a tiger, or perhaps a lion. Or maybe a large dog”. And by the time the first caveman has mimed his way through this and the action of “creeping up behind”, both have probably departed to the great cave in the sky. While useful and important, iconic communication clearly has its limits. 2.1.3 Symbolic communication In contrast to iconic and affective communication, we also have symbolic communication in which we give up the idea that the form must indicate the meaning. Rather we use a series of correspon- dences between form and meaning, in which the relationship is not direct. In symbolic systems, a tiger might be indicated by waving the left hand, a lion by waving the right. There is no reason why these forms should indicate what they do, it is merely a matter of convention. The advantage is that it is easier to devise a system where the forms are clear and distinct from one another - less confusion will arise. The disadvantage is that the fact that left-arm-wave means tiger, and right- arm-wave means lion has to be learned, if you don’t know, seeing the movement in isolation won’t give clue to the meaning. Despite the disadvantage in needing to learn, using conventions rather than icons can be hugely advantageous in that the conventions can be relatively brief; one noise or gesture may represent the tiger and this need not be acted our carefully each time. This brevity and clarity of form leads to a swiftness and precision seldom possible with iconic communication. Once a convention-based communication system is used, it soon becomes clear that it is the notion of contrast in form that is the key to success. To put it another way, once the form no longer needs to resemble the meaning, the communication system gains benefit from making the forms as distinct from one another as possible. To show this point, consider the following experiment. We grouped eight subjects into pairs, and in isolation, asked each pair to design a commu- nication system based on colour cards. The premise was that the pair were in a noisy pub, one individual was at the bar while the other was sitting down at a table. We said there were four basic concepts to communicate: “I would like a drink of water”, “I would like some food”, “I would like a beer” and “I would like to listen to some music”. Each pair was given a set of 100 different
- 35. Section 2.1. Types of communication 11 ❄ ● ❑ ❄ ❑ ● ❑ ❄ ❄ ● ● ❑ ❄ ❄ ❑ ❑ ● ❄ ❄ ● ❑ ❑ ❄ Figure 2.2 With three symbols in isolation only three meanings are possible. When we combine the symbols in sequences the number of meanings exponentially increases, while the ability to uniquely perceive the symbols is unaffected. coloured cards arranged across the spectrum, and asked to pick one card for each concept and memorise which colour was associated with which concept. In a simple test, it was clear that the pairs could remember the colours and their meanings, and could effectively communicate these simple concepts with this system. The real purpose of the experiment was different though. While three of the groups chose a blue colour for the concept “I want a drink of water”, for the other cat- egories there was no commonality in the choice. So “I want a beer” had two pairs chose green, one yellow and one black. Interestingly most gave some justification as to why each colour was chosen (green means “yes, lets have another”) and so on; but it is clear that this justification was fairly superficial in that each group choose quite different colours. This demonstrates one of the main principles of symbolic communication system, namely that the relationship between the form of the symbol (in this case the colour) and its meaning is in general arbitrary. These form/meaning pairs must be learned, but as we shall see, in many cases the “cost” of learning the pairing to an individual is less than the “cost” of potential confusion and imprecision that can arise from having the form resemble the meaning. This simple study demonstrates another vital point. In each case, the pairs chose quite different colours; so group A chose a blue, a red, a yellow and a green, while group B chose a red, a yellow, a light blue and a black. In no case did a group choose say a green, a light blue, a medium blue and a dark blue for instance. In other words, the groups chose a set of colours that had clear contrast with each other. In a way this makes obvious sense; picking colours that are significantly different from one another lessens the possibility that they will be confused, but it does demonstrate a crucial point in symbolic systems, namely that the systems are based on the notion of contrast between forms rather than the absolute “value” of each form. 2.1.4 Combinations of symbols This simple communication system works well because we only have four concepts we wish to convey. What happens when we need to communicate more concepts? We can see that the simple
- 36. 12 Chapter 2. Communication and Language colour contrast will eventually reach its limits. If say 20 concepts are required, it soon becomes a difficult task to associate each with just one colour. Setting aside the issue of whether the individuals can remember the colours (let us assume they have a crib sheet to remind them), a major problem arises because as more concepts get added, a new colour is associated with each one, and as the range of colours is fixed, each new added concept increases the likelihood that it will get confused with another concept which has a similar colour. While it may be possible for an individual to distinguish 20 colours, asking them to do so for 1000 colours seems unlikely to meet with success. We therefore need a separate solution. The obvious thing to do is use combinations of colour cards. For instance, we could associate the colour black with the concept not, pink with large, and then use these and one of the original colours to indicate a much larger range of concepts (so holding a black card, pink card and blue card would say “I don’t want a large drink of water”). Note that this is different from holding a single card that is a mixed combination of black pink and blue (which might be a murky brown). The reason why the combination system works is that each colour card on its own is clear, distinct and unconfusable, but by combining them, we can generate a large number of concepts from a small number of form/meaning pairs. The power of combination is considerable; in the limit, the number of complex forms exponentially multiples, so that if we have N simple forms and M instances of them, we have a total of NM complex forms; so 10 forms combined in 10 different ways would produce over 10 billion different complex forms. We call communication systems that use a fixed number of simple forms which can be com- bined into a much larger number of forms a symbolic language. Here we use the term “language” in both a very precise and general sense. Precise, in that it we use it only to mean this principle of combination, but general in the sense that all human languages make use of this principle in this sense, as are other systems, such as written numbers (we can create any number from a combina- tion of the 10 digits), mathematics and even email addresses. Human language often goes under the description of natural language and we shall use that term here to distinguish it from the principle of languages in general. One of the main distinctive properties of natural language is the sheer scale of things that it can express. While we can talk of other animals having sophisticated communication systems (bee dancing etc), none has anything like the scale of natural language, where a person might know say 10,000 words, and be able to combine these into trillions and trillions of sentences. 2.1.5 Meaning, form and signal In our discussion so far, we have been somewhat vague and a little informal about what we really mean by “meaning” and “form”. The study of meaning (often called semantics), is notoriously difficult, and has been the subject of debate since the earliest philosophy. It is an inherently tricky subject, because firstly it requires an understanding of the world (what the world contains, how the things in the world relate to one another and so on), and secondly, from a cognitive perspective, it is exceedingly difficult to peer inside someone’s brain and find out exactly how they represent the concept of “warm” or “drink” and so on. As we shall see, developing a precise model of
- 37. Section 2.2. Human Communication 13 semantics does not seem to be necessary for text-to-speech, and so we shall largely bypass this topic in a formal sense. Our notion of “form” also needs some elaboration. A proper understanding of this part of communication is vital for our purposes, but thankfully we can rely on an established framework to help us. The main issue when discussing form is to distinguish the general nature of the form from any specific instantiation it may have. Going back to our simple colour example; it is important to realise that when one individual holds up a red card, it isn’t the red card itself that is the important point, rather that red was used rather than the other choices of green, blue and yellow. So for example, a different red card could be used successfully, and probably any red card would do, as it is not the particular red card itself which is the form, but rather that the card has the property “red” and it is this property and its contrasting values that is the basis of the communication system. In some accounts, the term message is used instead of form. We use the term signal to refer to the physical manifestation of the form. Here signal is again used in a specific technical sense: it can be thought of as the thing which is stored or the thing which is transmitted during communication. The idea of “red” (and the idea that it contrasts with other colours) is the form; the physical light waves that travel are the signal. The above example also demonstrates that the same form (red) can have a number of quite different signals, each dependent on the actual mode of transmission. It is important to see that the relationship between meaning and form, and form and signal are quite separate. For instance, in our example we have associated the colours red, yellow, blue and green each with a concept. Imagine now the sender holds up an orange card. Most likely the receiver will see this signal and correctly interpret the physical light-waves into a form “orange”. After this however she is stuck because she can’t determine what concept is matched with this or- ange form. The signal/form relationship has worked well, but it has not been possible to determine the correct concept. This situation is similar to that when we read a new word - while are quite capable of using our eyes to read the physical pattern from the page and create an idea of the letters in our head, if we don’t know the form/meaning mapping of the word, we can not understand it. Finally, we use the term channel or medium to refer to the means by which we the message is converted and transmitted as a signal. In this book, we deal mainly with the spoken channel and the written channel. 2.2 HUMAN COMMUNICATION The above examination of communication in the abstract has of course been leading us to an explanation of how human communication operates. Taking spoken communication first, let us make a primary distinction between the two main aspects of this, which we shall call the verbal component and the prosodic component. We use the term “verbal” to mean the part of human communication to with words. In some books, the term “verbal” and “language” are synonymous, but because “language” has so many other, potentially confusable meanings, we will stick with the
- 38. 14 Chapter 2. Communication and Language terms “verbal” or “verbal component” throughout. The verbal component is a symbolic system in that we have a discrete set of words which can be arranged in sequences to form a very large number of sentences. In contrast, the prosodic component (often just called prosody) is different in that it is not primarily symbolic. We use prosody for many purposes: to express emotion or surprise, to em- phasise a word, to indicate where a sentence ends and so on. It is fundamentally different in that it is not purely discrete, does not generally have units which combine in sequences and so has a much more limited range of meanings that can be expressed. That said, prosody should not be taken as a poor relative of the verbal component; if we say “its wonderful” with words, but use a tone of voice that indicates disgust; the listener will guess we are being sarcastic. If the prosodic and verbal component disagree, the prosodic information wins. To a certain extent, these systems can be studied separately; in fact nearly all work in the field of linguistics is solely interested in the verbal component. The two do interact though; first in the obvious sense that they share the same signal (speech), but also in more complex ways in that prosody can be used to emphasise one particular word in a sentence. 2.2.1 Verbal communication The symbolic model of the verbal component states that the principle unit of form is the phoneme. Phonemes operate in a contrastive system of sounds, in which a fixed number of phonemes can combine to a much larger number of words (a vocabulary of more than 10,000 words is common, but many millions of words are possible). Next, these words are taken as the basic units, and they further combine into sentences. Verbal language is therefore seen as a combination of two systems; one that makes words from phonemes, and another that makes sentences from words. As the word/sentence system has quite a large set of basic units to start with, the number of combined units is vast. The two systems have different properties with regard to meaning. In the first system, the basic building blocks, the phonemes, have no direct (arbitrary, iconic or other) meaning, they are simply building blocks. These are combined to form words, and these do have meaning. The relation between a word’s meaning and the phonemes which comprise it is completely arbitrary and so the meaning and pronunciation of each word has to be learned. The second system operates differently, in that the basic building blocks (the words) do have inherent meaning, and the meaning of the sentence is related to the meaning of the components. Exactly how the sentence meaning relates to the word meaning is still an issue of considerable debate; one the one hand there is considerable evidence that it is a function, in the sense that if we have JOHN KICKED MARY1 we can substitute different people for JOHN and MARY (e.g. ANN KICKED GEORGE), or different verbs (e.g. JOHN KISSED MARY) and predict the sentence meaning from the parts. On the other hand, there is also considerable evidence that sentence meaning is not a simple function, in that JOHN KICKED THE BUCKET means John has died, and has nothing to do 1 From here we will use a small caps typeface when we refer to form/message such as JOHN and a courier typeface for the written signal, e.g. john.
- 39. Section 2.2. Human Communication 15 with kicking or buckets. From this basic model, we can now explore some of the fundamental features of language. In doing so, we follow Hockett [214], who stated the principle defining properties of verbal language. Among these are: Arbitrary Phonemes do not have inherent meaning but words do; a word is in effect a form/meaning pair in which the correspondence between the meaning and the form (ie. the sequence of phonemes) is arbitrary. Because of this, there is no way to guess the meaning of a word if it is not learned, so as a reader, you won’t know the meaning of the word PANDICULATION unless you are informed of its meaning. From this, it follows that words of similar form can have complete different meanings, so BIT and PIT mean quite different things despite only having a slight sound difference, whereas DOG and WOLF are similar in meaning, but share no sounds. Duality The duality principle states that verbal language is not one system but two; the first system uses a small inventory of forms, phonemes, which sequentially combine to form a much larger inventory of words. The phonemes don’t carry meaning but the words do, and this association is arbitrary. The second system combines the large number of words into an effectively limitless number of sentences. Productive The productive property means that there is an effectively unlimited number of things that one can say in a verbal language. It is not the case that we have a fixed number of messages and in communication simply choose one from this selection; rather many of the sentences we produce are unique, that is, the particular combination of words we use have never been used before. Of course many of the sentences we say are conventionalised (“How do you do?” etc), but most are unique. It might seem questionable if this is really the case as most people don’t feel they are being particularly “original” when the speak. However, a simple examination of the combinatorics of combining even a modest vocabulary of 10,000 words shows that a vast number of possibilities exist (certainly larger that the number of atoms in the universe). Discrete Both the basic types of linguistic unit, the phoneme and the word, are discrete. Phonemes are dis- tinct from one another and form a finite and fixed set. The set of words is less fixed (we can create new words), but nevertheless all words are distinct from one another and are countable. What is remarkable about the discreteness property of language is that the primary signal of language, speech, is not discrete and hence there is a mismatch between form and signal. Speech signals are continuous functions of pressure over time (see Chapter 10) and so speaking is a process of
- 40. 16 Chapter 2. Communication and Language converting a discrete phoneme representation into a continuous acoustic one. This process is re- versed during listening, where we recover a discrete representation from the continuous signal. As we shall see, this is a remarkably complex process, and is one of the main reasons why automatic speech synthesis and speech recognition are difficult. 2.2.2 Linguistic levels Another common way to look at verbal language is as a series of so-called linguistic levels. So far we have talked about the duality of form such that we have a phonemes which combine to form words which combine to form sentences. In our model, sentences, words and phonemes are considered the primary units of form, and suffice to describe the basic properties of verbal language. It is however extremely useful to make use of a number of secondary units, which we shall now introduce. Morphemes are units used to describe word formation or the internal structure of words. This describes for instance the connection between the three English words BIG, BIGGER, BIGGEST, and the study of this is called morphology. Next we have syntax, often defined as the study of the patterns of words within a sentence. Syntax describes the process of how nouns, verbs, prepositions and so on interact, and why some sequences of words occur but not others. The fundamental units of syntax are the syntactic phrase and sentence. With these levels, we can for a given sentence describe a syntactic hierarchy where a sentence comprises a number of phrases, each of these comprises a number of words, which then comprise morphemes, which then comprise phonemes. We can also consider the phonological hierarchy, which is a different, parallel hierarchy which focuses on the sound patterns in a sentence. In this, we have the additional units of syllables, which are a structural sound unit used to group phonemes together; and phonological phrases, another structural sound unit that groups words together within the sentence. Representations using these units can also be expressed in a hierarchy of sentences, prosodic phrases, words, syllables and phonemes. The difference between the two hierarchies is that the phonological one concentrates on the sound patterns alone, and is not concerned with any aspect of meaning, while the syntactic one ignores the sound patters, and concentrates on the grammatical and semantic relationships between the units. Next we consider levels between the form and the signal, which in spoken communication is an acoustic waveform. Here we have speech acoustics which involves working with the acoustic speech sound signal. Next we have phonetics, which is often described as the study of the physical speech sounds in a language. Phonetics is often studied independently of other linguistic areas, as it is seen as part of the system which links form and signal rather than as part of the meaning/form verbal linguistic system. Phonology is the study of sound patterns as they relate to linguistic use; in contrast to phonetics, this is considered part of the core linguistic system as phonology describes the form in the meaning/form system. That said, phonology and phonetics are closely related and phonology is sometimes described as the organisational system of phonetic units.
- 41. Section 2.2. Human Communication 17 Semantics is sometimes given a broad definition as the study of “meaning” in the general sense, but is sometimes given a much narrower definition, and is concerned only with certain as- pects of meaning, such as analytic truth, reference, and logic. Finally pragmatics concerns a num- ber of related areas including the aspects of meaning not covered by semantics, the interaction of language, a speaker’s knowledge and the environment and the use of language in conversations and discourses. Sometimes pragmatics is described as being “parallel” to semantics/syntax/phonology etc as it is the study of how language is used. It is important to realise that these terms are used both to refer to the linguistic level and the area of study. In traditional linguistics these terms are often taken as absolute levels that exist in language, such that one sometimes sees “turf wars” between phoneticians and phonologists about who best should describe a particular phenomenon. From our engineering perspective, we will find it more convenient to adopt these terms, but not worry too much about the division between each; rather it is best to view these levels as salient points in a continuum stretching from pragmatics to acoustics. 2.2.3 Affective Prosody We now turn to a different component of human communication. Broadly speaking, we can state that prosody has two main communicative purposes, termed affective and augmentative. The affective use of prosody can be used to convey a variety of meanings: perhaps the most straight- forward case is when is is used to indicate primary emotions such as anger or pain. Prosody is particularly good at this; when we hurt ourselves and want to tell someone we usually let out a yell rather than say “look, I’ve cut off my fingers” in a calm measured tone. The use of prosody to convey primary emotion is also largely universal; when listening to someone who speaks a dif- ferent language we find they use more or less the same form to indicate the same meanings. In fact this extends to animals too; we are often in little doubt as to what an animal snarl or growl is trying to convey. Beyond primary emotions we can consider a whole range of meanings that are expressible via prosody. These include secondary emotions such as angst or distrust, speech acts such as question, statement and imperative; modes of expression such as sarcasm or hesitancy and so on. Many of these phenomena are language specific (in contrast to primary emotion), and have a conventional or arbitrary character such that a speaker of that language has to learn the appropriate prosodic pattern in the same way that they learn the form/meaning correspondences for words in verbal language. In fact we can see these effects as lying on a scale with universal systems lying at one end and conventional arbitrary systems lying at the other. In many languages, a rising pitch pattern at the end of a sentence indicates a question; but this is not universally so. Ladd [269] shows that the pattern is reversed in Hungarian such that certain questions are indicate by a falling pitch. One of the central debates in the field of prosodic linguistics is just how far this scale extends. Crudely speaking, the argument concerns whether prosody is mainly a universal, non-arbitrary
- 42. 18 Chapter 2. Communication and Language system, with a few strange effects such as Hungarian question intonation; or whether it is in fact much more like verbal language in that it is mainly arbitrary, and has rules of syntax and so on. 2.2.4 Augmentative Prosody The second aspect of prosody is called augmentative prosody. The basic purpose of this is to augment the verbal component. Taking a simple example first, we see that many sentences are highly confusable with respect to syntactic structure. In sentences such as (1) BILL DOESN’T DRINK BECAUSE HE’S UNHAPPY we see that there are two completely different interpretations which depend on which syn- tactic structure is used. The speaker can choose to insert a pause, and if she does so, she will bias the listener’s choice as to which interpretation is intended, and hence facilitate understanding. In addition to helping with phrasing structure, this component of prosody is also used to draw attention to words by emphasising them. This use of prosody is called augmentative because it augments the verbal component and therefore helps the listener to find the intended form and meaning from the signal. In contrast to affective prosody, the use of augmentative prosody here does not add any new meaning or extra aspect to the existing meaning, it is simply there to help find the verbal content. One of the most interesting aspects of augmentative prosody is that speakers show considerable variation in how and when they use it. While one speaker may pause at a certain point, another speaker will speak straight on, and we even find that the same speaker will quite happily use a pause on one occasion and not on another. We can explain this variation by considering the range of conversational situations where augmentative prosody might be used. In general, when a speaker generates a message that is potentially ambiguous, it is often entirely obvious from the form of words used, the background knowledge and the situation what interpretation is intended, and in such cases no or little augmentative prosody is used. In other cases however, we may find that the speaker suspects the listener may have difficulty in interpre- tation, we find that here the speaker does indeed make use of prosody. In other words, prosody is usually only used when it is needed. Hence if we consider a sentence in isolation, it is nearly impossible to predict what prosody a speaker will use. A consequence of this is that it is clearly possible to use more or use less prosody for any given sentence. We can think of this as a scale of prosodic intensity, and this applies to both affective and augmentative prosody. In many cases, speakers utter sentences at the lower end of this scale, in which case we say they are speaking with null prosody or neutral prosody to indicate that there is no true prosody being used. 2.3 COMMUNICATION PROCESSES In this section, we will build a general model of communication that will serve as a framework in which to analyse the various communicative processes involved in spoken and written language.
- 43. Section 2.3. Communication processes 19 generation encoding understanding meaning message signal message meaning signal speaker listener transmission decoding Figure 2.3 Processes involved in communication between two speakers. For illustration, let us consider a dialogue between two participants, A and B. In a normal con- versation, A and B take it in turns to speak, and these are known as dialogue turns. Suppose, A decides to say something to B. To do this, A firstly generates a linguistic form, which we shall call the message. He then encodes this as a speech signal, which is a form which can be transmitted to B (lets assume they are in the same room). On receiving the signal, B first decodes the signal and creates her own message as linguistic form. Next she attempts to understand this message and so find the meaning. In line with our model of language explained before, we have three types of representation: the meaning, the form and the signal. In communicating, A and B use 4 separate systems to move from one to the other; these are: generation : meaning to form encoding : form to signal decoding : signal to form understanding : form to meaning This model works equally well for written language; A converts his thought into a message, then encodes this as writing and sends this to B (in this case the “sending” can be quite general, for example an author writing a novel and a reader subsequently picking it up). Once received, B attempts to decode the writing into a message, and then finally understand this message. A diagram of this is shown in Figure 2.3. 2.3.1 Communication factors We start our discussion by simply asking, why do the participants say the things they do? While “free-will” plays a part in this, it it is a mistake to think that “free-will” is the sole governing factor
- 44. 20 Chapter 2. Communication and Language in how conversations happen. In fact there are a number of guiding principles involved, which we shall now explore. The first concerns the understood conventions of communication. In human conversation, participants don’t just say everything on their mind; to do so would swamp the conversation. Rather, they only say what they need to in order to get their point across. We can state this as a case of effectiveness and efficiency; we wish to effectively communicate our message so that it is understood, but also to do so efficiently so that we don’t waste energy. In fact it is clear that conversation works by adopting a set of understood conventions that govern or guide the conversation. The most famous exposition of these principles was by Grice, who defined four main principles, of conversation which are now commonly known as Grice’s maxims [184]. These principles state for instance that it is expected that one participant should not be overly long or overly short in what is said; one should say just the right amount to make the thought clear. Grice’s maxims are not meant as prescriptive guides as to how conversation should progress, or even statements about how they do progress, but rather a set of tacit conventions, which we normally adhere to, and which we notice when they are broken. The second factor concerns the common ground that the participants share. This is a general term which covers the background knowledge, situation and mutual understand that the conver- sational participants share. When in the same room, they can point to objects or refer to objects, actions and so on, in a different way to how such things can be referred to if they are more distant. If one says “have a seat” and there is a chair in the room, this means something straightforward. The situation can significantly alter the intended meaning of an utterance. Take for instance the sentence “my car has broken down”. Now this can be taken as a simple statement of fact, but if someone says this upon entering a mechanic’s workshop, the implication is that the person’s car has broken down and he’d like the mechanic to fix it. If by contrast, this is the first thing someone says when arriving for dinner, the intention might be to apologise for being late because an un- forseen travel delay occurred. To take another example, if two experts in the same research field are talking, they can talk assuming that the other has a detailed knowledge of the subject matter. This is different to when an expert talks to a lay person. Finally, the knowledge of the channel helps form the message. Participants don’t generate messages and then decided whether to speak them or write them down; they know what medium is going to be used before they start. This allows the participants to best exploit the strengths of each form of communication; in spoken language extensive use may be made of prosody; in written language the author may take the liberty of explaining something in a complicated fashion knowing that the reader can go back and read this several times. 2.3.2 Generation In our model, the process of converting a meaning into a message is called message generation (elsewhere the term language generation is often used, but this is too potentially confusing for our purposes). To demonstrate how generation works, let us use the above mentioned communi-
- 45. Section 2.3. Communication processes 21 cation factors to consider situation where a participant in a conversation wants to know the score in a particular football game. We should realise that the participant has considerable choice in converting this thought into a message. For example, the participant can make a statement “I want to know the score” or a question “what is the score”; the participant has choices over mood, and may choose to say “can you tell me the score?”, may choose between active and passive “who has scored?” vs “have any goals been scored?”. The participant has choices over vocabulary, may well qualify the message, may decide on a casual, formal, polite, demanding tone and so on. In addition the participant can choose to emphasise words, pick different intonational tunes, perhaps make a pointing gesture and so on. The important point is that the participant have a wide range of options in turning the thought into the message. The result is the message, and a perfectly transparent communication system would send this to the receiver will all this detail and with no chance of error. It is sometimes assumed that messages themselves convey or represent the communicated meaning. While in certain cases this may be possible, we have just demonstrated that in most conversations, to extract (understand) the correct meaning one must know the situation and shared knowledge. 2.3.3 Encoding Encoding is the process of creating a signal from a message. When dealing with speech, we talk of speech encoding and when dealing with writing we talk of writing encoding. Speech encoding by computer is more commonly known as speech synthesis, which is of course the topic of this book. The most significant aspect of speech encoding is that the nature of the two representations, the message and the speech signal, are dramatically different. A word (e.g. HELLO) may be composed of four phonemes, /h eh l ow/, but the speech signal is a continuously varying acoustic waveform, with no discrete components or even four easily distinguishable parts. If one considers the message, we can store each phoneme in about 4-5 bits, so we would need say 2 bytes to store /h eh l ow/. By contrast, a reasonably high fidelity speech waveform2 would need about 20,000 bytes, assuming that the word took 0.6 seconds to speak. Where has all this extra content come from, what is it for and how should we view it? Of course, the speech signal is significantly more redundant than the message: by various compression techniques it is possible to reduce its size by a factor of 10 without noticeably affect- ing it. Even so, this still leaves about three orders of magnitude difference. The explanation for this is that the speech waveform contains a substantial amount of information that is not directly related to the message. First we have the fact that not all speakers sound the same. This may be a fairly obvious thing to say, but consider that all English speakers can say exactly the same message, but they all do so in subtly different ways. Some of these effects are not controllable by 2 Say 16 bits dynamic range and 16,000Hz sampling rate. (See Chapter 10 for an explanation)
- 46. 22 Chapter 2. Communication and Language the speaker and and determined by purely physiological means. Some are due to prosody, where for instance, the speaker can have different moods, can affect the signal by smiling when talking, can indicate different emotional states and so on. This fact is hugely significant for speech synthe- sis, as it demonstrates that encoding the message is really only part of the problem; the encoding system must do several other tasks too if the speech signal is to sound genuinely human-like to the receiver. 2.3.4 Decoding Decoding is the opposite of encoding; its job is to take a signal and create a message. Speech decoding is the task of creating a message from the signal; the technology that does this is most commonly known as speech recognition. The challenges for speech decoding (speech recogni- tion) systems are just the converse of speech encoding; from a continuous signal, which contains enormous amounts of information, the system must extract the small amount of content that is the message. Speech recognition is difficult because of the problem of ambiguity. For each part of a message, a word or a phoneme, if one studies the speech patterns produced from that one sees that there is an enormous amount of variability. This is partly caused by the factors we described above, namely speaker physiology, emotional state and so on, and partly from the encoding pro- cess itself, where a phonemes characteristics are heavily influenced by the other phonemes around it. This variability wouldn’t be a problem per se, but it turns out that the patterns for each phoneme overlap with one another, such that given a section of speech, it is very hard to determine which phoneme it belonged to. These aspects make speech recognition a difficult problem. Text decoding, also called text analysis, is a similar task where we have to decode the written signal to create a message. If the writing is on paper, we face many of the same problems as with speech recognition; namely ambiguity that arises from variability, and this is particularly true if we are dealing with hand writing. Computerised writing is significantly easier as we are dealing with a much cleaner representation in which the representation is discrete and each character is uniquely identified. For many types of message, the text decoding problem is fairly trivial. For instance, for simple text such as the man walked down the street, it should by possible for a simple system to take this and generate the correct message all the time. Significant problems do arise from a number of sources. First, ambiguity exists even in computerised written signals. One type is homograph ambiguity, where two different words have the same sequence of characters (e.g. polish can mean “from Poland” or “to clean”, console can mean “to comfort” or “a part of a machine”). One final point should be made when comparing text and speech decoding. In speech recog- nition, the problem is understood as being difficult, and so error rates of say 10% are often seen as good system performance on difficult tasks. As we have seen text decoding is fundamentally easier, but along with this so are expectations. So while for some tasks the vast majority of words may be decoded successfully, the expectation is also high and so error rates of 0.1% may be seen
- 47. Section 2.4. Discussion 23 as too hight. 2.3.5 Understanding Understanding is the process of creating a thought from the message. Understanding can be thought of as the reverse of generation and also bears similarity to decoding in the sense that we are trying to recover a thought from a representation (the message), which also includes information generated from other sources such as the background, situation and so on. In understanding, we also have an ambiguity problem, but this is quite different from the ambiguity problem in decoding. For example, it is perfectly possible when hearing a speech signal to decode it into the correct message, but then fail to understand it because of ambiguity. For instance, if someone says “did you see the game last night”, it is quite possible to correctly identify all the words, but fail to determine which game is being referred to. 2.4 DISCUSSION This chapter has introduced a model of communication and language that will serve as the basis for the next chapter, which looks at text-to-speech conversion in detail, and the subsequent chapters, which explore the various problems that occur and the techniques that have been developed to solve them. We should stress that what we have presented here is very much a model; that is a useful engineering framework in which to formulate, test and design text-to-speech systems. In building this model, we have drawn from a number of fields, including semiotics, informa- tion theory, psychology, linguistics and language evolution. These fields have overlapping spheres of subject matter, but unfortunately are often studied in isolation such that none really gives as complete and detailed a model of communication as we might wish for. Each of these separate disciplines is a vast field in itself, so we have only touched on each area in the briefest possible way. Considering semiotics first, we should look at the works of Ferdinand de Saussure, who is widely acknowledged to have invented the field of semiotics and is considered by many as the first modern linguist. Saussure was the first to emphasise the significance of the arbitrary nature of form and meaning in communication and proposed that the study of the relationship between the two should be the central topic in linguistics [391], [392]. The philosopher Charles Sanders Pierce also pioneered early work in semiotics and semantics, and gave us the framework of iconic and symbolic semiotics which is used here [350]. The second main area we draw from is communication theory or information theory, and from this we have the principles of encoding, decoding, channels, messages and signals. The pioneering work in this field was produced by Claude Shannon [400], who was originally interested in the problem of accurate transmission of information along telephone wires (he worked for the telephone company AT&T). It soon became clear that the implications of this work were much broader (Weaver [401]) and it has been widely
- 48. 24 Chapter 2. Communication and Language adopted as an approach to the analysis of signals arising from linguistic sources, such that this information theoretic approach is now dominant in speech recognition and many other fields. Saussure did not in general deal much with signals, and Shannon and Weaver did not deal much with meaning, but their work clearly intersects regarding the issue of form and message (the term form is used in linguistics and semiotics and message in information theory, but they both refer to the same thing.) The field of linguistics is often described as the science of language, but in the main this field focuses very much on just the meaning/form verbal part of language, as opposed to the entire process which includes participants, conversations, signals and so on. Linguistics generally approaches language without reference to use; the idea is that a language is an entity and it can be studied independently of any purpose to which it is put3. This approach is part tradition and part pragmatic; as linguistics can trace its lineage back to the philosophical musings of ancient Greek philosophers. In addition there is the practical justification that by isolating effects it makes them easier to study. In modern times, the field of linguistics has been dominated by Noam Chomsky[89], [90], [88], who in addition to many other innovations has championed the idea that the primary function of language is not communication at all, but rather forms the basis of the human mind; it evolved for this reason and its communicative purpose was a useful by product. Despite this mismatch between linguistics and our more practical needs, this field does provide the basics of a significant chunk of the overall picture of human communication, including the models of phonetics, phonology, words, syntax and so on that are central to our problem. Many linguistics text books tend to be heavily skewed towards one of the many incompatible theories within the discipline, so care needs to be taken when reading any book on the subject. A good introductory book which explains the “hard issues”, which is accessible to those with a scientific background and which doesn’t follow any particular theory is Lyons [290]. The study of human communication in terms of participants, conversations, background knowledge and so on is often performed outside the field of mainstream linguistics by psycholo- gists and philosophers. Good books on human conversation and language use include Clark [95] and Clark and Clark [96], Searle [395], Grice [184] and Power [361]. 2.5 SUMMARY Communication • A semiotic system has three levels of representation; meaning, form, and signal. • in affective communication, the meaning, form and signal are continuous and simply related. • in iconic communication, the form resembles the meaning in some way. Iconic systems can be either discrete or continuous 3 In fact in surveying more than twenty introductory text books in linguistics found in the Cambridge University book shop, I did not find even one that mentioned that language is used for communication!
- 49. Section 2.5. Summary 25 • in symbolic communication, the form does not resemble the meaning and the correspon- dences between these have to be learned. • Symbolic systems are discrete, and this and the arbitrary meaning/form relationship mean that communication of complex meanings can occur more effectively than with the other types. • Human language operates as a discrete symbolic system. Components of language • Human communication can be divided into two main components: – The verbal component describes the system of phonemes, words and sentences. It is a discrete symbolic system in which a finite number of units can be arranged to generate an enormous number of messages. – The prosodic component is by contrast mostly continuous, and in general expresses more basic meanings in a more direct • Affective prosody is when a speaker uses prosody to express emotion, speech act or other information which is conveyed sentence by sentence • Augmentative prosody is used to disambiguate and reinforce the verbal component. • Both types of prosody are optional, in that speakers can choose to use no prosody if they wish and just communicate via the verbal component. Writing in fact uses just this component. Communication • We can break communication into four processes in spoken communication: 1. generation is the conversion of meaning into form by the speaker 2. encoding is the conversion of form into signal by the speaker 3. decoding is the conversion of signal into form by the listener 4. understanding is the conversion of form into meaning by the listener • The equivalent processes exist for written communication also, where we have a writer and reader instead of a speaker and listener.
- 50. 3 THE TEXT-TO-SPEECH PROBLEM We now turn to an examination of just what is involved in performing text-to-speech (TTS) syn- thesis. In the previous chapter, we described some of the basic properties of language, the nature of signal, form and meaning, and the four main processes of generation, encoding, decoding and understanding. We will now use this framework to explain how text-to-speech can be performed. In TTS, the input is writing and the output speech. While it is somewhat unconventional to regard it as such, here we consider writing as a signal, in just the same way as speech. Normal read- ing then is a process of decoding the signal into the message, and then understanding the message to create meaning. We stress this, because too often no distinction is made at all between signal and form in writing; we here about “the words on the page”. More often an informal distinction is made in that it is admitted that real writing requires some “tidying up” to find the linguistic form, for instance by “text normalisation” which removes capital letters, spells out numbers or separates punctuation. Here we take a more structured view in that we see linguistic form as clean, abstract and unambiguous, and written form as a noisy signal that has been encoded from this form. The process of reading aloud then is one of taking a signal of one type, writing, and con- verting it into a signal in another type, speech. The questions we will explore in this chapter are: • How do we decode the written signal? • Does the message that we extract from the written signal contain enough information to generate the speech we require? • When someone is reading aloud, what are the relationships between the author, reader and listener? • What do listeners want from a synthetic voice? 3.1 SPEECH AND WRITING It is clear that natural (human) language developed first using sound as its medium. While humans clearly also communicate with facial and hand gestures and so on, these systems are not able to express the same wide range of expressions as natural language. Quite why human language 26
- 51. Section 3.1. Speech and Writing 27 evolved using sound as its medium is not clear, perhaps it was down to an ability to make distinct patterns more easily with the mouth, than say by waving hands. It is important to realise that this choice of sound may well have been a complete accident of evolution, today we communicate in many different ways with vision either by reading or by interpreting sign language. For most of human history, speech was the only (sophisticated) means of communication, but a second key invention, perhaps 10,000 years ago, was that of writing. To begin with, we need to look at the nature of speech and writing in more depth. We will do this from a number of angles, looking first at the physical differences between these signals, before moving on to look at how these differences influence communication. 3.1.1 Physical nature The speech signal is a continuous, acoustic waveform. It is created by the operation of the vocal organs in response to motor control commands from the brain. We only have a partial understand- ing of how this process works, but it is clear that the coordinated movement of the vocal organs to produce speech is extremely complicated. One of the main reasons why the relationship between the speech signal and the message it encodes is that speech is continuous and the form is discrete. In addition to this, speech is highly variable with respect to the discrete message as many other factors, including prosody, speaker physiology, a speaker’s regional accent and so on all affect the signal. A single message can generate many potential signals. Finally, speech is not permanent; unless a recording device is present, once something is spoken, it is lost for ever and if no listener is present or a listener does not hear the speech then the message that the speech encodes will never be decoded. Writing is fundamentally different from speech in a number of ways. Firstly writing is primarily seen as visual: traditionally, writing always took the form of visible marks on a page or other material, and today we are quite used to seeing writing on computer and other screens. This is not however a complete defining property; we consider braille a form of writing and its key benefit is of course that one does not need to be able to see to read it. It has been argued [199] that the main defining feature of writing is that it is permanent in a way that speech is not. Whether the writing is created with a pen on ink or impressions on a page, the key point is that once written it does not disappear. We can of course destroy what we have done, but this is not the point; the idea is that with writing we can record information, either so that we can access it at a later date, or send it over some distance (as a letter for instance). What then does writing record? Writing can be used to record natural language, but in saying this we should make it clear that writing attempts to record the linguistic form of a message, rather than the actual signal. That is, if we write down what someone says, we do not attempt to record all the information in the speech, such that a reader could exactly mimic the original; rather we decode the speech into a linguistic message, and then encode that. Of course quite often the message we wish to write has never been spoken, and so when writing we generate a linguistic form from our thoughts, and then encode this as writing without ever considering speech.
- 52. 28 Chapter 3. The Text-to-Speech Problem We explained in Chapter 2 that spoken language had a dual structure of sentences made of words, and words comprised of phonemes. We have a similar dual structure in written communi- cation, where sentences are made of words (as before), but where words are made of graphemes. Graphemes are in many ways analogous to phonemes, but differ in that they vary much more from language to language. In alphabetic languages like English and French, graphemes can be thought of as letters or characters (we will define these terms later); in syllabic writing like Japanese hira- gana, each grapheme represents a syllable or mora, and in languages like Chinese each grapheme represents a full word or part of a word. It should be clear that even apart from the visual and permanency aspects writing signals and speech signals are very different in nature. The most important differences is that writing is discrete and “close to the message”, while speaking is continuous, more complex, and does not encode the message in any trivial way. The complexities of the speech encoding and decoding arise because we are using an apparatus (the vocal organs) for a purpose quite different from their original one. In fact, one can state that it is a marvel that speech can be used at all for communication messages. In contrast, writing is a cultural invention, and was designed explicitly and solely for the purposes of communicating form. If we consider it a “good” invention, it should not be surprising that the encoding mechanism between linguistic form and writing is much simpler. This is not to say that the reading and writing process are trivial; it takes considerable skill for the brain to interpret the shapes of letters and decode messages from them; but this skill is one shared with other visual processing tasks, and not something specific to writing. The difficulty of decoding a written signal depends very much on the specific message. Often the relationship between the writing on the page and the words is obvious and transparent; but often it is complex enough for the decoding and encoding processes to be non-trivial, such that we have to resort to sophisticated and principled techniques to perform these operations. 3.1.2 Spoken form and written form We have just explained the differences between spoken and written signals. Turning now to form, we can ask how (if at all) do the spoken and written form differ? One way of looking at this is to consider the combined generation/encoding process of converting meaning into a signal. In general we can convert the same meaning into either a spoken signal or a written one, so we see that speech and writing (in general) share the same meaning representation but different signal representations. The question now is, how much of the overall language system do the two share? It is difficult to answer this conclusively, as to do so (in humans at least) would require a much more thorough understanding of the processes than we currently have. It is however possible to consider a model where the meaning to form systems for both spoken and written language share a common system from meaning to the second level of form. In other words, the meaning-to-word system is the same for both. The divergence occurs when we consider the primary level of form; in spoken language words are comprise of phonemes, in written language, they are comprised of graphemes. From this point until the signal, the two systems are separate. Depending on the
- 53. Section 3.1. Speech and Writing 29 language, the primary forms may be broadly similar (for instance in Spanish) such that one can deduce the pronunciation from the spelling and vice versa, or quite different, (for example in Chinese). While a few problems certainly occur, this model seems to fit the facts very well. Whether we write or say the sentence THE MAN WALKED DOWN THE STREET, there are six words and they occur in the same order in the spoken and written versions. This fact may be so obvious as to seem superfluous, but it is in fact vital to our discussion. One would find it difficult for instance to teach someone to read French text and speak English words: even if we learned that the pronunciation for chien was /d ao g/, we would find it difficult to understand why the written French sentence has different numbers of words in different orders from the spoken English version. The main point in this realisation of the commonality between the first level of form in written and spoken language is that it shows that if we can decode the form from the written signal, then that we have virtually all the information we require to generate a spoken signal. Importantly, it is not necessary to go all the way and uncover the meaning from the written signal; we only have to perform the job of text decoding and not text understanding. There are two qualifications we need to add to this. Firstly, some understanding may be required, or may make the job easier when we are resolving ambiguity in the written signal. For example, when encountering the homograph polish it may help in our disambiguation to know that the topic of the sentence is above people from Poland. Secondly, we may need to generate prosody when speaking, and this may not be obtainable from the writing. This issue is dealt with below (Section 3.5.5). This conclusion is so important for our system that it is worth restating again; by and large, the identity and order of the words to be spoken is all we require to synthesise speech; no higher order analysis or understanding is necessary. 3.1.3 Use Speech and writing are used for different purposes. When do we use one instead of the other? In certain cases we have no real choice; in face to face communication it is normal to speak: while we could write something down and pass it to our friend to read, this is in general cumbersome. When writing a book, or an article we normally use writing; that is the standard way in which material is stored, copied and distributed. Before the computer era it was fairly clear which communication mode was used for what purpose. Modern technology certainly gives us more choices; with telephones we can communicate over distance, whereas in the past this was only possible by letter. We can also now quickly send written messages electronically, by email, phone messaging, computer messaging and by other means. Computer or “instant” messaging is particularly interesting in that it assumes many of the characteristics of speech, but exists as a written signal. When phone messaging (“texting” or “sms”) first arrived, many though the idea absurd; why would you ever want to send a message in writing with a phone when you could simply call the person directly? In fact, many “functional” uses have been found; a common one is for someone to send an address as a text message as this
- 54. 30 Chapter 3. The Text-to-Speech Problem is easier for the receiver than having to listen and write the address down in case it is forgotten. Uses go beyond this though, and today we see “texting” as particularly popular among teenagers. Anecdotally, we believe that by using text, teenagers can think about, plan and edit their messages so as to seem more confident, funny and so on; this saves the embarrassment that may ensue from having to actually talk to someone. One major difference is that speech is in general spontaneous. During a normal conversa- tion, we usually speak with little planning or forethought; we can’t simply suspend the conver- sation until we have thought of the best possible way to say something. This accounts for the disfluencies in speech, such as hesitations, pauses and restarts. Writing is usually free of these effects, but it is important to realise that this is not because of any property of writing itself, but more because the writer can spend time over difficult sections, or can go back and correct what has been written. Another significant difference is that in spoken communication the speaker can often see or interact with the listeners. In face to face communication the speaker can see the listener and see how they react to what is being spoken; in addition to the listener responding directly (“Yes, I agree”), they may nod their head, give an acknowledgment or show confusion and so on. In response, the speaker can alter or modify their behaviour accordingly. This also happens over the phone, where verbal cues (“uh-huh”, “yeah...”) are given by the listener to help show the speaker their level of comprehension and agreement. This is less effective than with the visual cues; anyone who has taken part in a multi-person call can attest to the difficulty involved in sensing everyones understanding and participation in the conversation. If we compare this to a typical situation where written communication is involved, we see some important differences. For example, imagine someone is reading a newspaper on a train. Here the author of a particular column will not personally know the reader, in fact the author will probably know only a handful of people who read the column. The author and reader are separated in time and space and the communication is one way only (the reader may comment on the article to someone else, but this will not reach the author). Because of these factors, the situation with respect to common ground is very different. First, the author is writing for many readers, and all will have somewhat different degrees and types of background knowledge. Secondly, as no feed- back is possible, it is necessary for the author to “get it right first time”, there is no possibility of starting an article and then checking to see if the reader is following. These situational distinctions account for many of the differences in content between spoken and written language; for example, because the level of background information can not be known, writing is often more explicit in defining references, times and places and so on. Informally, we can describe the writing as being “more exact” or “less ambiguous”. Of course, these terms only make sense when considering the speech or writing in isolation; as explained above, most ambiguity is instantly resolved when the situation is taken into account.
- 55. Section 3.1. Speech and Writing 31 3.1.4 Prosodic and verbal content One of the most notable features of writing is that it nearly exclusively encodes the verbal compo- nent of the message alone: the prosody is ignored. This feature seems to be true of all the writing systems known around the world. Because of this it is sometimes said that written language is impoverished with respect to spoken language as it can only express part of the message. It is of course possible to represent effects such as emotion in written language, (one can simply write “I am sad”) but this is done by adding information in the verbal component and is different from the systematic use of prosody to indicate this emotion. Quite why prosody is not encoded in writing is not well understood. Possible explanations include the fact that it is often continuous, which creates difficulties if we are using a discrete coding system. Prosody is most powerful for emotive language; and while this is an important part of literature, the ability to express emotion directly is not needed so much for legal documents, newspapers, technical manuals and so on. We do however have a limited prosodic component in written language. Punctuation can be used to show structure; and underlined, italicised or bold text is used to show emphasis. These are clearly encoding augmentative prosody in that the punctuation and italics are there to help the reader uncover the correct verbal form from the writing. Affective prosody on the other hand is nearly entirely absent; there is no conventional way to shown anger, pain or sarcasm. The only aspect that is present is the ability to express sentence types such as question, statement and imper- atives with punctuation at the end of a sentence. It is interesting to note that in most early writing systems (e.g. Ancient Greek) no punctuation was used at all, but as writing evolved punctuation appeared and more recently italics and other font effects have become commonplace due to the wide availability of computer word processing. In the last few years, a system of “emotions” has become common in email and other electronic communication, in which punctuation is used to indicate happy faces :-) and other effects, which attempt to encode affective prosody. Quite why this trend is present is again not clear; while we can ascribe the rise of smiley faces to writing being used for more informal purposes, this doesn’t explain the trend from classical Greek to the English of 200 years ago with its commas, semi-colons and full stops. 3.1.5 Component balance From the outset, a person knows whether a message should be spoken or written: it is not the case that the person first generates the message and then decides what channel it should be sent in. This principle gives rise to another set of differences between spoken and written signals. Because he knows that a certain message is to be written, the person will, from the outset, structure the message so as to take account of the component balance available in that signal. So if for instance, if a writer wishes to convey something happy, he is quite likely to express this happiness in the verbal component, as the ability to encode affective prosody is not available. So he might say “I’m thrilled that you got that place” instead of perhaps “well done that you got that place”, which could give the impression of being somewhat unmoved. With appropriate expression, the “well done...”
- 56. 32 Chapter 3. The Text-to-Speech Problem version may sound very enthusiastic if spoken. The different use of components in written and spoken language is called component balance. Simply put, a speaker has a open choice of how much prosody or how much verbal content to put in a message; they will balance these against each other, to create a final signal that conveys the message they wish in a way that is balanced between efficiency and effectiveness. A writer on the other hand has only a very limited ability to use prosody; and so has to rely on the verbal component more. Because of this, he is likely to “beef-up” the verbal component, so that it is less ambiguous and more explicit than the spoken equivalent. In addition, the writer will “verbalise” any emotion if that is required (that is, explicitly explain the emotion by using words). It is this awareness of the prosodic deficiencies of written language and the ability to get around them that enables good writers to make themselves understood. It should be noted that the majority of people don’t really write anything more than the odd note or shopping list in their day to day lives. While we all learn to read and write at school, it is reading that is the essential skill to cope with modern life; only a few of us regularly write. This point is worth making because writing well, that is so that one communicates effectively, is actually quite difficult. One of the main aspects of this is that the ability to encode emotions, make writing unambiguous, and have a good picture of a potential reader in mind while writing, are all skills that require some mastering. We have seen this with the advent of email - many new users of emails send messages which are far too brief (e.g. “yes” in reply to something that was asked days ago) or that come across as being quite rude - the writer has written the words that he would normally use if he were speaking face to face, but without the prosody, the emotion (which could be humour or politeness) is lost and the words themselves seem very abrupt. 3.1.6 Non-linguistic content In our discussion so far we have equated speech as being the signal used to communicate linguistic messages acoustically and writing as the signal used to communicate linguistic visually or over time. This in fact simplifies the picture too much because we can use our speaking/listening and reading/writing apparatus for types of communication which don’t involve natural language. Just as we can use our vocal organs for whistling, burping, coughing, laughing and of course eating, breathing and kissing, we find a similar story in writing, in that we can use writing for many other purposes than just to encode natural language. This is a significant concern for us, in that we have to accurately analyse and decode text, and to do so we need to identify any non-linguistic content that is present. When considering “non- linguistic” writing, it helps again to look at the evolution of writing. The key feature of writing is that it is more permanent than speech it can be used to record information. Even the earliest forms of Egyptian and Sumerian are known to have recorded numerical information relating to quantities of grain, livestock and other “accounting” type information. Furthermore, there are several cultures who developed methods for recording and communicating accounting information without developing a system for encoding their language: the most famous of these is the Inca
- 57. Section 3.1. Speech and Writing 33 quipu system, in which numbers could be stored and arithmetic performed. This ability for writing systems to encode non-linguistic information is therefore an ancient invention, and it continues today in printed mathematics, balance sheets, computer programs and so on.[64], [19], [6] Consider for a moment numbers and mathematics. Without getting to far into the debate about what numbers and mathematics actually are, it is worth discussing a few basic points about these areas. Firstly, in most primitive cultures, there is only a very limited number system present in the language; one might find a word for one, two and three, then a word for a few and a word for lots. More sophisticated number communication systems (as opposed to numbers themselves) are a cultural invention, such that in Western society, we think of numbers quite differently, and have developed a complex system for talking and describing them. Most scientifically and mathemati- cally minded people believe that numbers and mathematics exist beyond our experience and that advancing mathematics is a process of discovery rather than invention. In other words, if we met an alien civilisation, we would find that they had calculus too, and it would work in exactly the same way as ours. What is a cultural invention is our system of numerical and mathematical no- tation. This is to some extent arbitrary; while we can all easily understand the mathematics of classical Greece, this is usually only after their notational systems have been translated into the modern system: Pythagoras certainly never wrote x2 +y2 = z2. Today, we have a single commonly used mathematical framework, such that everyone knows that one number raised above another (e.g. x2) indicates the power. The important point for our discussion is to realise that when we write x2 +y2 this is not encoding natural language; it is encoding mathematics. Now, it so happens that we can translate x2 +y2 into a language such as English and we would read this as something like “x to the power of two plus y to the power of two”. It is vital to realise that x2 + y2 is not an abbreviated, or shorthand form of the English sentence that describes it. Furthermore, we can argue that while mathematics can be described as a (non-natural) language, in that it has symbols, a grammar, a semantics and so on; it is primarily expressed as a meaning and a written form only; the spoken form of mathematical expressions is derivative. We can see this from the ease at which we can write mathematics and the relative difficulties we can get in when trying to speak more complex expressions; we can also see this from the simple fact that every mathematically literate person in the western word knows what x2 + y2 means, can write this, but when speaking has to translate into words in their own language first (e.g. EKS SQUARED PLUS Y SQUARED in English IX QUADRAT PLUS YPSILON QUADRAT in German). 3.1.7 Semiotic systems A semiotic system is a means of relating meaning to form. Natural language is a semiotic system, and probably has by far the most complicated form to meaning relationship, and as we have just seen mathematics is another. Most semiotic systems are much simpler than natural language as they are cultural inventions rather than biological processes (recall that while mathematics itself may be a property of the universe, the mechanism we use for its form and encoding are cultural
- 58. 34 Chapter 3. The Text-to-Speech Problem inventions). In addition to these examples, we also have computer languages, email addresses, dates, times, telephone numbers, postal addresses and so on. Just as with mathematics, a telephone number is not part of natural language, it is a complete meaning/form/signal system in its own right. If it were just a case of each semiotic system having its own writing system, we could design a separate system to process each. The problem arises because frequently we mix these systems in the same signal and using the same characters to do so. For example, in the above text, we saw the sentence (2) Pythagoras certainly never wrote x2 +y2 = z2. in which we clearly mixed natural language text with mathematical text. There are many other cases too; while the mathematics example uses different fonts and so on, many systems share the same characters as natural language writing, so that we get: (3) a. I was born on 10/12/67 b. It weighed 23Kg c. I’ll send it to paul@yahoo.com and so on. The point is that these numbers, dates and email addresses are encoded using the same signal as natural language writing. To be able to process this type of writing, we have to identify which parts of the text relate to which semiotic systems. 3.1.8 Writing Systems To the best of our knowledge1, the first writing system was that of Sumer which developed into Cuneiform. Perhaps the best known early writing system is the hieroglyphic writing of the Ancient Egyptians, although it is not entirely clear whether these two systems are related. Initially these systems were logographic in nature, meaning that one symbol usually rep- resented one word or one morpheme. Furthermore, the scripts were pictographic meaning and that the symbols were often iconic, ie pictorial representations of the entity the word stood for. As these systems developed, the iconic nature became less obvious, developing into a more abstract logographic script. Additionally alphabetic elements were developed in the language that allowed writers to “spell out” words. These developments should come as no surprise; in Section 2.1 we saw that it is quite natural for a communication system to start highly iconic and then progress to a state when the relation between meaning and form is more arbitrary as this general allows for experienced users to communicate more quickly and more accurately. Most writing systems can be described as belonging to one of the following groups, but in some cases, e.g. Japanese, the complete writing system uses more than one of these types. Logographic systems, as we have just seen use one symbols to represent words or morphemes. Syllabic systems use one symbol per syllable, such as in Japanese hiragana. Alphabetic systems 1 If only someone had taken the trouble to write down this key discovery!
- 59. Section 3.2. Reading aloud 35 generation transmission encoding decoding text text signal written written signal message meaning spoken signal Author Reader Listener meaning verbal message spoken signal verbal message speech encoding Figure 3.1 Basic model of reading aloud, which shows in the information flow between the author, reader and listener. use a small list of symbols which roughly correspond to the phonemes of a language. The latin alphabet is of course the most widely used, but its relation to phonemes varies greatly from lan- guage to language. Abjad systems are ones primarily based on consonants, vowels are usually not written. It is a basic fact of human language that most information that discriminates words is carried by the consonants, and so ignoring these in written language is possible (this is why the abbreviation for DOCTOR is dr and not oo). The semitic family of languages, including Hebrew and Arabic are abjads. Writing systems are an important consideration in TTS as in general we have to be able to design a system for any language, and this means being able to handle all writing systems. Prob- lems that have to be solved include, identifying the language, identifying the writing system (for example Japanese uses a logographic and a syllabic system), identifying the computer encoding used, the writing direction and many other issues specific to each language. 3.2 READING ALOUD We now turn to the problem of reading aloud, that is, how to read text and then speak what we have read. 3.2.1 Reading silently and reading aloud The process of reading, as we normally understand it, is one of taking some text, decoding it into a message, and then understanding it. For clarity, we call this process reading silently. By and
- 60. 36 Chapter 3. The Text-to-Speech Problem large, the process of decoding the text is relatively straightforward for us, and in say a newspaper, we can often successfully decode all the text without error. Understanding is more difficult, in that we may not completely understand an article, and we can perhaps think of various degrees of understanding that we might achieve with respect to a text. The other participant in this process is the author, and exactly the same factors govern the way she generates a message as with our conversation example of section 2; she assumes some background knowledge on behalf of the reader, makes an assumption about the reader’s situation, and is guided by the conversational principles outlined before. In addition to all these though, she generally assumes that the reader will be reading in the way just described; that is, decoding the writing into the message and then understanding this. This last point is highly significant because this carries with it an assumption that the spoken part of the communication process is not relevant to the task of normal, silent, reading. Examples of the effect of this include a situation where a journalist may include the name of a foreign person or place, but not have to explain how this should be pronounced. The word structure of the sentences can be different, because unlike speaking, there is no need to limit the length of phrase or sentences to chunks that are possible to say with one breath. Far more significantly though, there is little or no specification of prosodic information; it is assumed that the verbal content is enough to create the correct understanding, and emotional or other thoughts are communicated via the verbal component. These problems cause no problem for normal reading, as this speech isn’t involved in this process. Now consider the act of reading aloud, that is, reading the text and simultaneously speaking it. This is considerably more complicated in that it involves three agents; an author, a reader and a listener, and two types of signal, text and speech. A diagram of this process is shown in Figure 3.1. This type of reading of course does involve speech; and herein lies the problem, because in the majority of cases, the text that has been written has only been written to be read silently; it was never intended to be read aloud. The problem is that as we explained in Section 3.1.5, the component balance is now wrong; a message that was intended to be communicated by writing is now being communicated by speech. While sometimes this is unproblematic, often difficulties can arise from this mismatch, and we can class these difficulties into those to do with prosody and those to do with the verbal content and style. We now investigate these in detail. 3.2.2 Prosody in reading aloud We have seen already that the written signal contains little or no prosodic information, and when this is decoded, it is decoded into a message that has verbal content only; the prosodic content is generally missing. This is not a problem for reading silently, the reader can understand the thought without needing the prosodic information, and the assumption is that the author has made strong use of the verbal component to compensate for the lack of prosody. If the message is to be spoken aloud we face the issue of what and how much prosody to use. For messages which have little emotion and/or are highly propositional in content, there should not be much of a problem. We
- 61. Section 3.2. Reading aloud 37 decide to use neutral prosody, and the suprasegmental effects that need to be synthesised can be found from the verbal part of the message. For more emotive messages, or in cases where we think we need significant use of augmentative prosody, we have a fairly serious problem. This is because we have no means of knowing what prosody the speech should be encoded with; the message that we found from the text has no explicit clues as to what the prosody should be. The only solution is for the reader to generate their own prosody. This is very different from what they are doing with the verbal content; there they are simply decoding from one medium and encoding in another, no imagination, creativity or understanding is required. With prosody generation however, the message does not contain the necessary content, and so the speaker must generate this themselves. We don’t really know exactly how people perform this task, but it is plausible that some understanding of the message is undertaken, and from this an attempt is made to generate prosody that the speaker believes is appropriate for the verbal content. Another way of looking at this, is that they in some sense attempt to reproduce the prosody that the author would use if they had decided to speak the message rather than write it. It is crucial to realise the difference between the process of finding the message by decoding the writing, and the process of prosody generation. In the first, there is a clear right and wrong; the source generated this partial representation, it was clear and specific, and it is our job to find it. Any difficulties that may arise are solely because some distinctions in the structured representation are lost when the message is encoded in the signal (they are ambiguous). In the case of prosody generation, there can be no right or wrong; this information never existed in the first place and so we cannot recover it. 3.2.3 Verbal content and style in reading aloud Another issue concerns the situation where the reading process is taking place. In our canonical example, the author is writing with the intent that the reader silently reads the text. If the text is then read aloud, we face a potential problem in that the communication convention assumptions used to generate the message have been violated, and we are now converting the message into a signal that it wasn’t intended for. A simple example will demonstrate this. Many written documents are quite long; simplisticly we can say that this is necessary because the author has a lot to say and it may not be easy to get this across except in a long document (say a book). The author of course assumes that the reader will read this at a time that is convenient and at his own rate. Furthermore, while the author may not explicitly desire this, it is understood that the reader can skip ahead, glance through sections, or reread sections again. If we now consider the situation of a human or computer reading this document out loud, we see that many of these assumptions no longer apply; the speech progresses linearly from the start of the book, the rate of reading is more or less constant, and that the reader doesn’t directly decide when the reading should start and stop. Potentially, this can cause severe problems and we may find that although the system is reading and speaking the text perfectly adequately, the user is not satisfied with the experience because of these violations of communicative norms.
- 62. 38 Chapter 3. The Text-to-Speech Problem signal written spoken signal words decoding text encoding speech graphemes phonemes Figure 3.2 The common-form model showing the two processes of text decoding, which finds the words, and speech encoding from those words. For illustration we have shown the different primary levels of form for speech and writing as graphemes and phonemes to show that these are not directly connected. 3.3 TEXT-TO-SPEECH SYSTEM ORGANISATION The above discussion shed light on some of the issues of spoken and written communication, and the particular process of reading aloud. From our understanding of these processes, we can now define a model of how text-to-speech conversion by computer can be carried out. The next section describes the model we adopt throughout the book, and the sections after that describe alternative models and how these compare. 3.3.1 The Common Form model In the common form model, there are essentially two components; a text analysis system which decodes the text signal and uncovers the form, and a speech synthesis system which encodes this form as speech. The first system is one of resolving ambiguity from a noisy signal so as to find a clean, unambiguous message; the second system is one where we take this message and encode it as a different, noisy, ambiguous and redundant signal. The core idea is that writing and speech share a “common form” and that that, and that alone is the single and correct intermediate representation between the two types of signals. A diagram of this is shown in Figure 3.2. For purposes of basic explanation, we should take the word “form” here to mean “the words”. Hence, in analysis phase, we are trying to find the unambiguous, underlying words from the text; and in the synthesis phase we are attempting to create a speech signal from these words. In the basic common form model, we always read the words as they are encoded in the text; every word is read in the same order we encounter it. No understanding or other higher level analysis is performed, and as such everything is read with a neutral prosody. The key features of this model are; • There are two fundamental processes; text analysis and speech synthesis.
- 63. Section 3.3. Text-to-speech system organisation 39 • The task of analysis is to find the form, ie. the words, from the text. • The task of synthesis is to generate the signal from this form. • No understanding is required, an accurate decoding of the text into a linguistic form is suffi- cient. • Prosodic information can not in general be decoded from the text, and so everything is spo- ken with neutral prosody. 3.3.2 Other models A number of different models are found in TTS systems and we will briefly examine these now. These models are not mutually exclusive and many real life systems are a combination of two or more models. Signal-to-signal model In this model, the process is seen as one of converting the written signal into a spoken one directly. In such models, the process is not seen as one of uncovering a linguistic message from a written signal, and then synthesising from this, but as a process where we try and directly convert the text into speech. In particular, the system is not divided into explicit analysis and synthesis stages. Pipelined models Quite often the signal-to-signal model is implemented as a pipelined model were the process is seen as one of passing representations from one module to the next. Each module performs one specific task such as part-of-speech tagging, or pause insertion and so on. No explicit distinction is made between analysis and synthesis tasks. These systems are often highly modular, such that each module’s job is defined as reading one type of information and producing another. Often the modules are not explicitly linked so that different theories and techniques can co-exist in the same overall system. Text as language models In this model, the process is seen as basically one of synthesis alone. The text itself is taken as the linguistic message, and synthesis is performed from this. As the text is rarely clean or unambiguous enough for this to happen directly, a text normalisation process is normally added, as a sort of pre-processor to the synthesis process itself. The idea hear is that the text requires “tidying-up” before it can be used as the input to the synthesiser. Grapheme and phoneme form models This approach is in many ways similar to the common form model in that first a grapheme form of the text input is found, and this in then converted to a phoneme form for synthesis. Words are not central to the representation as is the case in the common form model. This approach is particularly attractive in languages where the grapheme-phoneme correspondence is relatively direct; in such languages finding the graphemes often means the phonemes and hence pronunciation can accurately be found. For other languages such as English, this is more difficult, and for languages such as Chinese this approach is probably impossible. If it can be performed, a significant advantage of the grapheme/phoneme form model is that, an exhaustive knowledge of the words in a language
- 64. 40 Chapter 3. The Text-to-Speech Problem is not necessary; little or no use is needed for a lexicon. This fact makes this approach still attractive for small footprint systems. Full linguistic analysis models The common form model is based on the idea that all we need to uncover from the text is primarily the identity of the words; we don’t make heavy use of any other linguistic form or meaning representations. Some systems go much further in terms of linguistic analysis and perform morphological analysis, part-of-speech tagging and syntactic parsing. To some extent, all these are useful for finding the words; the issue is really whether these should be performed as separate stand-alone tasks whose output feeds in to the word finding system, or whether we can find the words with a single, integrated approach. This issue is addressed in detail in Chapter 5. In addition to word identity detection, parsing and other types of linguistic analysis are often seen as being useful for helping with prosody. Complete prosody generation The common form model allows us to describe the three types of prosody independently. Suprasegmental “prosody” is modelled by the verbal component, and for utterances which have little affective prosody. This means that only augmentative prosody has to be explicitly generated. This contrasts with many systems which state that the F0, phrasing, stress and so on in an utterance are all directly determined by prosody, and so to generate speech we have to generate all these quantities with an explicit prosodic model. If this is the case, then the prosodic part of the system plays a much bigger role. Prosody from the text Following from the assumptions of the complete prosody model, we find that if every utterance requires a detailed prosodic specification, then this must somehow be generated. A common assumption is that the text does in fact contain enough information to determine prosody, and so many TTS systems have modules which try and predict prosodic representations directly from the text, often with the assumption that this is an analysis pro- cess with a right and wrong answer. 3.3.3 Comparison While some of the above systems may seem quite different from the common form model that we adopt, it should be realised that in many cases the differences can be somewhat superficial and that “under-the-hood” many of the same underlying operations are being performed. It formulates text analysis as a decoding problem, and this allows us to bring many of the powerful statistical text processing techniques to bear on the problem. This is a more powerful approach than the alternatives of either treating text as the synthesis input, which in some way must be normalised; or the approach of performing a full traditional linguistic analysis, which can often be errorful while the reason why we have chosen the common form model is that it simplifies the overall design of the system. The text decoding approach allows us to focus on a single well defined problem such that we are often able to perform this in a single, optimised system. Secondly, by avoiding the “complete prosody” assumption, we bypass many of the difficulties (and we believe impossibilities) that arise from trying to determine the prosodic content from the text. The common form approach is what one can also term an integrated approach. In many
- 65. Section 3.4. Systems 41 TTS systems, one often finds a huge list of modules arranged in a pipeline; a typical system of this kind might include modules for preprocessing, tokenisation, text normalisation, part-of- speech tagging, parsing, morphological analysis, lexicon, post-lexical rules, intonation, phrasing, duration, F0 generation, unit selection and signal processing. The problem with this approach is that if each module operates independently, it may only be marginally contributing to the overall process. Each module will have to make “hard decisions” to generate its output, and this can lead to a propagation of errors through the system. Furthermore, from an analysis of errors and performance, it may in fact be clear that many of these modules (e.g. a parser) would only be of marginal benefit even if their output was always correct. An integrated approach on the other hand aims to avoid these problems by optimising the system performance on a single metric. In the case of text decoding, the idea is to go from raw text to a word list in a single step. While some of the tasks performed by modules in a pipelined system may still be performed, they will be done implicitly, with few hard decisions until the final output is required. 3.4 SYSTEMS 3.4.1 A Simple text-to-speech system To demonstrate the common form model, let us now sketch how a text-to-speech system actually works. The input text arrives as a sequence of ascii characters, which can be of any length. To make the processing more manageable, we break input text into separate sentences using a sen- tence splitting algorithm. The input may only have one sentence, but we don’t know this so we always attempt to find sentence boundaries. For each sentence, we then further divide the input into a sequence of tokens, based on the presence of whitespace, punctuation and so on. Often tokens are just the written encodings of individual words, but can also be encodings of numbers, dates and other types. Next we find the semiotic class of each token. For non-natural language tokens, we use a separate system for each type to decode the text into he underlying form, and then using rules, we translate this form into a natural language form with words. For natural language tokens, we attempt to resolve any ambiguity and so find the words. We attempt a basic prosodic analysis of the text. Although much of the information we might ideally like is missing from the text, we do the best we can and use algorithms to determining the phrasing, prominence patterns and intonational tune of the utterance. This concludes the text and prosodic analysis phase. The first stage in the synthesis phase is to take the words we have just found and encode them as phonemes. We do this, because this provides a more compact representation for further synthesis processes to work on. The words, phonemes and phrasing form an input specification to the unit selection module. Actual synthesis is performed by accessing a database of pre-recorded speech so as to find units contained there that match the input specification as closely as possible. The pre-recorded speech can take the form of a database of waveform fragments and when a particular sequence of these are chosen, signal processing is used to stich them together to form a
- 66. 42 Chapter 3. The Text-to-Speech Problem single continuous output speech waveform. This is essentially how (one type) of modern TTS works. One may well ask why it takes an entire book to explain this then, but as we shall see, each stage of this process can be quite complicated, and so we give extensive background and justification for the approaches taken. Additionally, while it is certainly possible to produce a system that speaks something with the above recipe, it is considerably more difficult to create a system that consistently produces high quality speech no matter what the input is. 3.4.2 Concept to speech It is not strictly speaking necessary to start with text as input in creating synthetic speech. As we have clearly stated, our TTS model is really two processes, an analysis one followed by a synthesis one, and these are quite different in nature. As the analysis system will never be perfect, it is quite reasonable to ask whether there are situations where we can do away with this component and generate speech “directly”. Often this way of doing things is called concept-to-speech in contrast to text-to-speech. We shall stick with the generally used term “Concept-to-speech”, but we should point out that this can however mean a number of different things which we will now explain. Clearly, in our model, if we do away with text analysis, then the input to the system must be the input to the synthesiser itself, that is, the word-based form or message. We could call this “message-to-speech” or “form-to-speech” synthesis. The idea is that a message-to-speech system would be a component in a larger computer natural language system, and that the message would be generated by a natural language generation system. As this system knows whether it said PROJECT-NOUN or PROJECT-VERB, it can pass this directly into the synthesiser and so bypass any errors the text analysis system may make here. While this would certainly bypass any inadequacies of text analysis, it does suffer from a significant problem in that we don’t have a standard definition of what form actually is. Even with systems that conform to the common-form model, each may have a slightly different idea of what this form should take; one system may just use words, another may use syntax as well. We therefore face a situation of each system potentially exposing a quite different interface to an external system developer. A standard interface could be designed, but standardisations of systems often progress along a lowest common denominator strategy and may end up being somewhat ineffective. An alternative strategy is to propose a “meaning-to-speech” system, in which semantic rep- resentations are the input. Here the message generation and speech encoding parts would be combined into a single system. While this bypasses the nature of how the message should look, it opens up an even thornier problem in that we have even less agreement about how semantic representations should look. Such a system would have a significant advantage in that it could solve the prosody generation problem. As we are synthesising directly from meaning, it would in principle be possible to generate prosody in the signal directly, and so bypass problems of prosodic impoverishment in the text, or problems with how to represent prosodic form. A final solution is that we use a TTS system as before, but augment the input so as to reduce
- 67. Section 3.4. Systems 43 ambiguity and explicitly show the system where to generate prosodic effects. This can be done by a number of means, including the use of XML or other markup (explained in Chapter 17). While this will inevitably lack the power and fine control of the two above methods, it has the advantage in that the input is in a more standard form, such that a system developer should be able to get this working and switch from one system to another. It also has the advantage in that one can easily vary how close the system is to “raw” text or “clean” text. The message-to-form and augmented-text approaches are well covered in this book in the sense that they can be adapted from our main common-form TTS system. Meaning-to-speech is considerably more difficult and has only been researched in a very tentative manner to date. Mainly for this reason, we do not cover this approach in any detail here. 3.4.3 Canned Speech and Limited Domain Synthesis Today, most applications which use speech output do not use text-to-speech. Rather they make use of a set of recordings which are simply played back when required. These are called canned speech or pre-recorded prompts depending on the application. A typical way these are deployed is that a designer creates an exhaustive list of the utterances that are needed for an application. Quite often these utterances are specific to that application and a new set will be created and recorded for a different application. Once the utterances have been designed, a speaker is asked to read these, they are recorded and then stored as waveform files. The application then plays back the required file at the required time. A typical system of this kind might be deployed in a train station announcement system or an automated credit card booking system, where the users interacts with the system by speech recognition and automated dialogue control. Such systems initially compare unfavorably with TTS in that they require a new set of ut- terances for each application, compared to a TTS system which would just be deployed once and can be used for any application. Furthermore, the canned speech approach can only say a very fixed number of things, which can limit the scope of the application (for instance it would be very difficult to speak a user’s name). Finally, if the application has to be updated in some way and new utterances added, this requires additional recordings which may incur considerable difficulty if say the original speaker is unobtainable. Despite these apparent disadvantages, canned speech is nearly always deployed in commercial systems in place of TTS. Part of the reason behind this is technical, part cultural. Technically canned speech is perfectly natural and as users show ex- treme sensitivity to the naturalness of all speech output this factor can outweigh all others. In recent years, TTS systems have improved considerably in terms of naturalness, and so it is more common to find TTS systems in these applications. There are other non-technical reasons; canned speech is seen as a simple, low-tech solution whereas TTS is seen as complex and hi-tech. The upshot is that most system designers feel that they know where they stand with canned speech, whereas TTS requires some leap of faith. There may be purely business reasons also; while the canned speech approach incurs up front cost in an application, it is a one off cost, and does not increase with the size of deployment. TTS systems by contrast can often be sold like normal soft-
- 68. 44 Chapter 3. The Text-to-Speech Problem ware where extra licences (and hence cost) is required when the size of the deployment increases. The main technical drawback of canned speech is that it can only say a fixed number of things. This can be a severe drawback in even simple applications where for example a telephone number is to be read in an answer machine application. A common solution is to attempt to splice together recordings of individual words or phrases so as to create new utterances. The result of such operations varies greatly from acceptable (but clearly spliced) to comically awful. That said, even the resultant poor naturalness of this is often chosen over TTS in commercial situations. Faced with the choice between fully natural but inflexible canned speech, and somewhat un- natural but fully flexible TTS, some researchers have proposed limited domain synthesis systems which aim to combine the benefits of both. There are as many different approaches to this. Some systems attempt to mix canned speech and TTS. Black and Lenzo [52] proposed a system for clev- erly joining words and carrier phrases, for use in applications such as a talking clock. The phrase splicing approach of Donovan et al [139] used recorded carrier phrases and cleverly spliced with unit selection synthesis from recordings of the same speaker. A somewhat different approach is to use what is basically the same system as for normal unit selection synthesis but to load the database of recordings with words and phrases from the required domain [440], [436], [394], [9]. 3.5 KEY PROBLEMS IN TEXT-TO-SPEECH In this section we identify the main challenges in high quality text-to-speech. First we describe four areas which we believe are the key problems that have been recognised as the main goals of TTS system building (but note, the problems are not always described in this way). Then we go on to describe two additional problems, which are starting to attract more attention and which we believe will be central to TTS research in the future. 3.5.1 Text classification with respect to semiotic systems As we explained, it is a mistake to think that text is simply or always an encoding of natural language. Rather, we should see text as a common physical signal that can be used to encode many different semiotic systems, of which natural language is just one (rather special) case. There are two main ways to deal with this problem. The first is the text-normalisation approach, which sees the text as the input to the synthesiser and tries to rewrite any “non-standard” text as proper “linguistic” text. The second is to classify each section of text according to one of the known semiotic classes. From there, a parser specific to each classes is used to analyse that section of text and uncover the underlying form. For natural language the text analysis job is now done; but for the other systems an additional stage is needed, where the underlying form is translated into words. Let us consider the semiotic class approach. Assume for a moment that we can divide an input sentence into a sequence of text tokens, such that the input sentence
- 69. Section 3.5. Key problems in Text-to-speech 45 (4) Tuesday September 27, 10:33 am ET would be tokenised as (5) Tuesday September 27, 10:33 am ET Semiotic classification is therefore a question of assigning the correct class to each of these tokens. This can be done based on the patterns within the tokens themselves (e.g. three numbers divided by slashes (e.g.10/12/67) is indicative of a date) and optionally the tokens surrounding the one in question (so that if we find “1967” preceded by “in” there is a good chance that this is a year). Once the class has been identified, it is normally quite straightforward for classes other than natural language to analyse this token to find its underlying form. In dealing with times for instance, the text token might say 10:45am and from this we would find a form something like (hours=10, minutes=45, time of day=morning). Finally, we can translate this into words. In one sense this is often straightforward as there are often simple well-accepted rules which describe this. In our time example, the word version could be read as TEN FORTY FIVE A M, and it is fairly easy to see how the form-to-words rules would be written for this example. Two issues complicate the matter however. Firstly, often there are many different legitimate ways to translate the same form; with our time example we could have (6) a. A QUARTER TO ELEVEN B. A QUARTER OF ELEVEN C. A QUARTER TO ELEVEN IN THE MORNING D. TEN FORTY FIVE IN THE MORNING E. TEN FORTY FIVE F. TEN FORTY FIVE A M and so on. All are legitimate, but as the system can read only one, we have to be sensitive to which one this should be. Sometimes preferences depend on the user; in particular a British English speaker may prefer one, an American speaker another. In addition, the context and genre of the text may have an effect; in an airline reservation system it would be useful to make absolutely sure the user knows the time is in the morning, in a business appointment system this may be obvious.
- 70. 46 Chapter 3. The Text-to-Speech Problem The important point to realise is that there is often no single correct answer to the translation issue, and so systems need to be sensitive to the user expectations. A further difficulty with translation is that for some content there is no commonly agreed rendering in words. While times, dates and currencies normally present no problem, classes such as email addresses, computer programs and other technical information has no agreed reading. For example, taking the email address: (7) pat40@cam.ac.uk We can easily parse this into user = pat40 top level domain = uk secondary domain = ac local domain = cam In effect as all email programs do just this. The problem is how should we read this? Should pat40 be read P A T FOUR ZERO, P A T FORTY, or PAT FORTY? In summary then, there are three tasks involved here. 1. Identify which semiotic class a token belongs to. This is often the most difficult task. 2. Analyse each token to find its underlying form. Apart from the case of natural language, this is often quite easy as the semiotic systems are cultural inventions and have been designed to be easy to understand. 3. Translate into natural language. It is often quite easy to generate at least one legitimate translation but sometimes difficult to generate the particular version that a user expects. 3.5.2 Decoding natural language text While most semiotic systems have a fairly simple and direct signal-to-form mapping, this in gen- eral is not so for natural language. Once the text has been properly classified in terms of its semiotic class, the general level of ambiguity with respect to form is greatly reduced; it is only in the area of natural language where significant ambiguity still remains. Ambiguity occurs in many different guises. The most basic is homograph ambiguity, where two words share the same form. In cases such as (8) lets now project the image and (9) the project was running behind schedule we see that two separate words PROJECT-VERB and PROJECT-NOUN both share the same letter sequence, but have different meanings and different pronunciations. In addition to homograph ambiguity, syntactic ambiguity is also reasonably common. In the example (from [363])
- 71. Section 3.5. Key problems in Text-to-speech 47 (10) Police help dog bite victim we (at least) two different possible syntactic patterns that could give rise to these tokens: (11) a. (POLICE HELP DOG) BITE VICTIM b. POLICE HELP (DOG BITE VICTIM) The question when building a TTS system is just how much of this ambiguity we actually need to resolve. A comprehensive approach would attempt to resolve all the ambiguity in the text and create a rich, structure for the form. In recognition that the more we attempt the more likely we are to make an error, and also in recognition of the basic engineering principle of efficiency, we could also take a minimalist approach, which would adopt the principle that we should only resolve those ambiguities that affect pronunciation. The word SET is used in many different ways in English, but it is always spelt the same and always pronounced the same, and so we can ignore the multitude of meanings and uses. If we know that some text could be two or more underlying forms, but realise that all these forms are spoken the same way, then we shouldn’t bother resolving the ambiguity. 3.5.3 Naturalness We take the view that a major goal of TTS research is to make the system as natural sounding as possible. By natural we mean that the system should sound just like a human; and so eventually we would like our TTS system to be able to create the full range of human speech. Sometimes the question is raised as to whether we really want a TTS system to sound like a human at all. The concern is raised that if the system sounds too much like a human listeners will get confused and mistake it for a person. We think that in reality this is highly unlikely. All the time we hear “perfect” recordings of people’s voices, view photographs, watch films and so on, and are rarely confused. In fact, most of us are used to hearing recorded voices on the telephone in the form of people’s answering machines, introductory messages or recorded prompts in IVR systems. Rarely do these confuse us. So no matter how good a system is, it will rarely be mistaken for a real person, and we believe this concern can be ignored. We still have to ask however whether people really want natural sounding voices: wouldn’t they be happy with a “robotic” sounding one instead? (which would after all be much easier for us to produce). This is really an empirical question that needs to be resolved from user trials. From what we know however, the answer is over whelming in that people give much higher satisfaction and acceptance ratings to natural sounding systems. In fact we can go so far as to say that most listeners are extremely intolerant of unnaturalness to the extent that they will refuse to use non-natural sounding systems regardless of what other benefits are provided. Quite why this should be so is somewhat of a mystery; after all in many situations we are quite happy with visual caricatures of people as evidenced by cartoons and animation. Somehow we don’t consider a drawing of Homer Simpson a pathetic rendition of reality, but we are insistent on him having a natural voice, such that in the Simpsons (and all other cartoons) a real actor is used for his voice.
- 72. 48 Chapter 3. The Text-to-Speech Problem What then makes one voice sound more natural than another? Again, this is a question that we are not in a position to answer fully, but we can provide a sketch of the factors involved. Firstly, any system that produces obviously non-human artefacts in the speech will readily be judged as unnatural. It is all too easy to generate speech with pops, clicks, buzzes and an endless variety of other mechanical sounds. Even speech that is absent from any of these “error” type of sounds can readily sound unnatural. While speech exhibits considerable variability in the realisation of a particular phoneme, this is controlled by a number of factors including phonemic context, position in sentence, supra-segmental influence and so on. Getting this variation just right is a key task. Beyond this, we can consider a number of issues which are related to individuality. In normal speech, any given instance of speech is obviously speech from someone; there is no such think as speaker-neutral speech. So in a sense, we are completely conditioned to the fact that speech comes from people, all of whom sound slightly different. So when listening to a synthesiser, one naturally thinks that it is from someone, and to sound natural, this someone must have a believable sounding voice. This aspect of naturalness has virtually nothing to do with the message being spoken, it is entirely a property of the system that encoded the speech. Hence we must make the synthesiser sound like someone, either a copy of a real person’s voice, or the creation of a new voice that could pass for a real person unproblematically. 3.5.4 Intelligibility: encoding the message in signal The final central task is intelligibility. For our purposes, we define intelligibility as the ability of a listener to decode the message from the speech. As such, this does not include any measure of comprehension directly, but the idea is that if the listener can decode the message properly they should be able to understand it with the same ability as they could the same message spoken by a human. Because of that we can sometimes measure intelligibility with comprehension tests; this is usually done with content that may be hard to decode, but is often not difficult to understand (e.g. a simple sentence address or telephone number). Of all the TTS problems, intelligibility is probably the easiest to solve. In fact, it is possible to argue that this problem was solved some time ago, as was possible to generate fairly intelligible sounding speech in TTS systems from the late 1970s onwards. In fact, it can be shown in tests that the intelligibility of a modern TTS system is often only marginally better than much older systems such as MITalk [10], once any text errors have been dealt with2. From this it is wrong to conclude that little progress has been made on this issue. Crudely speaking, the story is that while the intelligibility of older systems was reasonably good, research focused on improving naturalness which was uniformly rated as being very poor with these systems. Hence the idea in the research community (though rarely stated as such) was to improve naturalness without making intelligibility any worse. 2 Today’s systems are massively more accurate with respect to text analysis, and so any text processing errors in systems such as MITalk have to be eliminated if a comparison of synthesis encoding with respect to intelligibility is to be performed.
- 73. Section 3.5. Key problems in Text-to-speech 49 The relationship is by no means deterministic or simple, but it is possible to see that in many synthesis paradigms there is an inverse correlation between naturalness and intelligibility. To achieve high intelligibility, formant systems and early diphone systems used very “safe” speech, which was well articulated and close to the average realisation for the message in question. In a sense, this speech was too safe, and containing none of the fluency, slurring, rhythm and heavy coarticulation found in real speech. As these effects are added, there is a chance that they will be added inappropriately, and in so doing will actually cause a loss of intelligibility. The trick therefore is to add in extra naturalness by synthesising variation more appropriately, but to do so in a way that does not cause unnatural variation to occur. 3.5.5 Auxiliary generation for prosody We now consider the first of the more advanced problems; specifically how to create speech that encodes the full range of prosodic effects and not just the neutral prosody of most of today’s systems. Recall that this is difficult because while the text encodes the verbal message that the au- thor generated, it does not do the equivalent with prosody. The author of course (subconsciously) knows this. As such, it is not the case that he generates a message with verbal and prosodic content as he would do if speaking, rather he generates a message that only contains verbal content; the prosody was neither generated or encoded, and so it we can never recover this from the writing. How then are we to generate prosodic content if we can’t uncover it from the text? Firstly, let us separate the issue of generating a signal which contains a particular prosodic effect (for example surprise, or accentuation of a particular word) from the issue of deciding how to do this automatically from the text. The issue of the realisation of a particular prosodic form is by no means easy, but it is certainly more tractable than the second issue. In essence it is no different from other parts of the synthesis problem in that we can collect data with this effect, study how supra-segmental features vary with respect to this effect and so on. It is really the second issue; that of how to determine automatically which prosodic effects to use from the text that is the more difficult problem. We term this process auxiliary generation to show that it is a generation (and not text decoding) issue, and that this is done as a separate, auxiliary task to the process of verbal decoding and encoding. How then are we to approach this problem of auxiliary generation? First let us state that just because there is no definite right or wrong, this doesn’t imply an “anything goes” policy. One would not think highly of a system that read a message “we are regret to inform you of the sad demise of Mr...” with a humorous or flippant paralinguistic tone. Furthermore, we probably wouldn’t think too highly of a system that placed emphasis as follows “THE share price is AT and all TIME high because of uncertainties caused BY the forecast of PROFESSOR Merton...” and so on. Now there is nearly always some context which we can construct that makes the emphasis of any particular word seem reasonable. However, using the same knowledge and context that we used to decode the writing, we can see that some emphasis patterns seem more appropriate than others. This leads us to the proposal that the approach to prosody generation should be one of assumed intent, which means that we
- 74. 50 Chapter 3. The Text-to-Speech Problem have to make a guess as to what the author might have specified in the structured representation if he had intended to speak rather than write the message. This is by no means a certain or easy task but it does give us a reasonable basis on which to solve this problem. We should make two final points about this. Firstly, as we are making an assumption as to what the author might have done, there is no clear right or wrong about this task. Secondly, we have to from the outset build the notion of generation choice into our model of auxiliary completion: just as the author can choose to represent a thought in a number of different ways, we have to assume that he could have exercised choice over the prosodic and paralinguistic components. Therefore, as well as being uncertain because we are making an assumption, it is also perfectly valid to assume that there are a number of equally valid choices. 3.5.6 Adapting the system to the situation The issue here is that unlike the “normal” communicative situations of reading silently or having a conversation; in the case of reading aloud we have three agents; the author, the listener and the reader (i.e. the TTS system). Most text was never written with the intention that it be read aloud, and because of this, problems can occur in that faithful readings of the text can lead to situations where the reader is speaking something that the listener can not understand, has no knowledge of or is not interested in hearing. When people take on the role of the reader, they often stray from the literal text, using explanatory asides, paraphrasing, synonyms and other devices to make the author’s message understood. Just how faithful the reader is to the text depends very much on the situation; we sometimes describe a good reader as someone who has a good awareness of both the listener and the intent of the author and as such can steer a good compromise between faithfulness and acceptability. Very few TTS systems themselves make any serious attempt at solving this issue; it is nor- mally seen as a problem for the application to resolve such issues. Even so, it is possible for the TTS to facilitate better control; this can be done by providing natural and easy to control ways to change the speed of the speaking, to have a phrase repeated, a phrase skipped and so on. 3.6 SUMMARY Speech and writing • We can consider both speech and writing/text as signals which encode a linguistic, or other, message. • The primary functional difference is that writing is an invention used to record messages; as such it is significantly “closer” to the message than is the case with speech. • Most writing is written with the intention that the reader reads it silently. This can present difficulties if this writing is then read aloud.
- 75. Section 3.6. Summary 51 • Not all writing encodes linguistic messages, it is used to encode messages in many other semiotic systems also. Reading aloud • To a large extent, we can perform the task of reading aloud, by decoding a written signal into a message, and the re-encode this into a speech signal. • In many cases, no prosody is required, but in others a more sophisticate approach is to attempt to generate prosody that is appropriate for the message. This can not however by decoded from the text as that information was never encoded there in the first place. • These observations lead us to build our system within the framework of the common form model. This has two fundamental processes. 1. Text analysis; a decoding process that finds the message from the text 2. Speech synthesis; a encoding processes that creates a signal from the message. Text to Speech Key Challenges • We can identify four main challenges for any builder of a TTS system: 1. Semiotic classification of text 2. Decoding natural language text 3. Creating natural, human sounding speech 4. Creating intelligible speech • We can also identify two current and future main challenges 1. Generating affective and augmentative prosody 2. Speaking in a way that takes the listeners situation and needs into account.
- 76. 4 TEXT SEGMENTATION AND ORGANISATION The next three chapters of the book deal with how to extract linguistic information from the text input. This chapter can covers various pre-processing issues, such as how to find whole sentences in running text or how to handle text various markup or control information in the text. Chapter 5 describes the main processes of text analysis itself, such as how to resolve homograph ambiguity. Finally, Chapter 6 describes how to predict prosody information from an often impoverished text input. In many ways, this subject shares similarities with text analysis. There is an important difference however in that while we can view text analysis as a decoding problem with a clear right and wrong, prosody prediction has no strict right and wrong as we are attempting to determining prosody from an underspecified input. 4.1 OVERVIEW OF THE PROBLEM The job of the text analysis system is to take arbitrary text as input and convert this to a form more suitable to subsequent linguistic processing. This can be thought of as an operation where we try to bring a sense of order to the often quite daunting range of effects present in raw text. If we consider (12) Write a cheque from acc 3949293 (code 84-15-56), for $114.34, sign it and take it down to 1134 St Andrews Dr, or else!!!! we can see that this is full of characters, symbols and numbers all of which have to be interpreted and spoken correctly. In view of our communication model (Chapter 2), we can more precisely define the job of the text analysis system to take arbitrary text and perform the task of classifying the written signal with respect to its semiotic type (natural language or other, e.g. a date), decoding the written signal into an unambiguous, structured, representation, and in the case of non-natural language, verbalising this representation to generate words. Importantly, the text analysis process only involves finding an underlying word form from the input, it does not attempt any understanding or further analysis. A simple overview of the processes involved in text analysis and prosody prediction is given below; 52
- 77. Section 4.2. Words and Sentences 53 1. Pre-processing: possible identification of text genre, character encoding issues, possible multi-lingual issues. 2. Sentence splitting: segmentation of the document into a list of sentences. 3. Tokenisation: segmentation of each sentence into a number of tokens, possible processing of XML. 4. Text analysis (a) Semiotic classification: classification of each token as one of the semiotic classes of natural language, abbreviation, quantity, date, time etc. (b) Decoding/parsing: finding the underlying identities of tokens using a decoder or parser that is specific to the semiotic class. (c) Verbalisation: Conversion of non-natural language semiotic classes into words. 5. Homograph resolution Determination of the correct underlying word for any ambiguous natural language token. 6. Parsing Assigning a syntactic structure to the sentence 7. Prosody prediction Attempting to predict a prosodic form for each utterance from the text. This includes (a) Prosodic phrase break prediction (b) Prominence prediction (c) Intonational tune prediction In this chapter, we concentrate on a number of preliminary, document processing and archi- tectural issues. Firstly we attempt to define the precise goal of most of these operations, which is to uncover the underlying linguistic form. Mostly, this involves defining a precise notion of what we mean by a word and a sentence and how it relates to its spoken and written form. Secondly, we examine many of the initial tasks mentioned above, specifically those of classifying documents, handling markup, splitting documents into sentences and splitting sentences into tokens. The to- kens which are an end result of this are then used as the input to the further text analysis modules. Finally, we discuss how to store and communicate information within a TTS system. 4.2 WORDS AND SENTENCES This section examines the notion of just what we mean by a word and a sentence. We may think we are so familiar with these entities that we don’t have to consider it further but as we saw in Chapter 2 we nearly all make one basic confusion which is to confuse words with their written form. The next sections therefore attempt to define the notions of words and sentences with some rigour.
- 78. 54 Chapter 4. Text Segmentation and Organisation 4.2.1 What is a word? In the model of text analysis used in this book, we always maintain a clear distinction between a word and its written form. The basic idea is that each word is a unique linguistic entity, which can be expressed in either speech or writing. To do so, it must be expressed as a form, and so we talk of the spoken and written form of a word. People rarely confuse a word with its pronunciation, but often the identity of a word and its written form are confused, so we will take extra care to explicitly show the difference. This is highly beneficial because we can then see text analysis as a process of determining words from text, rather than a process of say assuming that there are words actually in the text, such that all text analysis essentially becomes a process of tidying up these words (e.g. expanding abbreviations). Making text analysis a process of uncovering words from superficial text is a cleaner and more powerful formulation. The idea behind our use of underlying words is to try to define some notion of word that captures the essential essence of the linguistic unit of word, but is free from distinctions that are not directly linked to word identity itself. Hence while the capitalisation difference between Captain and captain may provide some useful information, in essence, these are written variants of the same word. We use the concept of word to reflect this, and so we state that the CAPTAIN is the underlying word of Captain and captain. Next, we use this concept to help us with the ambiguity problem, namely that sometimes difference words have the same written form. If we consider the text form Dr we see that this is highly ambiguous as it can mean DOCTOR or DRIVE, or less commonly DOOR (4Dr car). Dr is an abbreviation, ie an orthographicly shortened form of a written word. Abbreviations are often created by leaving characters out (usually vowels) of the original word, and the omission of these characters can create the ambiguity. So the full written form of the words DOCTOR and DRIVE are not confusable, but the abbreviated forms are and so we term these abbreviation homographs. Note that the abbreviation only happens in the written form of the language: when people speak they say /d aa k t er/ and /d r ay iv/ in its full form, and not /d r/. We also have words such as polish which is ambiguous (someone from Poland and some- thing to shine your shoes with), but not because of a short cut: there is no other way to write this word. Such words are called true homographs as the ambiguity is not caused by writing style or abbreviation, but rather due to the defined spelling of the word. There are two ways in which true homographs come into a language. First, there are many words such as polish in which it is a pure coincidence that the two words have the same written form. These are called acci- dental homographs. Secondly, we have words which are related, such as record in I want to record this and just for the record. These words are clearly related and form a verb/noun pair and such homographs are called Part-of-speech homographs. We finish this section with a note on the definitions and terminology we will use for the remainder of the book. The term word will always be used for the unambiguous, underlying identity of a word, and never used to indicate its written or spoken form. Words will always be marked in a SMALL CAPS FONT. The terms writing, orthography and text are used more
- 79. Section 4.2. Words and Sentences 55 or less interchangeably, and all text will be display in a courier font. A Token is a term used for the written form of a word (for reasons explained below). The written form of a word is often described in terms of graphemes, letters or characters. Here we will use the term character throughout, as it represents the normal computer science view as a single unit of writing. Characters are always defined in terms of a standardised computer character set such as ASCII, and so when working with a character set each character is simple a member of that set. The term grapheme is often taken as the minimal unit of written language, but is quite confusing in some respects because often common character sequences such as ch and th in English are considered graphemes. As we only deal with computerised sequences of characters as input, we will use the term character throughout, and avoid grapheme and letter except to name the field known as grapheme-to-phoneme conversion (This is the standard name and we will stick to it; in our scheme strictly speaking this would be termed character-to-phoneme conversion.) We will reserve the term letter to mean the word which we used to speak a character, so that we speak of “the letters ABC” we that we are talking about three individual words called LETTER-A, LETTER-B and LETTER-C. 4.2.2 Defining words in text-to-speech A word is a linguistic device that relates form and meaning. In spoken language, words are composed of sequences of phonemes, and we call this the form of the word. The key concept of words is that the relation between form and meaning is arbitrary: we have /d ao g/ being the form of DOG and /k ae t/ being the form of CAT but there is nothing canine about the particular phone sequence /d ao g/, virtually anything would suffice. The association of forms and meanings has arisen over the years during the development of the language; and it is important to note that people need no knowledge of how a word developed or where it came from to use it properly. With the advent of written language, words where given an additional form to the spoken one. In alphabetic languages, there is often an clear relationship between the spoken and written form, as we see in the two forms look and /l uh k/. Often though this relationships isn’t so clear and both the spoken and written form of a word have to be learned. In logographic writing systems (such as Chinese) there is no connection whatsoever between the spoken and written forms. We can represent the relationship between spoken form, written form and meaning as follows: meaning WORD orthography pronunciation type of table or seat BENCH bench /b eh n ch/
- 80. 56 Chapter 4. Text Segmentation and Organisation In a “perfect” language1, every word would have a unique meaning, written form and spoken form; upon encountering a written form, there would be no doubt as to what it meant, upon wishing to express a meaning, only one spoken or written form could be chosen. In such a language, enumerating the words in a language would be trivial. Of course, this is not the case, and mainly by historical accident many words have arisen which share one of more of the written form, spoken form or meaning. Because of this, identifying the set of words is not trivial, and we have to make decisions about whether bank, polish and record are one word or two. We will now study this issue in some detail with the aim of determining an exact definition of word that is best suited for text-to-speech. In doing so, our basic method is to use form and meaning distinctions between pairs of potential words. Idealistically, we can think of the following exercise as one where we have all the form-meaning pairs in a language and one by one compare them and make a decision as to whether they are the same word or not. Let us begin by saying uncontroversially that we can assume that words which are different in orthography, phonetics and meaning are in fact different words. So if we have: animal xa dog /d ao g/ fruit xb apple /ae p ax l/ it is same to assume that these are different words and so xa 6= xb. Let us also say, trivially, that if two pairs have (exactly) the same meaning, written form and spoken form that they are in fact the same word. Now let us consider words in which at least one form is the same, such as: shine, clean xa polish /p aa l ih sh/ from Poland xb polish /p ow l ih sh/ It is highly desirable that these be considered different words; it is only a quirk of orthography that they could be considered the same in any case; they clearly have different meanings and different sounds, and it would be a bad mistake to speak one with the spoken form of the other. We call words which have the same written form homographs, and the process of determining the word given only the written form is known as homograph disambiguation. Now consider pairs which have the same spoken form but a different written form such as: 1 meaning of course, one that would be easy for engineers to deal with.
- 81. Section 4.2. Words and Sentences 57 animal xa bear /b eh r/ naked xb bare /b eh r/ Again it is intuitive to think of these as two different words; they mean different things, are spelt differently and are used differently. Pairs such as these are termed homophones and as with the homograph example, it is only a historical quirk that these have the same spoken form, they are not related in any other way. Consider now the following: medic doctor doctor /d aa k t er/ medic doctor Doctor /d aa k t er/ medic doctor Dr /d aa k t er/ As with bear and bare these have different written forms but the same spoken form. But should we really regard them as different words? In general this is inadvisable as in doing so we wouldn’t be able to make use of the fact that apart from written form, these act identically in all other ways. The essential difference between the doctor cases and the bear and bare is in the meaning: in one case the meaning is the same while in the other it is different. Rather than count these as different words we say the different text forms are written variants of a single word, DOCTOR (we can list more such as Dr., dr. and so on). A word such as DRIVE, which shares some written variants with DOCTOR is considered a different word because it differs in meaning and spoken form: type of street drive drive,Drive,Dr,Dr. /d r ay v/ Thus words can be considered homographs if only one of their written forms is the same. A single word can also have phonetic variants, and this is found in words such as EITHER and ECONOMIC:
- 82. 58 Chapter 4. Text Segmentation and Organisation finance, money ECONOMIC economic /iy k ax n aa m ax k/, / eh k ax n aa m ax k/ choice EITHER either /iy dh er/, / ay dh er/ In general, written variants are more common that pronunciation variants because they are usually the result of creativity on the part of writing practices rather than aspects of the language itself. In traditional linguistics, pairs of words which have the same meaning but different surface forms are called synonyms so that we have HIGH and TALL or QUICK and SPEEDY. Our intuition tells us that these are indeed different words that just happen to have the same meaning. In modern linguistics, the existence of synonyms is disputed in that there is always some context where a significant difference between the words seems to arise, so that she is very tall seems to mean something different from she is very high. There are a number of words though which do seem to have exactly the same meaning in that they refer to the same entity, but which have both different written and spoken forms. Consider hercales, pronounced /h eh r k ae l iy z/ and hercules pronounced /h eh r k y ax l iy z/. These clearly relate to the same (mythical) person, but they have different surface forms. Should we consider these as the same word? To answer this, we look at the issue of the author’s intention and state that while it is hardly a disaster if one reads hercules but says /h eh r c aa l iy z/, the author clearly intended one form over the other, and we should stick with his wishes. Most examples of this type relate to different surface forms of a name (e.g. Peking and Bejing) and while they in some regards the same, they may carry subtle intended differences which may be important with respect to the author’s intentions. Hence we add to our definition of word the distinction that units in which both surface forms are different are different words, regardless of meaning. Finally we consider cases where the written and spoken form are the same, but where we can argue there is more than one underlying word. If we look at sing and sang it is clear that according to the definitions given above that although related (one is the past tense of the other) these are two different words. But what about words such as cut? This too has a present and past tense, but only one surface form. Other words have the same form but more than one part of speech: walk can be a noun or a verb, but in both cases it is spelt the same and pronounced the same. As a final case, we can consider words such as bank which, even as a noun, has two distinct meanings (bank of a river and somewhere to keep money). Pairs such as bank are
- 83. Section 4.2. Words and Sentences 59 termed homonyms, but traditionally this term covers both homographs and homophones; words which have difference in meaning and in which both surface forms are the same we shall call pure homonyms. Distinctions such as this pose the most difficultly for our definition of word. The problem is basically this; while it is clear that in general use these words are different, for text-to- speech purposes the distinction is largely unnecessary. cut is pronounced /k uh t/ regardless of its tense, and bank is again pronounced the same way regardless of meaning. It is important to realise that we should be wary of taking a position of “lets make the distinc- tion anyway, just in case”. The more distinctions we make, the bigger the feature space becomes and as problems with dimensionality and sparse data are ever present, making unnecessary dis- tinctions may harm the performance of subsequent modules. If we consider words such as SET, we find more than fifty different uses (which cover many meanings) in the Oxford English Dictionary (a chess set, set in stone, a mathematical set, a film set and so on) and we run a sever risk of unnec- essarily exploding the feature space for little gain. Because of this, we will complete our definition criteria by saying that words which have identical written and spoken form will be considered the same word, regardless of meaning. We can justify this on the grounds that collapsing this distinc- tion is acceptable for text-to-speech purposes, but it should be noted that for other purposes, say machine translation between one language and another, distinguishing between these words may be vital. But in such systems we could probably make other assumptions, for instance completely ignoring differences in spoken form in a text based machine translation system. We can summarise our definition in the following table, which describes the effect of simi- larities or differences between pairs of potential words. The table compares the meaning, written form and spoken form, gives a name to each, and shows whether we consider such pairs as the same or difference words. meaning written form spoken form name distinction different different different different different CAT, DOG different different same homophones different BEAR, BARE different same different homographs different BASS-FISH, BASS-MUSIC same different different synonyms different HIGH, TALL different same same pure homonyms same BANK same different same phonetic variants same EITHER, /iy dh er/ or /ay dh er/ same same different orthographic variants same LABOUR: labor, labor, same same same identical same 4.2.3 Scope and morphology Our discussion so far has assumed we already have our units delimited in some way. That is, we have assumed we know where one word stops and the next word starts. When first studying phonetics and speech, most people are surprised to learn that there are no gaps between the words
- 84. 60 Chapter 4. Text Segmentation and Organisation in speech; we are naturally think that language is writing, so that we think of words as well delimited units (at least for speakers of European languages). In normal speech however, the flow of speech is continuous with no acoustic boundaries marked. Furthermore, while some phonemes are more likely to occur near word beginnings and ends, there really is no way to bottom-up find word boundaries from acoustic evidence. The sceptical observer might well then ask, do words, as distinct units in time, really exist, or are they simply a invention of linguists (both traditional and modern)? These problems of scope (that is where one word stops and the next starts) can be seen as problems of morphology, the study of how words are composed. In practice, the main issue concerns whether we should explicitly regard all the inflected forms of a lexeme as words in their own right, or whether we should only consider the as a true word. A lexeme (e.g. BIG) is defined as the most basic form of a word, and its inflections are composed from this so that we have BIG, BIGGER and BIGGEST. In addition we have derivational morphology, which explains how words such as CREATION are formed from CREATE. In earlier text-to-speech systems it was often considered wasteful to list all the forms, rather the lexeme was stored and the other forms were derived. This approach of course is more expensive computationally as processing has to be used to derive the forms from the lexeme. Today it really is a matter of engineering expediency which approach is chosen as it is quite possible to store all the forms of a lexeme. 4.2.4 Contractions and Clitics We will now turn to the issue of whether longer units form words in their own right. First let us consider contractions, which are written forms created from particular combinations of two words. Common examples are: (13) I’ve, can’t, won’t ,they’ll, he’d Contractions are not abbreviations. Abbreviations are words which have a shortened written form only, such as (Dr 7→ DOCTOR). By contrast, contractions have shortened forms in both the written and spoken form, so they’ll, formed from the contraction of THEY and WILL is pronounced /dh ey l/ and not /dh ey w ih l/. We can state more properly that contractions are shorted spoken combinations of words, whose properties are reflected in writing conventions. The question with contractions is whether we should regard the contraction as one word or two. First note that the contraction isn’t always simply a case of concatenating a first written form to a shortened second written form. While they’ll is created this way, won’t meaning WILL NOT is not formed this way, and similarly with aint. While don’t can be formed from the concatenation of do and not followed by the replacement of o with ’, phonetically the process is different in that the vowel of DO, /uw/ changes to /ow/ in don’t. Furthermore, sometimes the apostrophe occurs at the place where we would break the forms (e.g. they’ll) but sometimes not e.g. don’t would break into do and n’t, not don and ’t). As with previous problems, we settle this issue by deciding what we think is best for TTS. While it is common in natural language parsing to expand contractions and regard them as two
- 85. Section 4.2. Words and Sentences 61 words, in TTS, they are spoken exactly as a single word would be, and for this reason we regard all these forms as single words. In further support of this we find that contractions can be very ambiguous with respect to the second word, for example he’d can either be HE HAD or HE WOULD. If we chose to expand he’d we would have to decide which word the ’d was a form of. Luckily in text-to-speech, both forms are written the same and are pronounced the same (/h iy d/), so we don’t have to resolve the ambiguity. Contractions come close to forming a closed class, in that they vast majority of contractions are formed from pronouns plus auxiliaries (I’ll, he’d) or with negations can’t, won’t, but it is in fact possible to form contractions from normal nouns as in: (14) The roads’ll be busy tonight (15) The boy’s got to learn (16) The team’ve really shown their strengths which means the set is not truly closed. Clitics have similar written forms to contractions in that they use a apostrophe, and that the written pattern follows through to the spoken form (unlike abbreviations). Clitics differ though in that they are not contractions of two words, but rather function more like an affix joined to a normal word. There is only one use of a clitic in English, which is the genitive ’s, as used in Paul’s book. Other languages have much richer use for clitics. In this sense, clitics can be thought of as affixes; indeed some argue that it is only a historical quirk of English printing that the generative apostrophe is used at all; there is no structural linguistic reason why we shouldn’t write Pauls book. As with contractions, we view a form with a clitic as a single word. A final type of modified word is the shortened form where a normal word has been short- ened in some way. Examples include lab (from laboratory), phone from telephone and so on. These are distinct from abbreviations in that both the written and spoken form are short- ened. This is an important distinction because we assume that when an author writes lab he does in fact mean just this and if he were to speak it aloud he would say /l ae b/ not /l ax b ao r ax t ao r ii/. Despite the commonality in meaning, we therefore regard LAB and LABORATORY as two separate words. 4.2.5 Slang forms An additional type of contraction is what we term slang forms which include forms such as gonna, wanna, gotcha and so on. Despite the lack of the apostrophe that we find with the other contractions, these are similar to contractions in that the written form clearly represents the combination of two words as a single word. The issue here is again whether to represent gonna as two words (GOING and TO) or one. At first glance, the answer to this may seem to be the same as with the other contractions we looked at, such as I’ve etc, and that they should be regarded as a single word. Recall however that our definition of a contraction was one of two words becoming a single surface form in both the written and spoken form. The problem with gonna is that it is
- 86. 62 Chapter 4. Text Segmentation and Organisation not clear exactly what the status of the spoken form is. In reality, the sequence <GOING, TO> is realised in a wide variety of fine spoken forms, and it is only an accident that two (going to and gonna) have become standard in the orthography. While gonna may be a genuine case, there are actually quite a number of less standard forms such as gotcha (GOT YOU), (whadaya WHAT DO YOU) and so on. If it wasn’t for the fact that these forms span two words, we would describe them as orthographic variants of the same word, in the way we would with nite and thru and so on. Weighing up these competing arguments and the discussion on pronunciation variation that we will have in Chapter 8, we decide that gonna and other similar forms should be represented by two words and should not be regarded as bona fide contractions. 4.2.6 Hyphenated forms So far our discussion has centred on whether forms which we normally delimit by whitespace should be considered as full words or further broken down, and we have seen that apart from a few exceptions, whitespace in conventional writing is actually a quite reliable guide as to word boundaries. But there are certain cases where we can argue that forms separated by whitespace should in fact be treated as a single word. Most of this involves issues to do with hyphens and compounding. Consider the following examples: (17) a. last ditch, last-ditch b. spokes woman, spokes-woman, spokeswoman c. web site, web-site, website d. new line, new-line, newline e. steam boat, steam-boat, steamboat f. vice chancellor, vice-chancellor g. lock nut, lock-nut, locknut Here we have a different set of words in which we can have the same two words combined to form up to three written forms; either as separate written forms separated by whitespace, as forms sepa- rated by a colon, or as a single written form. These patterns raise a number of issues. Firstly, there can be a huge number of forms such as SPOKESWOMAN, in which two relatively common words combine to form a single token. While we could again put every form of this type in the lexicon as is, there are likely to be several unknown forms. Secondly, words like spokeswoman act like separate words for the purposes of inflection; the plural of spokeswoman is spokeswomen and not spokeswomans, and so the orthographic form uses the irregular plural form of one of its constituents. Thirdly, as shown in the list of examples, the written conventions actually used vary considerably, and one finds that the web site, web-site and website forms are all common, when it is clear that they are all just written variant forms of the same word. Finally, many of these words act like a single phonological word. Of course, in all cases, there is no silence between the words in when spoken fluently, but there is also the problem of stress. In normal adjective noun patterns (e.g. large site) the stress normally goes on the second word,
- 87. Section 4.2. Words and Sentences 63 but in cases such as web site it goes on the first, as would be the case in single words such as PARASITE. This helps show the reason over the variability and behavior of the written form; in speaker’s minds this acts like a single word. There is no real way to resolve this issue as we feel we must allow words like blackbird to be represented by a single word, but it is also impossible to know where to draw the line. In summary, we use the following guidelines and rules to delimit words: • Words include free stems and all inflected and derived forms of a stem • Contractions and forms containing clitics are single words • Slang forms created from two words are not single words • Compounds, hyphenated words and multi words are considered single words on a case by case basis. • In conventional writing, whitespace and is quite a good guide to word boundaries. 4.2.7 What is a sentence? As with our discussion with words, we find that there are many possible definitions of a sentence, and we need to investigate this a little further before discussing any algorithms which process sentences. It is frequently asked whether people really “speak in sentences” at all. When listening to politicians being interviewed, they have an uncanny ability to avoid ever ending a sentence; if they did they could be interrupted. As they wish to hog the floor, they avoid ending a sentence at all costs (to the extent that even more rubbish is spoken). In spontaneous dialogue, it is again uncertain whether sentences are the main unit. Thankfully this is one area where our cross-over between written and spoken language does actually help. While it is unclear whether people engaged in interviews or dialogues use sentences, it is clear that the vast majority of conventional writing is indeed clearly written in sentences. Furthermore, when people read aloud, they effortlessly find the sentence boundaries and adopt the appropriate sentence-final prosody, such that the boundaries are preserved in the spoken signal. For our purposes, the situation is complicated somewhat by the use of non-natural language writing. In an email header, there are no sentences explicitly marked, and it is part of the verbalisation process that we find suitable sentence boundaries. Again, this problem is one we share with a human reader - when they are given an email header and told to read it, they also suffer from not knowing exactly how to go about this. Nonetheless, our task in sentence splitting is therefore to find the written sentence boundaries in conventional and all other types of writing. In the linguistic literature, many different ways of defining what a sentence is have been proposed. One approach is what we can term the discourse approach, which states something like, a sentence is the smallest possible answer to a question, or that it is the smallest unit that can bear functional meaning and so on. Another approach is to define sentences using more formal mechanisms; this can take the form of saying that a sentence is a linguistic unit with a subject- predicate structure, or even more formally, a sentence is the full expansion of any formal grammar
- 88. 64 Chapter 4. Text Segmentation and Organisation (this final point of course doesn’t attempt to define sentences “in nature” but just says that they are a logical by product of a formal grammar). A third definition based on writing, which would define a sentence as something like “the words between two full stops” or, “it starts with a capital letter and ends with a full stop”. With certain caveats (see below) we will adopt this definition as we can’t rely on using any deeper linguistic analysis to assess the internal coherence of a sentence. 4.2.8 The lexicon In all modern TTS systems we make extensive use of a lexicon. The issue of lexicons is in fact quite complicated, and we discuss this in full in Chapter 8 when we start to talk about pronunci- ation in detail. For our current purpose, its main use is that it lists the words that are known to the system, and that it defines their written form. A word may have more than one written form (labour and labor) and two words may share the same written form polish etc. It is by using a lexicon as a place to define what is a possible word and what its written forms may be, that we can use the decoding model we are adopting in this book. 4.3 TEXT SEGMENTATION Now that we know what we are looking for (underlying words) we can turn our attention to the problem of how to extract these from text. While in principle this could be achieved in a single process, it is common in TTS to perform this in a number of steps. In this section and the next we deal with the initial steps of tokenisation and sentence splitting which aim to split the input se- quence of characters into units which are more easily processed by other processes which attempt to determine the word identity. subsequent 4.3.1 Tokenisation Formally, we can think of the input as a sequence of characters which encode a sequence of words: < w1,w2,w3,...wN >7→< c1,c2,c3,...cM > and so our problem in text analysis is simply one of reversing this process so that we can uncover the word sequence from the character sequence, as in the following: (18) The old lady pulled her spectacles down →< THE, OLD, LADY, PULLED, HER, SPECTACLES, DOWN > In most cases of conventional writing, the words don’t overlap with respect to the characters, so we can also represent this relationship as: (19) THE z}|{ The OLD z}|{ old LADY z }| { lady PULLED z }| { pulled HER z}|{ her SPECTACLES z }| { spectacles DOWN z }| { down
- 89. Section 4.3. Text Segmentation 65 We can clearly see that the written sentence is easily segmented by using whitespace as a delimiter. This is only possible in this because the incidence of whitespace in the input always indicates the presence of a word boundary. There are numerous cases however where this simplistic approach fails. In a sentence such as: (20) In 1997, IBM opened a Polish lab problems occur because a section of the text may be ambiguous as to its word: (21) polish → POLISH COUNTRY or POLISH SHINE problems where whitespace is insufficient to act as the only guide to segmentation: (22) IBM →< LETTER I, LETTER B, LETTER > cases where the token does not map to an word: (23) , → NULL and cases where both ambiguity and segmentation problems exist: (24) a. 1997 →< NINETEEN, NINETY, SEVEN > b. or < ONE, THOUSAND, NINE, HUNDRED, AND, NINETY, SEVEN > c. or < ONE, NINE, NINE, SEVEN > so we see in this last example that the process of converting sentences into words is one of both segmentation and classification of the written form of the sentence. As the classification and seg- mentation interact, it is not possible to do one then the other. Rather, our approach is to first perform a provisional segmentation into potential written forms called tokens, followed by a sec- ond step whereby we examine each token in turn and resolve any ambiguity. This first process is called tokenisation; the step which then generates the words from the tokens is called text analysis and is the subject of the next chapter So in our first example, the tokenisation process would perform the following mapping: (25) The old lady pulled her spectacles down → < the, old, lady, pulled, her, spectacles, down > The advantage of dividing the problem this way is that it makes it much easier to perform the text analysis step. By tokenising (segmenting) the text, we are making it possible for the text analysis algorithms to focus on one token at a time: while the analysis algorithms may examine the tokens on either side of the one in question, only a few are usually consulted, and this makes the writing of rules and the training of algorithms quite a bit easier. 4.3.2 Tokenisation and Punctuation In languages like English, it is obvious that the presence of whitespace in the input is a very good guide to tokenisation. In most conventional writing, splitting on whitespace will often leave a
- 90. 66 Chapter 4. Text Segmentation and Organisation one-to-one mapping between tokens and words. The difficulties soon arise however: when we encounter sequences of capital characters such as IBM we could guess that this will be a letter sequence and therefore go someway towards the classification and rewrite process by splitting this into three tokens. But there are several forms such as NATO and UNICEF which don’t follow this convention; furthermore, sometimes normal words are put into upper case for effect: lets GO!. Dealing with punctuation properly is one of the main issues in tokenisation. In sentences such as (26) So, he went all that way (at considerable cost) and found no-one there we see the classic uses of punctuation. Firstly it creates a boundary between two tokens which might otherwise be assumed to be linked, or it creates a link between two tokens which might otherwise be assumed to be separate. So a full stop ends a sentence, a comma ends a clause and both create a boundary between the tokens that precede and follow them. Within this use, we have status markers, so that while ., ? and ! all indicate the end of a sentence, they each assign a different status to that sentence. Hyphens and apostrophes link tokens that might otherwise be assumed to be more separate (assuming that whitespace would be used if there were no hyphen or apostrophe). More difficult problems arise with technical language. First the rules of using whitespace in conjunction with punctuation are often different; so that in a time we find 10:34am in which there is no whitespace after the colon. Furthermore, some of the verbalisations of technical language speak the punctuation as words, so that the words for yahoo.com are <YAHOO, DOT, COM>, where the . is spoken as DOT. It helps in cases like these to make some distinctions. First we need to distinguish underlying punctuation from text punctuation. Underlying punctuation is the abstract notion of what we normally think of as punctuation in conventional writing, and text punctuation is its written form. These are clearly the punctuation equivalent of words and tokens. So, while we may have two types of underlying punctuation DASH and HYPHEN there may in fact be only one written form in our character set. This use of one character for two underlying punctuation marks causes ambiguity and this must be resolved. In addition, as we have just mentioned, punctuation is sometimes spoken in technical language. We deal with this phenomenon by using our model of verbalisation, so that in yahoo.com the sequence is analysed firstly as an email address, then verbalised into natural language and then spoken. In such cases, we will represent the word that the punctuation is spoken as WORD COMMA, WORD COLON and so on, to distinguish from the silent form of COMMA and COLON. 4.3.3 Tokenisation Algorithms We will now describe a basic tokenisation algorithm: 1. Create an empty list for the tokens to be placed in. 2. Define a start point before the first character of the sentence.
- 91. Section 4.3. Text Segmentation 67 3. Move forwards through the sentence, examine each character and the character following it, and decide whether there should be a token boundary at this point 4. If yes, append an entry to the token list and copy (without modification) the characters from the start point to the boundary point. 5. Reset the start point to be at the boundary point 6. If at the end of the sentence stop, otherwise go to step 3. The heart of the algorithm is therefore the part which decides whether a boundary should be placed between two characters of the input. This works on the following criteria: 1. Split on whitespace, but don’t include the whitespace itself in any token. 2. Split on punctuation, create a separate token for each punctuation character. 3. Split when a contiguous non-whitespace sequence changes from one character grouping to another (e.g. 10cm → 10 cm). 4. Don’t split in any other cases. Sometimes more complex (or more accurate) tokenisation algorithms are used. The amount of resource put into tokenisation to some extent depends on the sophistication of the process that follows it. If this can handle errors or noise in the tokenisation, then a fairly simple algorithm should suffice. If however the subsequent algorithms require more accurate tokenisation a number of approaches can be applied. The deterministic approach can be extended by considering greater numbers of cases and writing rules by hand which deal with each case in turn (e.g. email addresses, times, dates and so on). Alternatively, a data-driven approach can be taken where we label a corpus with token boundaries and train an algorithm to learn the correspondence. Such an approach is a classic use of a data-driven technique in speech and language processing, and discussion of this will be deferred until the general discussion of this in Section 5.2.1. 4.3.4 Sentence Splitting Many of the algorithms in TTS work a sentence at a time. This is because most linguistic units smaller than this (words, syllables) etc are heavily influenced by their neighbours which makes autonomous processing difficult. Sentences on the other hand don’t interact with each other much, and apart from some specific phenomena, we can by and large process each sentence indepen- dently without problem. The input to the TTS system is not necessarily in sentence form however, and in many cases we are presented with a document which contains several sentences. The task of sentence splitting then is to take the raw document and segment it into a list of sentences. While sentence splitting is not the most complex of tasks in TTS, it is important to get right and this is mainly due to the fact that sentence-final prosody is one of the phenomena that listeners are most sensitive to. Generating the high quality sentence-final prosody is hard enough in its own right, but if the sentence boundary is in the wrong place, then the system has no chance. For conventional writing we make use of the fact that in most cases the writer has clearly marked the sentence boundaries; and so our job is simply to recover these from the text. Perhaps
- 92. 68 Chapter 4. Text Segmentation and Organisation ironically then, our algorithm is based on the lay notion of finding upper case characters at begin- nings of sentences and period characters at the end, and marking sentences as what lies between. The situation is not quite that simple; as we know the period character can be used for a variety of purposes, as can upper case characters. The task is therefore one of finding instances periods and related characters, and classifying them as to whether they indicate sentence boundaries or not. A basic sentence splitting algorithm for conventional writing can be defined as follows: 1. Start a new sentence at the beginning of the input 2. Search forwards through the input to find instances of the possible end of sentence characters ., ! and ?. 3. Using the immediate context of these characters, determine whether a sentence boundary exists; if so complete the current sentence and start a new one. 4. End the current sentence at the end of the document (regardless of whether the last characters are ., ! or ?.) The heart of the algorithm is the decision as to whether an instance of an text form of ., ! or ? constitutes a sentence of not. The most difficult case is . as this can be used as part of an abbreviation, a number an email address and so on. In conventional writing, a space always follows the full stops at a sentence end, and the next sentence always starts with an upper case letter. In fact the only conventional other use of . is in abbreviations, so that we might find: (27) Dr. Jones lives in St. James Road In fact, in most modern writing, the use of . in this sort of abbreviation is not common. Profes- sional publishers make use of style manuals to help authors with various issues in writing and layout. Conventions differ, but instructions from one academic journal include: • Do not use periods in titles, hence Dr Jones, Mr Mark Johnson Jr • Do not use periods in letter sequences, hence USA, IBM • Do use periods for abbreviations in personal names, hence J. R. R. Tolkein (note that a space follows the period) • Do use periods for latin letter sequences, hence e.g. and et. al. These guidelines show that the incidence of periods in abbreviations is probably less com- mon that expected, but there are still potentially confusing situations. Remarkably, many style guides are aware of this and even advise writers “do not start a sentence with an abbreviation”. Of course, real life is never so simple, and even in conventional writing, one sees frequent variations on the above guidelines, such that we must in principle allow for a period after any abbreviation (For example, one sees JRR Tolkein, J.R.R. Tolkein, J R R Tolkein and J. R. R. Tolkein all on writing from the Tolkein Society alone). Usage patterns are potentially sig- nificant, in that while we may allow a period after any abbreviation, a probabilistic algorithm could make use of the fact that they are rarely found in some cases. And of course, if these really were universal conventions they wouldn’t have to be spelled out in a style guide, authors would simply know.
- 93. Section 4.4. Processing Documents 69 For less casual conventional writing, the situation can be more difficult. In email messages, one finds that some writers do not use an upper cases characters at the start of a sentence, or that new-lines are used as sentence boundaries and no ending punctuation is found. We also come across the problem in email that the form of the headers is intended for email programs to parse and process and not for reading aloud. So the notion of “sentence” is unclear here. To get around this, we simply have to decide in what form we will read this information and so what form the sentences will take in these genres. One approach is to have each line as a sentence, another is to somehow process the header into a spoken list or other structure. 4.4 PROCESSING DOCUMENTS 4.4.1 Markup Languages It is useful to be able to control the behavior of the TTS system at run time. While global parame- ters can be set in a number of ways, often it is useful to have very precise instructions which affect specified words. This is most effectively done by interleaving instructions to the TTS system with the actual text to be spoken. In older TTS systems, it was common for each to have its own type of markup. So for example in the Bell Labs system [411],, there are special sequences of escape characters that help the user specify what behaviour the system is to follow. The disadvantage of this is that these special characters would mean nothing to a different TTS system, so a document marked up for one TTS system would not work with another. To get around this, Taylor and Isard [435], [444] proposed the speech synthesis markup language (SSML) which aimed to provide a general set of markup tags which could in principle be used by any TTS system. This was extended by a number of further projects including the spoken text markup language by Sproat et al [411], and the java speech markup language developed by Andrew Hunt at Sun Microsystems. These served as the basis of the synthesis part of the W3C standard VoiceXML (also called SSML, but not to be confused with the original). All these schemes are based on the idea of generalised markup, first developed in the type- setting industry. There, the idea of markup had been around for a long time, where a typesetter would mark some text to be printed in 12 point, or in italics or so on. The problem with such markup is that instructions is that they are physical, meaning that they are very specific to the type of printing being used. If someone wished to print the document in a bigger font for instance, the insistence on using 12 point (which may have seemed large in a 10pt document) means that the intended effect is not achieved. Generalised markup attempts to get round this by splitting the type setting into a two stage process. The author, or editor, adds logical markup to the text, for example by indicating that something is a heading, or requires emphasis. In the second stage, this is then rendered by a process much more specific to the actual printing mechanism being employed. As the document itself only contains logical instructions, it is much more portable, and can be printed on a number of devices and in a number of styles.
- 94. 70 Chapter 4. Text Segmentation and Organisation The speech synthesis markup languages followed a similar philosophy, so instead of spec- ifying that a word should be spoken at 150Hz (which would have different effects for different speakers), instead more abstract or “logical” instructions are used, which for instance state that a particular word is to be emphasised. A few simple examples will now be explored. tag Example Explanation <s> This tag can be used <s> to override the decisions of the <s> sentence splitting algorithm. This tag indicates that a sentence break should be placed at this point. It is a good way for the author to over ride any possible shortcomings in the sen- tence splitting algorithm. <emphasis> Sometimes we <emphasis>need</emphasis> to give a word special emphasis. This is used to give particular empha- sis to a word or group of words. Note how abstract or “logical” this tag is; no reference is made to the phonetics of how to render the emphasis as height- ened pitch, longer duration and so on. <voice> <voice gender="female" age="20"> This is a female voice </voice> <voice name="mike"> And this is a male one <break/> Hope you’re OK. </voice> Many TTS systems have the capabil- ity of speaking in more than one voice. This tag can switch between voices in by specifying general attributes or spe- cific voice names: In an SSML document, the tags are interleaved with the text to be spoken. It is of course important that the two are distinguished, and so the first stage in processing an SSML document is to perform an XML parsing step. There are a number of standard parsers which do this and which subsequently produce a structure in which the text and tags are clearly separated. From there, it is up to the TTS system itself to decide how to interpret each tag. The tags vary considerably in their purpose, and hence they require processing by different modules in the TTS system. The standard way of dealing with this is to do a partial processing of the XML initially, and then leave hooks in the tokens which tell subsequent modules what to do. As can be seen with the <s> tag, XML, tokenisation and sentence splitting can interfere, and so it is often a good idea to perform all the required processing in a single step. 4.4.2 Interpreting characters Since computers have been around they have been used to store text, but as computers only operate in ones and zeros, text can not be represented directly. Rather a character encoding scheme is used which maps internal computer numbers to characters. So long as the same convention
- 95. Section 4.5. Text-to-Speech Architectures 71 is maintained by the readers and writers, chunks of memory can be used to represent encoded text. The issue for a TTS system is to identify the character encoding being used and process it appropriately. Partly due to the sheer diversity of the worlds writing systems and partly due to historical issues in the development of character encoding schemes, there are several ways in which characters can be encoded. One of the earliest and most common encodings was ASCII (the American Standard Code for Information Interchange), which gave a character value to the first 127 values in an 8-bit byte of memory (the final bit was left unused). Ascii is of course a standard that was developed in the United States, which could only represent the 26 characters of the standard English alphabet. Most of the worlds languages of course require different characters and so ascii alone will not suffice to encode these. A series of extensions were developed primarily with use with other European characters, often by making use of the undefined eighth bit in ascii. The ISO 8859 standard defines a number of encodings. For example ISO 8859-1 encodes enough extra characters for most western European languages such as French and German. ISO 8859-2 encodes eastern European languages, while ISO 8859-7 encodes Greek characters. These are all 8-bit encodings are can be used without trouble so long as it is known which is in use. All contain ascii as a subset so have the advantage of backwards compatibility. Many languages, in particular south east Asian languages have far more characters than can be included in 8 bits, and so cannot be accommodated in single byte encoding schemes. A number of incompatible wide character encodings were developed for these languages before a universal standard unicode was developed in the early 1990s. This provides an encoding for virtually all languages, although some controvosies (largely political rather than technical) still prevent its completely universal adoption. Unicode simply defines a number for each character, it is not an encoding scheme in itself. For this, a number of different schemes have been proposed. UTF-8 is popular on Unix machines and on the internet. Is is a variably width encoding, meaning that normal ascii remains unchanged but that wider character formats are used when necessary. By contrast UTF-16 popular in Mi- crosoft products, uses a fixed size 2 byte format. More recent extensions to unicode mean that the original 16 bit limitation has been surpassed, but this in itself is not a problem (specifically for encodings such as UTF-8 which are extensible). 4.5 TEXT-TO-SPEECH ARCHITECTURES The various algorithms described here all produce output, and this has to be passed to the next module in the TTS system. As the next module will produce output too, it makes sense to have a general mechanism for storing and passing information in the system, This brings us to the issue of how to design our text-to-speech architecture. Most TTS systems have adopted a solution whereby a single data structure is passed between each module. Usually, this data structure represents a single sentence, so the TTS system works by first splitting the input into sentences by the sentence splitter, forming a data structure containing
- 96. 72 Chapter 4. Text Segmentation and Organisation the raw text for each sentence, and then passing this data structure through a series of modules until a speech waveform is generated. We will term this the utterance structure. In many early systems, the utterance structure was no more than a string that was passed between modules. In doing this systems could adopt either an overwrite paradigm or a addition paradigm. For instance in an overwrite system if the input text was The man lived in Oak St. The text normalisation module might produce output like: the man lived in oak street And the subsequent pronunciation module might then produce output like: dh ax m a n l i v ax d ax n ou k s t r ii t The point is, in each case, the module takes its input and only outputs the particular results from that module. In an addition paradigm, each module adds its output to the output from the previous modules. So in the above example, the output from the text normalisation module might produce something like: The|the man|man lived|lived in|in Oak|oak St|street .|null And the subsequent pronunciation module might then produce output like: The|the|/dh ax/ man|man|/m ae n/ lived|lived|/l ih v ax d/ in|in|/ax n/ Oak|oak|/ow k/ St|street|/s t r iy t/ .|null The advantage of addition systems is that it is often hard to know in advance of development what each module requires as input, and if we delete information we run the risk of supplying a module with impoverished input. For example, we might find that the phrasing module (which is called after the text normalisation module) requires punctuation, which can clearly not be used if the information has been deleted. The trouble with string based addition systems is that they can become unwieldy rather quickly. Even in the example above, with a short sentence and the operation of only two modules, we see that the string is rather difficult to read. Furthermore, each module will have to parse this string, which becomes increasingly more complicated as the string becomes richer. It is common therefore for modern systems to adopt an approach which follows the addition paradigm, but does so in a way that allows easier and clearer access to the information. Two of the most widely used formalisms are the heterogeneous relation graph or HRG formalism [441],
- 97. Section 4.5. Text-to-Speech Architectures 73 used in the Festival system, and the delta system [201], used in a number of commercial and academic systems. Both these are based on the idea of building data structures based on linguistic items2 An item can be any single linguistic unit, including a word, phone, syllable, pitch accent or other. The item is a feature structure which is a data structure common in computational linguistics which defines a collection of properties as an unordered list of key-value pairs also known as attribute-value pairs. Attributes must be atomic and drawn from a finite set of defined values. Attributes must also be unique within the feature structure, so that when given an attribute, at most one value will be found. From a computer science perspective, feature structures are very similar to associative arrays or maps, and can also be though of as lookup tables or dictionaries. In mathematics, they can be thought of as finite partial functions; that is, functions which have a finite domain, and for which not all domain values are defined. An example feature structure is given below: word : NAME abuse1 POS noun TEXT abuse PRON /@buws/ (4.1) In the above example, all the values are atomic. It is also possible to have feature structures themselves as values, which is useful for grouping particular types of information together: phone : NAME p STRESS 1 DISTINCTIVE FEATURES VOICED false MANNER stop PLACE bilabial (4.2) (4.3) These items represent a single linguistic entity. As there are usually several of these in each sentence we need an additional data structure to store these. In the HRG formalism, relations are used to group items of a similar type. In the standard HRG formalism, we have three types of relation: list, tree and ladder. Many types of linguistic data are best stored in lists, and this is how word, syllable and phone information is normally stored. Other types of linguistic data are most 2 For clarity, we will adopt HRG terminology throughout our explanation.
- 98. 74 Chapter 4. Text Segmentation and Organisation (c) (b) (a) Figure 4.1 The three types of relation: a) list relation, b) tree relation and c) ladder relation S VP NP (a) (b) this is parent daughter1 daughter2 parent daughter2 an example daughter1 this is next previous next next previous previous an example daughter2 Figure 4.2 a) list relation containing word items. b) syntax relation containing syntax items . suited to trees, and this is how syntax, prosodic phrasing and syllable structure can be stored. Less common is the use of ladders, which are used to represent connections between linguistic data that is not organised hierarchically, as in a tree. The ladder structure can be used to implement the data structures of autosegmental phonology [179], which is used for instance to represent the relationship between syllables and intonation accents. The three types of structure are shown in Figure 4.1. One way of configuring the data structures is to have a separate relation for the output of each
- 99. Section 4.5. Text-to-Speech Architectures 75 3 2 1 6 9 5 8 7 4 this is Syntax Word an example NP VP S Figure 4.3 Example of a word and syntax relation in which the items in the word relation are also in the syntax relation. module, and so overwriting the output of previous modules is avoided. While straightforward, this configuration can makes the utterance structure very dependent on the particular TTS setup, so that a module that performs pronunciation in one step would generate one relation, whereas one that performs lexical lookup, letter-to-sound rules and post-lexical processing as three separate mod- ules would generate three relations. An alternative is to separate relations from particular modules and instead use relations to store information for each linguistic type, so we can have a “phone” re- lation which comprises a list of phone items, a syllable relation, word relation, intonation relation and so on. The relations are graphs of nodes arranged in the form of lists, trees and ladders. Each node can be linked to an item, and in this way the relation can show the contents of the items and the relationship between them. Named links such as next(), previous(), daughter() and parent() and so on are used to navigate through the relation. Examples of a list and tree relation are shown in Figure 4.2. One of the key features of the HRG architecture is that items can be in more than one relation at once. For example, word items can be be in both the word relation and the syntax relation. This is useful because often we wish to process the words left to right as a list: the word relation is used for this. Sometimes, however, we want to access the syntax relation, which is a tree in which the words form the terminal nodes. Having the same information in the two different places is unappealing as an update to one item requires an update to the other, and this is often awkward to achieve. Figure 4.3 shows the syntax and word stream together. This ability to include items in multiple relations means that usually any information in the utterance structure can be accessed
- 100. 76 Chapter 4. Text Segmentation and Organisation s word phoneme structure 1 syllable structure 2 syllable syllable nucleus rhyme coda onset ch r @ coda k u r t onset rhyme nucleus syl structure syl Figure 4.4 Thing from any other item by navigating through the relations. An alternative to the HRG formulation is the Delta formulation [201]. This mainly differs in that the connection between the relations (called streams in Delta) is determined by a chart. An example of a delta structure is shown in Figure 4.4. Here, each stream/relation is a list, but navigation is performed by going to the edge of the item and ascending or descending through the chart. The delta system is somewhat simpler and often easier to use than the HRG formalism, but is somewhat less powerful in that it cannot represent general tree information. 4.6 DISCUSSION 4.6.1 Further Reading The ontological status of linguistic entities was a prime concern of linguistics in the middle part of the 20th century, but surprisingly has not been the subject of more recent work. Good references that discuss the nature of words, sentences, phonemes and so on include Lyons [290] and Bloom- field [57]. Tokenisation in general does not receive much attention in the literature as it is always assumed to be a relatively minor (and simple) task. The original SGML book is Goldfarb [178], which describes in detail the justification for markup languages based on logical rather than physical markup. XML is by is inherently tied in with the development of the web, and so good references for it can be found on the W3C website (www.w3c.org). The original SSML and justification for its use in TTS can be found in Taylor and Isard [444]. As the current version of SSML was developed as a web standard and references to this can again be found on the W3C website.
- 101. Section 4.6. Discussion 77 4.6.2 Summary Words and sentences • It is important to distinguish words which are unique linguistic entities from tokens which are their textural instantiation. • The chapter attempts to provide a rigourous definition of the notion of word so that the process of finding these from the text has a clear target. • Homographs are sets of words which share the same text form. There are three reasons why homographs arise: 1. The same lexeme being used as a noun or verb which have different pronunciation (e.g. project) 2. From abbreviated forms of two words producing the same written form (e.g. DOCTOR and DRIVE producing Dr). 3. By historical accident (e.g. BASS-MUSIC and BASS-FISH). • We use a series of tests to define our set of words by looking at distinctions in written form, spoken form and meaning. • Sentences have a number of possible valid definitions. We use one based on the writing as this is most helpful to TTS. Text Segmentation • Top make subsequent processing easier we require that the input be segmented into sentences and tokens. • It is hard to divide the input into sequences of characters which align one-to-one with the words, hence tokenisation often produces a provisional segmentation of the text. • Deterministic algorithms are often accurate enough to produce acceptable tokenisation and sentences. Documents • A number of markup languages exist for speech synthesis, based on XML the successor to XML. • These markup languages can be used to give extra instructions to the TTS system to override default behaviour. • In modern systems, this markup takes the form of logical rather than physical instructions. Text-to-speech Architectures • All TTS systems are composed of a series of modules and all systems have to pass data from one module to the next. • Early systems typically used strings for this purposes
- 102. 78 Chapter 4. Text Segmentation and Organisation • More modern systems use sophisticated structures where each type of linguistic information (e.g. words, syllables, phonemes) can be accessed separately. • The HRG system and delta system are two popular systems for this purpose.
- 103. 5 TEXT DECODING: FINDING THE WORDS FROM THE TEXT 5.1 OVERVIEW OF TEXT DECODING The task of text decoding is to take a tokenised sentence and determine the best sequence of words. In many situations this is a classical disambiguation problem: there is one, and only one correct sequence of words which gave rise to the text, and it is our job to determine this. In other situations, especially where we are dealing with non-natural language text such as numbers and dates and so on, there may be a few different acceptable word sequences. So in general, text decoding in TTS is a process of resolving ambiguity. The ambiguity arises because two or more underlying forms share the same surface form, and given the surface form (i.e. the writing), we need to find which of the underlying forms is the correct one. There are many types of linguistic ambiguity, including, word identity, grammatical and semantic, but in TTS we need only concentrate on the type of ambiguity which affects the actual sound produced. So while there are two words which share the orthographic form bank, they both sound the same and so we can ignore this type of ambiguity for TTS purposes. Tokens such as record can be pronounced in two different ways, and so this is the type of ambiguity we need to resolve. ’In this chapter, we concentrate on resolving ambiguity relating to the verbal component of language. At its core, this means identifying the correct sequence of words for every sentence, but we shall also consider in brief the issue of resolving ambiguity in the syntactic structure of a sentence also, as this is thought by some to have a lesser but still significant impact on speech. In the next chapter, we concentrate on the corresponding process for prosody, where we attempt to find a suitable underlying prosodic representation from the text. This process boils down to the following tasks: 1. Determine the semiotic class (number, date, natural language etc) for each token. 2. For each sequence of tokens in the same semiotic class find resolve any ambiguity and find their underlying forms 3. For non-natural language classes, perform a verbalisation process, which converts their un- derlying forms into a sequence of words which can be spoken. Generally, non-natural language tokens are quite easy to decode; the difficulty is not in 79
- 104. 80 Chapter 5. Text Decoding: Finding the words from the text knowing how to decode the token 1997 once we know it is a year, but rather in determining it is a year in the first place. With natural language tokens, even once we have established that they are indeed natural language, we still can face considerable difficulty in determining their underlying form as different words can share the same text form. While there are many subtle issues involved in text decoding, in essence then it reduces to two main problems; semiotic classification, i.e. the assignment of a semiotic class to each token, and homograph disambiguation where we determine the correct word for ambiguous natural language tokens. Both these tasks involve assigning a category to a token, and as we shall see, both can be solved with very similar algorithms. Because of this, we now turn our attention to a general overview of text disambiguation. 5.2 TEXT CLASSIFICATION ALGORITHMS 5.2.1 Features and algorithms In both semiotic classification and homograph disambiguation, our goal is to assign a label to each token, where the label is drawn from a pre-defined set. This process is one of classification, which as we shall see, crops up all the time in TTS to the extent that we use a number of basic approaches again and again for various TTS problems. More formally, we say that for each unit (tokens in this case) we wish to assign a label li drawn from a set of labels L = {l1,l2,...lN}. Our decision is based on analysing a number of features each of which has set of legal values. These serve as information as to which label is best in each case. This process can be defined as a function C C : F 7→ L which maps from a feature space F to a label space L. Let us consider a simple example where we wish to disambiguate the token read which can encode the past tense word READ-PAST, pronounced /r eh d/ or the present tense/infinitive word READ-PRESENT, pronounced /r iy d/. In this case, our label set comprises two labels such that L = {l1 = READ-PAST,l2 = READ-PRESENT}. If we now consider some example sentences: (28) a. I was going to read it tomorrow. b. I read it yesterday. c. It is vital that young people learn to read at the correct stage. d. It read him the riot act. We see that the other words/tokens in the sentence, and the way they interact with the to- ken read gives us strong clues as to which word is correct in each case. These “clues” can be formulated as features, so in the above examples we might have features and values defined as:
- 105. Section 5.2. Text Classification Algorithms 81 Name Feature legal values F1 Time token TOMORROW, NOW, YESTERDAY F2 Is the preceding word TO ? yes, no F3 Is the sentence part of a narrative yes, no Hence a classification algorithm working on this is defined as C(F1,F2,F3) 7→ l1 or l2 Below we describe a range of algorithms which operate in this way. Some are explicitly formulated by a human designer and by convention we call such algorithms hand written (although obviously it is the brain which really writes the rules). In these approaches, a human designer has considered the effect of the features and has by one means or another written a system which performs the above mapping. The other common approach is data driven where the mapping is learned by analysis of a database of examples. Often this database is labelled, meaning that it comprises a set of pairs of feature values and labels, with the idea that an algorithm can be trained to reproduce this correspondence. Not all data driven algorithms work in this way, some are unsupervised meaning that they look at the inherent patterns in the data and attempt to come up with a mapping that way. An important sub-class of data-driven algorithms are statistical or probabilistic. Regardless of which approach is taken, the goal is effectively the same, and all encounter the same problems to be overcome. The difference between hand written and data driven is not as large a difference as it is often portrayed in that even if a human designs the rules, they presumably do this based on experience of the problem, in which they have encountered data with these features and labels. Conversely, using an arbitrary data driven algorithm on data generally produces poor results; better results are obtained by pre-processing the features and constraining the model in some way, which of course requires knowledge of the problem. The basic problem in all classification algorithms, no matter what type, comes from the issue of how features interact. In general, the features can not be considered independent, such that we count the evidence from one feature, then the next and so on and then combine these to make a final decision. If we could do this, then the design of classification algorithms would be a simple matter. Rather the features tend to interact, so in the above example, instead of examining a separate situation for each feature, we have to consider all combinations of feature values. If we had say three values for the first feature (TOMORROW, NOW, YESTERDAY)), and yes/no values for the second and third, this would give use a total of 3×2×2 = 12 possible feature combinations. The fundamental issue in classifier design is how to deal with this feature combination issue. On the one hand, features hardly ever operate independently, so the effect of particular combina- tions needs to be taken into account. On the other, it is generally impossible to deal with ever feature combination uniquely as there are simply far too many. If one uses hand written rules there are too many cases to deal with, if one uses a statistical algorithm the amount of data required is too large. The worst case is where the feature combinations are completely unpredictable, in which case the only answer is a lookup table which lists every combination of feature values and
- 106. 82 Chapter 5. Text Decoding: Finding the words from the text specifies a label: F1 F2 F3 Label YESTERDAY yes yes READ-PRESENT YESTERDAY yes no READ-PRESENT YESTERDAY no yes READ-PAST YESTERDAY no no READ-PAST NOW yes yes READ-PRESENT NOW yes no READ-PRESENT NOW no yes READ-PAST NOW no no READ-PRESENT TOMORROW yes yes READ-PRESENT TOMORROW yes no READ-PRESENT TOMORROW no yes READ-PAST TOMORROW no no READ-PRESENT Our above example is quite small in fact, more realistically we might have 10 or 20 features that we used and in such cases the number of feature combinations might be in the thousands or even millions. In general of course, the more features we use, the more accurately we should be able to make our prediction. But as we add more features, the number of feature combinations increases and we become less likely to observe a full set. This problem is known in statistics as the curse of dimensionality and gives rise to the phenomenon known as data sparsity meaning the inability to obtain sufficient examples of each feature combination. While using fewer features will result in less data sparsity, it also runs the risk that the classifier will be too crude and miss out on important distinctions. Thankfully, the pathological case where every feature combination operates differently is very rare; and in general while features rarely operate independently, they do have well behaved effects, such that we can build models, or make assumptions about feature interactions that allow us to avoid the lookup table case. As we shall see, all the classification approaches below work by using some knowledge or assumptions about how the features interact in order to reduce the number of parameters than need to be specified. One common approach is to transform the original features into a new feature space in which the new features do in fact operate independently. Another approach is to determine which features interact most, model the dependencies between these and then assume that the other features do in fact operate independently. In our above example, we may decide for instance that while F1 and F3 often interact, F2 operates independently, which means we can reduce the size of the lookup table to six entries.
- 107. Section 5.2. Text Classification Algorithms 83 F1 F3 Label YESTERDAY yes READ-PAST YESTERDAY no READ-PRESENT NOW yes READ-PRESENT NOW no READ-PRESENT TOMORROW yes READ-PAST TOMORROW no READ-PRESENT which can then just be combined with the value for F2. In general, hand written rules can be quite successful when the number of features is low and the training data is small. A knowledgeable designer can quite often use insight to design a classifier that gives reasonable performance. This becomes increasingly difficult as the number of features increases and the interactions between these get more complicated. Another way of looking at this is to say as we acquire more training data, it becomes possible to use more features and possible to model the interactions between these more easily. The complexity of this soon exceeds human capabilities, and so a data-driven algorithm will tend to deliver better performance. 5.2.2 Tagging and word sense disambiguation Semiotic classification and homograph disambiguation are close enough in terms of formalism that by and large the same techniques can be used for either problem. We find that for these, and related problems, we have two main types of technique. The first is known as tagging and is the generalisation of the subject of part-of-speech (POS) tagging. In POS tagging, the goal is to assign a part-of-speech tag c, drawn from a pre-defined set C to each token t. Hence for a sentence: (29) The man took a walk in the park We assign a POS tag to each word as follows (30) The/det man/noun took/verb a/det walk/noun in/prep the/det park/shelf Where the ambiguous token walk, which could be a noun or a verb, has been assigned a noun label. In general then, a tagger assigns a label to every token, and these are drawn from a single set C. The second type of algorithm performs word sense disambiguation (WSD) . The difference between this and tagging is that in word sense disambiguation, we define multiple sets of labels L1,L2,...,LM and choose a label from one of these sets for only some of the tokens. In natural language processing, WSD is often used for semantic disambiguation to resolve cases such as: (31) I took the money to the bank where we are only really concerned with the disambiguation of the token bank. This has 2 senses BANK MONEY and BANK RIVER and these are particular to that token only (i.e. these labels are meaningless to the token took). In TTS we are of course only concerned with homonyms if they affect pronunciation, so that instead we would look at examples such as:
- 108. 84 Chapter 5. Text Decoding: Finding the words from the text (32) He played bass guitar with the aim of resolving it to (33) He played bass/BASS-MUSIC guitar Similar techniques can be used for both tagging and WSD; the reason these are considered two different problems is that one is mainly concerned with syntactic ambiguity, the other with semantic ambiguity. In the following sections, we example a number of algorithms. In these we use homograph disambiguation for all the examples; but this is purely for reasons of exposi- tion. In general, all these techniques can be used for semiotic classification, and other types of classification described in Chapter 6. 5.2.3 Ad-hoc approaches While we like to imagine that beautiful well constructed theories underly our algorithms, we frequently find that when it comes to text classification many systems in practice simply use a hodge-podge of rules to perform the task. This is particularly common in approaches to semiotic classification in genres where the amount of non-natural language text is very low, and so only a few special cases (say dealing with numbers) are required. As these systems are by their very definition ad-hoc we can’t give a comprehensive guide to how these work, but often these are configured as a series of regular expressions which try to find a pattern (e.g. if the number is 4 digits long and has the form 19xx or 20xx is is a year). These rules are often ordered somewhat arbitrarily, in many cases with the most recently written rules last. Sometimes the output is directly rewritten over the input and sometimes systems keeps running in a repeated fashion until no rules fire. For text genres dominated by natural language text, and in situations where very little training data is available, this approach may be adequate for semiotic classification. This is less likely to work for homographs, but again depending on the language, the level of errors may be acceptable. 5.2.4 Deterministic rule approaches A more principled approach is to define a general rule format, a rule processor or engine and a clearly defined set of rules. This at least has the advantage of engine/rule separation meaning that the way the rules are processed and the rules themselves are separate, which helps with system maintenance, optimisation, modularity and expansion to other languages. One common type of rule is the context-sensitive rewrite rule, which takes the form A → B/d/C meaning token A is converted (“rewrites to”) d when preceded by token(s) B and followed by token(s) C. In general, the contexts B and C can be any length. This rule has the disadvantage that only the very immediate context can be used, and so a second type of rule, called a collocation rule is also popular and takes the form:
- 109. Section 5.2. Text Classification Algorithms 85 A → d|T which reads A converts to d if a trigger token T is present in the sentence. The main difference is that T does not have to be near to A. Often this approach is called a bag of features approach as it uses evidence based on other tokens without considering their order in the sentence. Token collocations are the most common type to consider to use as we can observe the tokens directly, but it is also possible to use words as triggers also. These are obviously not directly observable so we have to perform some classification before we can determine the words, and of course even once found, our algorithm may not have chosen the word correctly. If we consider some sentences with an ambiguous token , we can see quite clearly how collocation information can help (from Yarowsky [507]). ... it monitors the lead levels in drinking... LEAD-METAL ... median blood lead concentration was ... LEAD-METAL ... found layers of lead telluride inside ... LEAD-METAL ... conference on lead poisoning in ... LEAD-METAL ... strontium and lead isotope zonation ... LEAD-METAL ... maintained their lead Thursday over ... LEAD-AHEAD ... to Boston and lead singer for Purple ... LEAD-AHEAD ... Bush a 17-point lead in Texas, only 3 ... LEAD-AHEAD ... his double-digit lead nationwide. LEAD-AHEAD ... the fairly short lead time allowed on ... LEAD-AHEAD One simple rule based systems works as follows. Firstly, we require a lexicon which gives the orthography for each word. With this, we can easily determine check whether two words in the lexicon have the same orthography, and from this compile a list of tokens which have two or more word forms. Next we write by hand a set of rules which fire (i.e. the activate) on the presence of the trigger token, for example: concentration LEAD-METAL levels LEAD-METAL singer LEAD-AHEAD dog LEAD-AHEAD At run-time, we then move through the sentence left-to-right examining each token in turn. When we find one that is ambiguous (that is, it appears in our list), we look for trigger tokens at other positions in the sentence which will form a collocation with the current token. In longer sentences we may however find several collocation matches, and so exiting the list on finding the first may result in an incorrect decision. One extension then is to search the entire list and choose the word that gives the most matches. An alternative is to consciously order the list in some way, with the idea that the choices at the top of the list are more definite indicators than those else where. For example, if we find bass followed immediately by guitar, we can be
- 110. 86 Chapter 5. Text Decoding: Finding the words from the text almost certain that this is BASS 1 (rhyming with “base”) and not the type of fish. Another way of using the rules is to run the most specific rules first, so if we have a rule which has several trigger tokens, or a very rare trigger token (e.g. strontium) these fire before more general rules. The idea here is that very specific rules are much more likely to be appropriate for that word only. There are fundamental difficulties with regard to this and similar hand written rule ap- proaches. Firstly, it can be extremely laborious to write all these rules by hand. Secondly, we face situations where we have an ambiguous token but no trigger tokens that are present in our collocation list. Both these problems amount to the fact that the mapping from the feature space to the label is only partially defined. Finally, we face situations where say 2 rules match for each word - which do we then pick? For these reasons, we now consider machine learning and statistical techniques. 5.2.5 Decision lists One of the weaknesses of the simple collocation list is that it doesn’t tell given any indication of how strong a collocation might be. While we can use ordering (as in the case of bass guitar), finding the correct ordering can be difficult and can still cause problems when contradictory rules fire. An alternative is then to use statistics to give us an indication of how strong or how likely a particular collocation is. One such method is to use decision lists, introduced by Rivest [373], and used in TTS by Yarowsky [506], [507], Gale et al [169] [125] and Sproat [411]. The basic idea of decision lists is to determine the strength of the link between a token and its trigger, and this is done by measuring the frequency of occurrence. In the following table (taken from [507]), we see the raw counts which have been found for the token bass: Trigger BASS-FISH BASS-MUSIC fish 0 142 guitar 136 0 violin 49 0 river 0 48 percussion 41 0 salmon 0 38 from this it is clear that the trigger words are very strong disambiguation indicators, but also that some triggers occur more frequently that others. From these counts, we can calculate the conditional probability that of seeing a word given the collocation, e.g. P(BASS-FISH|guitar,bass) P(BASS-MUSIC|guitar,bass) We then compute the log-likelihood ratios for each collocation: abs log P(BASS-FISH|collocationi) P(BASS-MUSIC|collocationi)
- 111. Section 5.2. Text Classification Algorithms 87 which then give us a sorted decision list: Log likelihood Collocation token word 10.98 fish BASS-FISH 10.92 stripped BASS-FISH 8.87 sea BASS-FISH 9.70 guitar BASS-MUSIC 9.20 player BASS-MUSIC 9.10 piano BASS-MUSIC At run time, we simply start at the top of the list and progress downwards until we find the colloca- tion which matches our input sentence. We then take the word that results from that as the answer. There are many subtle modifications, which for instance smooth the probability estimates, count all the evidence (not just the best case) and which consider the conditional probability of a word being the answer when we have several triggers present. Yarowsky notes however that the simple approach is often good enough that many of these modifications don’t add any extra accuracy. 5.2.6 Naive Bayes Classifier The naive Bayes classifier is a simple and popular approach in word sense disambiguation. In this, we attempt to estimate the probability that we have word wi is the correct one for a token based on a number of features: P(W = wi| f1, f2,...fN) (5.1) This is difficult to estimate directly so instead we make use of Bayes rule which gives us a more tractable formulation: P(W = wi| f1, f2,...fN) = P(W = wi)P(f1, f2,...fN|wi) (5.2) In this, the expression P(f1, f2,...fN|wi) is easier to calculate than the expression of Equation 5.1 because all we need do now is find a word that gives rise to an ambiguous token and then count occurrences of the features f1, f2,...fN around this word. As we mentioned in Section 5.2.1, in general the interactions between a set of features f1, f2,...fN mean that calculating P(f1, f2,...fN) or P(f1, f2,...fN|w) can be quite complicated because we have to consider every possible combi- nation of features. This approach is called the naive Bayes classifier because it makes the huge simplifying assumption that none of these features interact, such that the are statistically indepen- dent. This means that the expression P(f1, f2,...fN) can be simplified: P(f1, f2,...fN) = P(f1)P(f2),...P(fN) = = N ∏ i=1 P(fi)
- 112. 88 Chapter 5. Text Decoding: Finding the words from the text which means that all we need find are the probabilities that a word is found with a each feature individually. P(W = wi| f1, f2,...fN) = P(W = wi) N ∏ i=1 P(fi) When an ambiguous token is found, the classification is performed by calculating the above ex- pression for each word and choosing the answer as the one with the highest probability. This algorithm has the advantage that the features can be quite heterogeneous; for instance as well as simply the tokens themselves (as with the decision list case), we can also use POS tags, position in sentence and any other features we think might help. While the algorithm is obviously crude in the sense that it ignores feature dependencies, it does at least give a defined mapping from point in the feature space. 5.2.7 Decision trees Given the simplicity of the naive Bayes classifier it performs surprisingly well. However, the assumption that the features are independent is in general too crude and better performance should be obtainable by a system which models interaction between features. To see the problem of feature dependence we might see that the token Morgan is a strong indicator that we are have with BANK MONEY and not BANK RIVER. We might also find that the token Stanley is just as strong an indicator. But it is clearly false to assume that these tokens operate independently of one another and that it is just coincidence that they appear in the same sentence. Morgan Stanley is of course the name of a real bank and those two tokens very frequently co-occur. Once we have used one feature, seeing the second adds very little extra information that we are dealing with BANK MONEY. In general though, we have seen that it is intractable to assume that every combination of features is unique and needs to be calculated separately: we simply never have enough training data to estimate such models accurately. One popular solution to this is to use decision trees. These model interactions between features but do so in a particular way so as to concentrate on learning the influence of the most important features first, such that rare or undiscriminating feature combinations can be ignored with some safety. We will explore decision trees in full in Section 15.1.9; for now we will give a quick overview and show how they can be used for text classification. At run time, a decision tree asks a series of questions in order, and based on the outcome of these delivers a label. In essence then, this is no more than series of nested “if then else” constructions in a computer program. For our cases, one set of questions might be “if the token is lead and the following token is contains the string poison and if the token pollution or heavy is also found in the sentence then mark this as LEAD METAL”. Any series of questions can be formulated as a tree, and in fact decision trees can be written by hand. We do however have a number of algorithms which automatically build trees and it is
- 113. Section 5.2. Text Classification Algorithms 89 because of these that the technique is popular. The decision tree growing works as follows: 1. Create an initial group containing all the instances of one ambiguous token. 2. Design a set of questions based on our features which we think will help in the word identi- fication 3. For each question (a) form two new groups based on the value of the feature (b) measure the combined impurity of the new groups 4. Find the question which gives the biggest reduction in impurity 5. Delete this question from the list of features to be examined. 6. Form two new groups based on this question. 7. Repeat steps 3 to 6 on each new group until the stopping criteria have been met. The impurity of a cluster can be formulated in a number of ways as we shall see in Section 15.1.9. Here even a simple measure will do, for example the ratio of word-1 over word-2. The stopping criteria usually involves specifying a minimum decrease in impurity. The decision tree gets round the problem of modelling all the feature combinations by effectively clustering certain feature combinations together. This is not always ideal, but it certainly does allow for more accu- rate modelling than naive Bayes, as we can see that the same feature can appear in different parts of the tree and have a different affect on the outcome. 5.2.8 Part-of-speech Tagging A somewhat different approach to the problem is to use a tagging algorithm. This type of al- gorithm arose from the problem of part-of-speech (POS) tagging which was concerned with finding the sequence of syntactic categories or parts of speech for each token in a sentence. These algorithms are of interest to us in two ways. Firstly, many homographs are part-of-speech homo- graphs. In English, these are often noun/verb pairs where the same token can be used as a noun or verb from the same semantic root. In fact a huge number of words in English have noun/verb pairs, so that can have GO FOR A WALK and I LIKE TO WALK. In most cases these are syntactic homonyms meaning that while they may operate differently with regard to syntax, they are spelt and pronounced the same way. A few however are spelt the same way but pronounced differ- ently and these are the cases which concern us. The idea of POS-tagging is that if we find the part-of-speech from the text we will know which word we are dealing with. A second interest in POS-tagging is that this technique can be used to assign any type of labels to the tokens, not just parts-of-speech. Hence we could use a tagging algorithm for general disambiguation purposes. POS-tagging has been a problem of concern in NLP for some time. There it is often used as a first required step before syntactic parsing is performed, but it is also used in other applications and seen as a problem in its own right. Many approaches have been tried but probably the most
- 114. 90 Chapter 5. Text Decoding: Finding the words from the text NOUN VERB ADJ DET 0.00 0.01 0.38 0.61 0.32 0.23 0.07 0.38 Figure 5.1 Example transition probabilities from a simplified 4 state POS tagger. The probabilities exiting a state always add up to 1 the a an some that now walk 0.00 sing plane hill 0.003 0.001 0.002 0.00 0.00 0.00 0.00 0.00 0.00 0.002 DET NOUN the a an some that now walk 0.63 0.22 0.07 0.05 0.03 0.00 0.00 sing plane hill 0.00 0.00 0.00 VERB the a an some that now walk 0.00 sing plane hill 0.00 0.00 0.00 0.00 0.00 0.002 0.00 0.001 0.003 Figure 5.2 Example observation probabilities for 3 POS tags, showing the probability of observing a token given the state.
- 115. Section 5.2. Text Classification Algorithms 91 common approach is to use a hidden Markov model. These are dealt with fully in Chapter 15, but for now we can give a quick overview explaining their use for tagging purposes. An HMM tagger operates probabilistically to find the most likely sequence of tags (labels as we have previously called them) L for an input sequence of tokens T. We write this as L̂ = argmax l P(L|T) Rather than compute the probability that a token generates a tag P(l|t), an HMM is a generative model which computes P(t|l), that is, the probability of seeing a token given the tag. We can calculate one from the other by concatenating the HMM models to form sequences and then apply Bayes rule: P(L|T) = P(L)P(T|L) P(T) Where P(L) is called the prior and is the probability of a particular sequence of tags occurring. P(T) is called the evidence, that is, what we actually see, and P(T|L) is called the likelihood. An HMM can be viewed as a network of states, where each state represents one POS tag. In a tagger, all the states are connected, and the connection between each pair of states is given a transition probability. This represents the probability of seeing one POS tag follow another. A simple example of this is shown in Figure 5.1 and in this example we see that the probability of a noun following a determiner is high, whereas the probability of a determiner following another determiner is very low. In addition, each state has a set of observation probabilities which give the probability of that state generating each token. An example of this is seen in Figure 5.2 for the determiner POS tag. As we would expect, tokens such as the have high probabilities, whereas content tokens such as walk have low probabilities meaning it is highly unlikely that this tag would generate that token. We will consider the full issue of how to train an HMM in Section 15.1.8. For now, let us simply assume that we can calculate the transition probabilities and observation probabilities by simply counting occurrences in a labelled database. To see how the tagging operates consider the issue of resolving the classic POS homograph record. This can be a noun or a verb, and a trained HMM would tell us for instance: P(t = record|l = NOUN) = 0.00065 P(t = record|VERB) = 0.0045 Which basically states that in our training data one in every 1538 nouns that we encountered was the token record, and one in every 220 verbs we encountered was record. In addition to this we have the prior model which gives us the probability that a sequence of tags will occur independent of any actual observations. In general this would take the form:
- 116. 92 Chapter 5. Text Decoding: Finding the words from the text P(L) = P(l1,l2,...,lN) that is, the probability of seeing all the tags in the sentence. We then wish to find the probability of a particular tag given all its predecessors: P(li|l1,l2,...,li−1) We would never have enough data to count all these occurrences, so we make an assumption based on the chain rule of probability (see Appendix A). This states that a sequence as above can be approximated by considered only a fixed window of tags before the current one: P(li|li−2,li−1) or P(li|li−1) In this model we might find for instance that the probability of two particular tag sequences are P(NOUN|DET) = 0.019 P(VERB|DET) = 0.00001 which indicates that the chance of a verb following a determiner is much much smaller than the chance of a noun following a determiner. Given an input token sequence which includes ... the record ... we would be certain that the token the would be classed as a determiner, and so when we can combine the two models and calculate the probability we see that the token record is a noun rather than verb. In a real HMM tagger system, we have to determine these probabilities from data. In general, this data has to be labelled with a POS tag for each word, but in cases where labelled data is scarce, we can use a pre-existing tagger to label more data, or use a bootstrapping technique where we use one HMM tagger to help label the data for another iteration of training. The prior probability is calculated by counting occurrences of sequences of tags. This is modelled by an n-gram model which gives the probability of finding a tag ti given a fixed length sequence of previous tags ti−1,ti−2.... The longer the tag sequence the more accurate the prior model, but in general sparse data problems prevent us from having an n-gram of more that 3 or 4 tags in length. We also have to build a likelihood model, and we do this by defining a separate model for every tag. Each of these models then defines a probability distribution over the set of tokens, giving the probability of seeing any one token given the tag. At run-time, a search is conducted through all possible tag sequences to find the single most likely, and this is given as the answer. Another advantage of the HMM approach is that it is amenable to computing every
- 117. Section 5.3. Non-Natural Language Text 93 possible sequence of tags and choosing the one with the highest probability. This is performed using the Viterbi algorithm, explained in Section 15.1.7. We will return to the entire subject of HMMs in Chapter 15 where we discuss the general nature of the likelihood models, the prior n-gram models, the search algorithms, the training algorithms and other practical matters. 5.3 NON-NATURAL LANGUAGE TEXT We now turn to the problem of dealing with text which does not directly encode natural language. In our model of communication we have a number of semiotic systems which are used to com- municate information. Natural language is the richest and often the most common of these, but we also have numbers, dates, times, computer information, money, addresses and so on. These systems are separate from natural language, but in TTS we are required to speak them, and this can only be done if we convert data in this form into natural language. This is a two stage problem where first we must analyse the text to determine its units and structure; this is done by a process of semiotic decoding. Then we have to convert this into natural language, and this process is called verbalisation. But we have an additional problem in that as these systems share the same characters as natural language and hence each other, the text forms of these systems often overlap, leading to considerable ambiguity. Hence before we attempt decoding and translation, we first have to perform semiotic classification to resolve this ambiguity and determine which system each token of the input belongs to. 5.3.1 Semiotic Classification Semiotic classification can be performed by any of the techniques described in Section 5.2.1 above. First, we define which classes we believe our system will encounter. We can immediately identify some which crop up regardless of text type and genre, and these include: cardinal numbers 25, 1000 ordinal numbers 1st, Henry VIII telephone numbers 0131 525 6800, 334-8895 years 1997, 2006, 1066 dates 10 July, 4th April 10/12/67 money $56.87, 45p, Y3000, $34.9M percentages 100%, 45% measures 10cm, 12 inches In addition, we have a number of classes which are more genre specific; for example we may have street addresses in a telephone directory application, computer programs in a screen reader for the blind and any number of systems in specialist areas such as medicine, engineering or construction:
- 118. 94 Chapter 5. Text Decoding: Finding the words from the text emails pat40@cam.com urls http://mi.eng.cam.ac.uk/ pat40 computer programs for (i = 0; i n; ++i) addresses 4318 East 43rd Street real estate CGH, sng whit fem So it is important to have a good knowledge of the application area in order to ensure a good and accurate coverage of the type of semiotic systems that will be encountered. It should be clear that a simple classifier based on token observations will only be sufficient in the simplest of cases. If we take the case of cardinal numbers, we see that it would be absurd to list the possible tokens that a cardinal class can give rise to; the list would literally be infinite. To handle this, we make use of specialist sub-classifiers which use regular expression generators to identify likely tokens from a given class. One way of doing this is to run regular expression matches on tokens and use the result of this as a feature in the classifier itself. So we might have (34) Match [0-9]+ ? as a rule, which will return true if the token is composed entirely of an arbitrary number of digits. More specific regular expressions can also be used which handle numbers of a fixed length (e.g. years in the modern era being 4 digits long,) or for real numbers which contain a decimal point. In addition to these open class rules, we also see that semiotic classes form a hierarchy. So the money class is composed of the number class and usually a currency symbol (e.g. $). A further issue is that often several tokens form a coherent semiotic group, such that if we have (35) 8341 Belville Dr we have three tokens that together form a single address. Hence we need an ability to group the tokens after or during classification. The process of semiotic classification can therefore be summarised as: 1. Define the set S = {s1,...sN} of semiotic classes 2. For each decide if this is a closed or open class. 3. Closed classes can be enumerated, open classes require regular expression matches. 4. For each token: (a) run the open class expressions and record the answer for each as a feature (b) Using the open class features and the normal features, run a classifier (e.g. the naive Bayes classifier), record the most probable class si for each token. 5. Group consecutive tokens of the same class together 6. Perform class-specific decoding and verbalisation To take a simple example, suppose we have the sentence: (36) You should meet him at 3:30pm at 1994 Melville Drive and we have the set of semiotic classes as just defined. For each open class we would have a regular expression as follows:
- 119. Section 5.3. Non-Natural Language Text 95 natural language [A-Za-z]+ cardinal numbers [0-9]+ telephone numbers 0131 525 6800, 334-8895 years 1[0-9][0-9][0-9] | 2[0-9][0-9][0-9] time hours : minutes part-of-day hours [0-9] | 1[0-9] | 2[0-4] minutes [0-6][0-9] part-of-day am | pm addresses cardinal number proper-name street-type proper-name [A-Z][a-z]+ street-type Dr| St| Rd| | Drive| Street | Road For each token we run each regular expression and record its answer as a feature: text natural language cardinal year hours minutes part-of-day proper-name You 1 0 0 0 0 0 1 should 1 0 0 0 0 0 0 meet 1 0 0 0 0 0 0 him 1 0 0 0 0 0 0 at 1 0 0 0 0 0 0 3 0 1 0 1 0 0 0 : 0 0 0 0 0 0 0 30 0 1 0 0 1 0 0 pm 1 0 0 0 0 1 0 at 1 0 0 0 0 0 0 1994 0 1 1 0 0 0 0 Melville 1 0 0 0 0 0 1 Dr 1 0 0 0 0 0 1 The results of these matches can now be used as features, along with general features such as “is this the first word in the sentence” (which is useful for telling a normal natural language word from a proper name). All the features are then given to the classifier which makes a final decision on semiotic class. With regard to the issue of grouping, we can use a combination of the classifier and regular expression matches to ensure that a sequence of tokens such as 3, :, 30, pm are seen as a single time. A more sophisticated approach is available if we make use of the HMM tagging algorithm as one of the features of an HMM is its ability to include multiple tokens in a single class, and also, via the Viterbi search algorithm, ensure that the most probably global sequence of classes is found, which helps resolves problems where two classes claim overlapping tokens.
- 120. 96 Chapter 5. Text Decoding: Finding the words from the text 5.3.2 Semiotic Decoding Once identified, each token or sequence of tokens must be decoded to determine its underlying form. Luckily for nearly all classes except natural language, this is fairly straightforward issue. In fact many classes (e.g. email addresses or computer programs) have been specifically designed to ensure that the underlying form can be found quickly and deterministically. We have call this process decoding as it is seen as the process of reversing an encoding of the original form into text. This is often also be called parsing in computer science, but we reserve this term for the specific problem of uncovering syntactic structure in natural language. We require a separate decoder for each semiotic class, but due to the simplicity of each, this in is general is not too difficult to design by hand. Let us take the example of dates to see how a decoder would work. Let us take the Gregorian calender system (i.e. the date system we nearly always use) as an example. This is particularly illustrative as a non-natural language semiotic class in that 1. It has a number of text forms, e.g. 10/12/67 or 10th December ’67 2. It is used nearly universally, regardless of which natural language is being used. 3. There are a number of ways to verbalise a date Gregorian dates have a day a month and optionally a year and the task of the decoder is to deter- mine the values of each of these from the text. The same date can give rise to a number of different text encodings, for example: (37) a. 10 December 1967 b. December 10 1967 c. 10th of December 1967 d. 10th of December ’67 e. 10/12/1967 f. 10/12/67 g. 12/10/67 In each of these cases, the date is exactly the same, and so here we see how a single date can give rises to multiple text encodings. Despite these possible differences, finding the day month and year is normally possible with a little care (the only really difficult case being the handling of the convention often found in the United States, where the pattern is month/day/year rather than the more common day/month/year). A typical way of actually performing the decoding a date is to use a list of rules, each of which decodes one particular date format. The rules are run in order, until a match is found. 5.3.3 Verbalisation The final step in handling non-natural language text is to convert it into words and this process is often called verbalisation. If we take our decoded date example, we see that we have values for
- 121. Section 5.3. Non-Natural Language Text 97 the three fields of day month and year, and with these we can generate a sequence of words. For dates we have a number of conventions, such that the date day=10, month=12, year=December can be spoken in English as: (38) a. TEN TWELVE SIXTY SEVEN b. THE TENTH OF DECEMBER SIXTY SEVEN c. THE TENTH OF DECEMBER NINETEEN SIXTY SEVEN d. DECEMBER THE TENTH NINETEEN SIXTY SEVEN e. DECEMBER TENTH NINETEEN SIXTY SEVEN f. DECEMBER TENTH SIXTY SEVEN g. DECEMBER THE TENTH IN THE YEAR OF OUR LORD NINETEEN SIXTY SEVEN and so on. In this case the issue of verbalisation is not difficult per-se - a simple template approach can be used to create the word sequence. First we define two functions ordinal() and cardinal() which convert a number into a word string representing the type of number. For example, for cardinal() we have: 1 → ONE 2 → TWO 3 → THREE 13 → THIRTEEN 32 → THIRTY TWO 101 → ONE HUNDRED AND ONE 1997 → ONE THOUSAND NINE HUNDRED AND NINETY SEVEN and for ordinal(): 1 → FIRST 2 → SECOND 3 → THIRD 13 → THIRTEENTH 32 → THIRTY SECOND 101 → ONE HUNDRED AND FIRST 1997 → ONE THOUSAND NINE HUNDRED AND NINETY SEVENTH next we define a function year(). A simple version of this which covers all years between 1010 and 1999 is: (39) cardinal(y1y2) cardinal(y3y4) which when give 1387 for instance will produce the string THIRTEEN EIGHTY SEVEN. For the years immediately after the millennium the situation is more tricky, where we require a function of the type: (40) cardinal(y1000) AND cardinal(y4)
- 122. 98 Chapter 5. Text Decoding: Finding the words from the text which when given 2006 would generate (41) TWO THOUSAND AND SIX Here we start to see the first case of why verbalisation can be tricky as some people prefer this rule not to have the AND so that the function should produce (42) TWO THOUSAND SIX and there is some debate over whether 2013 will be pronounced TWENTY THIRTEEN or TWO THOUSAND AND THIRTEEN. With the year() function we can now define a template for verbalis- ing the full date. One template would then be: (43) THE ordinal(day) of month-name(month) year(year) giving (44) THE TENTH OF DECEMBER NINETEEN SIXTY SEVEN In general, verbalisation functions for cases like these are not too difficult to construct, all that is required is a careful and thorough evaluation of the possibilities. The main difficulty in these cases is that there is no one correct expansion; from 38 we see at least 7 ways of saying the date, so which should we pick? While sometimes the original text can be a guide (such that 10/12/67 would be verbalised as TEN TWELVE SIXTY SEVEN) we find that human readers often do not do this, but rather pick a verbalisation that they believe is appropriate. The problem for us as TTS system builders is that sometimes the verbalisation we use is not the one that the reader expects or prefers, and this can be considered an error. For some each semiotic classes, we can specify a degree of conventionality with regard to verbalisation. We can state that ordinals and cardinals verbalisations are very conventional in that nearly all readers verbalise a number the same way (although there may be slight differences between American and other types of English.) With dates, we see that these are again quite conventional, but that there is quite a range of acceptable forms. We find that there is a significant problem with some other semiotic forms in that the degree of convention can be quite low. If we consider the email addresses (45) pat40@cam.ac.uk we might find that any of the following are acceptable verbalisations: (46) LETTER-P LETTER-A LETTER-T FORTY AT CAM DOT LETTER-A LETTER-C DOT LETTER- U LETTER-K LETTER-P LETTER-A LETTER-T FOUR ZERO AT CAM DOT LETTER-A LETTER-C DOT LETTER-U LETTER-K PAT FORTY AT CAM DOT LETTER-A LETTER-C DOT LETTER-U LETTER-K PAT FORTY AT CAM DOT AC DOT UK The problem being, that the particular owner of this address may only consider one of these as “correct” and will be dissatisfied if an alternative version is used.
- 123. Section 5.4. Natural Language Text 99 5.4 NATURAL LANGUAGE TEXT Having dealt with non-natural language text, we now turn our attention to the issue of how to find the words from the text when that text does in fact encode natural language. For many words, this process is relatively straightforward; if we take a word such as WALK, we will have its textual form walk listed in the lexicon, and upon finding the a token walk in the text we can be fairly sure that this is in fact a written representation of WALK. At worst we might be presented with a capitalised version Walk or an upper case version WALK but these are not too difficult to deal with either. If all words and all text behaved in this manner the problem would be solved, but of course not all cases are so simple, and we can identify the main reasons for this: Alternative spellings A number of words have more than one spelling. Often in English this is related to American vs non-American conventions so that we have tyre and tire and honour and honor. In many cases these differences are systematic, for instance with the suffix -ISE which can be encoded as organise or organize. Often it is possible to list alternative spellings in the lexicon. Abbreviations Many words occur in abbreviated form, so for instance we have dr for DOCTOR and st for STREET. In many cases these are heavily conventionalised such that they can be listed as alternative spellings, but sometimes the writer has used a new abbreviation or one that we may not have listed in the lexicon. Unknown words There will always be words that are not in the lexicon and when we find these we have to deal with them somehow. Homographs are cases where two words share the same text form. In cases where the words differ in pronunciation, we must find the correct word. In natural language homographs arise for three reasons: 1. Accidental homographs arise because completely words such as BASS-FISH and BASS- MUSIC just happen to share the same text form bass. 2. Part-of-speech homographs arise from words which have a similar sematic root, but where their use as a noun or verb gives them a different pronunciation. For instance PROJECT-NOUN and PROJECT-VERB both share the same text form project but are pronounced differently. 3. Abbreviation homographs arise because as we use fewer characters in abbreviations, we are more likely to use the same text encoding as for another abbreviated word. Hence we find that while DOOR, DOCTOR and DRIVE all have the quite separate full text forms of door, doctor, DRIVE, but all share an abbreviated text form dr. Mis-spellings Often if a text form of a word is misspelled a human reader is quite capable of guessing what word was intended, e.g discovery and ideally a TTS system should be able to cope with such instances.
- 124. 100 Chapter 5. Text Decoding: Finding the words from the text This shows why we need a more sophisticated model than a simple isomorphic correspondence between words and their text encodings. 5.4.1 Acronyms and letter sequences At this stage it is worth mentioning two other types of word/token relationships that are often cited as problems for TTS. Acronyms are words formed from sequences of other words, usually (but not always) by taking the first letter of each. Examples include NATO, UNICEF SCUBA and AIDS. In our approach, these are treated as normal words so that we have a word NATO in our lexicon and this has a pronunciation /n ey t ow/; the fact that this was historically formed from other words or is normally found written in upper case is of no real concern. (In fact, with use many acronyms really do become indistinguishable from normal words such that RADAR is nearly always spelled in lower case and few people realise that it is an acronym, and fewer still what the original words were). In lay speak, the term acronym is also used for a different type of formation which we shall call a letter sequence; examples of these include IBM, HTML, HIV. The difference between these and acronyms is that letter sequences don’t form new pronunciations; they are simply spoken as a sequence of letters, hence we have IBM as LETTER-I LETTER-B LETTER-M giving a combined pronunciation /ay b iy eh m/. It doesn’t really matter if we regard IBM as being a single word with pronunciation /ay b iy eh m/ or a token which represents three words (LETTER-I LETTER-B LETTER-M). Potentially however, the productivity of letter sequences is very high, so regardless of whether we keep them in the lexicon or not, we still need a means of identifying new ones and pronouncing them as letter sequences rather than as normal words (i.e. we don’t want to encounter a token RAF and think that this is just an upper case version of an unknown word, and then pronounced it as /r ae f/). A solid solution to the problem of upper case tokens is to assume that in a well developed system all genuine acronyms will appear in the lexicon. Upon encountering a capitalised token, we first check if it is in the lexicon and if so just treat it as a normal word. If not, we then split the token into a series of single character tokens and designate each as a letter. There is a possible alternative, where we examine upper case tokens and attempt to classify them either as acronyms or letter sequences. Such a classification would be based on some notion of “pronouncability” such that if the token contains vowel characters in the certain places then this might lead us to think that it is an acronym instead of a letter sequence. Experience has shown that this approach seldom works (at least in English) as there are plenty of potentially pronounceable tokens (e.g. USA or IRA that sound absurd if treat then as acronyms (e.g. /y uw s ax/ or /ay r ax/). If a mistake is to be made, it is generally better to pronounce an acronym as a letter sequence rather than the way round.
- 125. Section 5.4. Natural Language Text 101 5.4.2 Homograph disambiguation The main problem in natural language text decoding is dealing with homographs, whether they be accidental, part-of-speech or abbreviation homographs. Many systems choose to handle each of these separately, and in particular abbreviation homographs are often with at the same time or in the same module as semiotic classification. Typically this is done in systems which don’t have an explicit semiotic classification stage. An alternative approach is to handle these in a single module. Regardless, the general approach is to use one of the token disambiguation techniques described in Section 5.2.1. Here we discuss some of the specifics of homograph disambiguation. Turning to POS homographs first, we see that the classic way to resolve these is to use a POS- tagger of the type described in Section 5.2.2. One problem with this approach is that unfortunately POS-taggers only have a certain accuracy (96%-98% are commonly quoted figures [299]), and often it is in the hard cases of resolving noun/verb pairs where the errors are made. Recall that the POS tagger has two components, one which models sequences of tags and one which generates tokens given a tag. The tag sequence n-gram is a very powerful constraint and nearly always gives the correct answer in cases such as: (47) The/det project/noun was/aux behind/prep schedule/noun (48) The/det problem/noun is/aux how/wrb to/inf project/verb the/noun missile/noun accurately/adv but in sentences such as (49) The new images project well on to the wall POS-taggers often struggle as the n-gram sequence of tags alone is not sufficient to disambiguate in this case. Part of the problem with POS taggers is that they are often formulated as general purpose NLP tools, and as such are not focused on the goals of how they might be used in a particular application. One simple improvement then is to adapt the tag set so that the discriminative power in the algorithm is geared towards the distinctions we really care about. An alternative approach is to use one of the word sense disambiguation algorithms as these can be more easily focused for each homograph case. A further advantage in using the WSD approach is that we can use this single approach for all type of homograph, not just the POS homographs that the tagger deals with. To give an example, consider the abbreviation homograph Dr which can encode DOCTOR or DRIVE. Both of these are often collocated with a proper noun, but differ in their position relative to that proper noun. A simple feature that distinguishes many examples is then whether the Dr occurs before or after the proper noun: (50) Dr + proper-noun → DOCTOR (51) proper-noun + Dr → DOCTOR for example (52) Dr Williams → DOCTOR WILLIAMS
- 126. 102 Chapter 5. Text Decoding: Finding the words from the text (53) Williams Dr → WILLIAMS DRIVE In deterministic rule system this can be used as is, but as the rule is not 100% accurate an alternative is to use is as a feature in a WSD classifier. We can summarise the procedure for building a homograph disambiguation system as fol- lows: 1. Build a lexicon which records all the text forms of each word 2. Compare text forms and construct a list of homograph tokens 3. For each form (a) search for cases of these in a large text corpus (b) by hand mark the correct word (c) find collocation trigger tokens by finding the tokens which most frequently co-occur with the word (d) write feature rules which may describe any more elaborate relationship between word and trigger tokens 4. Train WSD classifier 5. Run over test corpus, examine errors and repeat training process based on findings. 5.4.3 Non-homographs Most tokens are not in fact homographs, and we will now outline how these non-homograph tokens are processed. 1. Check whether the token is present in the lexicon as the text form of a known word. 2. If it is found, take that word as the correct answer and no further processing is conducted. 3. If in the lexicon and upper case, label it as a letter sequence. Regarding this last rule, It is nearly impossible to sure whether an upper case token is an acronym or letter sequence, but analysis of data shows that unknown letter sequences are much more common than unknown acronyms and hence the most accurate strategy is to assign it as a letter sequence. Furthermore an acronym spoken as a letter sequence is deemed a lesser error than a letter sequence spoken as a single word. If the token is not upper case and not in the lexicon we have the choice, it could potentially be a normal unknown word, an abbreviation for a known word, a different spelling for a known word or a misspelling of a known word. The probability of each of these cases mainly depends on the quality of our lexicon. If we have a very accurate comprehensive lexicon then it is less likely that we have a missed a normal word or alternate spelling. This therefore increases the chance of the token being a mis-spelling. If the system has a spell correction system, it should be run at this point. Assuming the spell corrector has a confidence score we can then decide whether the token is a mis-selling of a known word or an unknown word. The final case occurs when we believe we have a token for a genuinely unknown word. This is assigned a new unique label (e.g.
- 127. Section 5.5. Natural Language Parsing 103 UNKNOWN-43), and the token and label are added to the run-time lexicon. The pronunciation is determined by the grapheme-to-phoneme convertor, which we will describe in Section 8.4. These set of operations can be found in the right half of Figure 5.3. 5.5 NATURAL LANGUAGE PARSING Our concern so far has been the task of identifying the word sequence from the token sequence, and to a large extent this is all the text decoding that we are required to do. Once we have un- ambiguously identified the word sequence, we can use the lexicon and grapheme-to-phoneme conversion to generate a sound representation which can then be input to the synthesiser itself. It can be argued however that the word sequence alone is not enough, and that in addition it is advantageous to conduct further processing to determine the underlying structure of the sentence. Semantic analysis is generally thought too difficult and/or unnecessary, but a number of systems perform a syntactic analysis which finds the correct or best syntactic tree for the input sentence. There are several justifications for wishing to perform this additional level of analysis. 1. Proper parsing can resolve part-of-speech ambiguity problems that finite state (ie HMM parsers) fail on, such as Example 49 above. 2. The syntax can help determine sentential prominence patterns (explained in Section 6.8). 3. The syntax gives a basic grouping of words, and these groupings strongly influence the prosodic phrase structure of the utterance (explained in Section 6.7). 5.5.1 Context Free Grammars In Chomsky’s early work [88], he proposed that the syntax of natural languages like English could be described at a basic level by a context free grammar (CFG). A CFG defines a number of rules of the form (54) S → NP VP NP → DET NP VP → V NP VP → PP PP → P NP The rules operate generatively meaning that they are applied by starting with a basic top node (S standing for “sentence”) and then applying rules in turn until a string of terminal nodes (ie words) are generated. Chomsky’s idea was that this finite set of these rules could generate the near infinite set of sentences in natural language. In addition, the operation of the rules defined the structure structure of a sentence. The operation of context-free rules is easily represented by a syntactic tree:
- 128. 104 Chapter 5. Text Decoding: Finding the words from the text S NP Det the N man VP V walked PP PREP down NP Det the N street Formally, a grammar is a 4-tuple 1. σ a set of of terminal symbols 2. n a set of non-terminal symbols 3. P a set of re-write rules 4. S an initial (non-terminal) symbol Chomsky soon made many additions to this basic system and the field of modern linguistics has dispensed with the idea that simple context free rules alone describe natural language. In lan- guage engineering however, CFGs are still widely used as they are (reasonably) tractable models of syntax. The problem of finding the syntactic structure of a sentence is called parsing and a parser is an algorithm which assigns this structure to arbitrary text. All parsers work with respect to a grammar and effectively try to assign the best or correct structure to a sentence. Our problem is of course that we are starting from the other perspective in that we have the words already, and wish to find which rules generated them. The job of the parser is to therefore find which rules can have generated the input, and from these rules we can find the tree and hence the syntactic structure of the sentence. The first problem that we encounter in parsing is that any simplistic or exhaustive search for the correct set of rules is prohibitively expensive. Parsers can therefore be thought of as search algorithms which find a tree or trees for a sentence in reasonable time. The second problem in parsing is that any non-trivial set of rules will generate more than one parse that fits the input, and in many cases an considerable number of perfectly legal parses will fit the input. A simple example is: (55) POLICE HELP DOG BITE VICTIM which could have the parse
- 129. Section 5.5. Natural Language Parsing 105 S NP N Police VP V help N dog VP V bite N victim meaning the police were helping the dig to bite the victim, or S NP N Police VP V help AP AP Adj dog Adj bite N victim meaning the police were helping a victim of a dog bite. The Cocke-Younger-Kasami (CYK) algorithm is one of the most widely used parsing al- gorithms. It can (in general) efficiently parse text and is easily adaptable to probabilistic formula- tions. Its basic operation can be thought of as dynamic programming for context free grammars. This algorithm considers every possible group of words and sets entries in a table to be true if the sequence of words starting from i of length j can be generated from the grammar. Once it has considered sequences of length 1, it goes on to ones of length 2, and so on. For the sequences greater than length length 1, it breaks the sequence into all possible subdivisions, and again checks to see if a rule in the grammar can generate one of these subdivisions (this is why the grammar has to be in Chomsky normal form). If a match is found, this is stored in the table also. This process is continued until the full set of rules (ie. parses) is found. Deterministic parsers have no real means of choosing which parse is best, so probabilistic parsers (explained below) are usually preferred as these at least show that one parse is more probable than another. 5.5.2 Statistical Parsing The CYK algorithm as just explained will find the set of legal parses for a sentence, but the problem is in general that we wish to find just one parse, not several. We of course want the “true” parse and by this we mean we want what the author originally intended when the sentence was written.
- 130. 106 Chapter 5. Text Decoding: Finding the words from the text The major weakness with this type of parsing is that it only considers syntactic factors, semantics is completely ignored. That is why a parser often returns parses that are “correct” but seem silly. Some more advanced models of syntax do incorporate semantic information [78], [177], [312], but in general most syntax models, especially in engineering, only make use of syntactic rules. A sensible alternative then is instead of trying to find the parse which the author intended, or which makes most sense, we can try and find the most likely parse. That is, we build a probabilistic model of the grammar and use the parser to find the most likely parse. There are a number of ways this can be done, but the most common is to assign probabilities to the rules themselves so that we say a non-terminal expands to each of its productions with a certain probability. When the parsing is complete, each set of rules has an associated probability and the highest can be chosen as the right answer. A number of specialist databases called tree-banks are available for training PCFGs. These have trees and nodes labelled by hand, and from these a grammar can be trained and probabilities assigned to each rule. For unlabelled data (that is sentences which we have no trees for) the inside-outside algorithm can be used. This is the PCFG equivalent of the Baum-Welch algorithm used to train hidden Markov models, and is explained in Section 15.1.8. We can therefore identity three common approaches to parsing: 1. Define a CFG by hand and use a deterministic search algorithm (i.e. the parser) to find the parse or parses which best fit the input 2. Define a CFG by hand and use a probabilistic search algorithm to find the parse or parses which is most likely given the input 3. Learn the CFG from data and use this with a probabilistic search algorithm to find the parse or parses which is most likely given the input. 5.6 DISCUSSION Figure 5.3 shows the flowchart for the text decoding system described in this chapter. At this stage, we will discuss a number of alternate formulations to the one presented here. In the (distant) past, text forms such as Dr were often see as abberations as far as text and lan- guage processing was concerned to the extent that these were often called text anomalies [276] . The “real job” of text/language processing was to perform syntactic and morphological process- ing. For instance in the MITalk book 6 pages are devoted to text anomalies and 48 to syntactic, morphological and phonological processing. As can be seen from this chapter, in modern systems, the opposite balance is seen. Another view is that this problem is one of text normalisation in which the a form like Dr should be converted into its true text form doctor. In such systems there is often no formal distinction between the concept of words and tokens. A more recent ap- proach is to define such text forms as non-standard words (NSWs) [413]. This approach takes the notion of these text forms much more seriously, and for example notes that the text forms for a word are quite productive and can’t simple be listed. The NSW approach uses decision tress
- 131. Section 5.6. Discussion 107 In homograph list? In lexicon? Upper Case token? mis spelling possible spell corrector convert to letter sequence homograph disamgiuator lexical lookup natural language verbalisation verbalisation verbalisation verbalisation decoder decoder decoder decoder email time number date date number time email yes yes yes yes no no no grapheme to phoneme conversion words no (genuine unknown word) classifier semiotic tokens non−natural language Figure 5.3 Flow of entire text analysis system (only four semiotic decoders/translators are shown for succinctness. to perform disambiguation and is in many ways similar to the approach taken here. Two notable differences are that it again doesn’t make an explicit distinction between the nature of words and tokens, and secondly, the processes of non-natural language and natural language disambiguation are performed in a single step. We argue that there are a number of advantages to the text decoding model used here. Tak- ing the issue of natural language decoding first, we see that this follows the information theoretic
- 132. 108 Chapter 5. Text Decoding: Finding the words from the text source/channel paradigm where we have a clean, discrete underlying form that has generated the noisy surface form that we observe. This is a very general and powerful model and has been shown successful in many speech and language applications. It facilitates the ability to have abbrevia- tions, misspellings and ambiguity all handled in a single unified powerful statistical framework. Probabilistic models can be built for the general process of text encoding that will facilitate suc- cessful decoding. In addition, once the words are found in this model, all problems of ambiguity are resolved, which greatly simplifies the work of say the pronunciation module. The model of semiotic systems described here makes explicit the notion that text encodes other information that natural language information. This is well understood in the field of semi- otics but is nearly universally ignored or brushed over in speech and language processing. There are two main advantages we see in adopting this model. First, it makes explicit the difference between analysis and verbalisation. In analysing text tokens (for example the time 3:30pm) there is one, and only one, correct answer which is hour=3, minutes=30, part-of-day=afternoon. When it comes to verbalisation however, there are many ways in which we can render this as words, e.g. THREE THIRTY, HALF PAST THREE, HALF THREE etc and different users will have different pref- erences depending on their dialectal preferences and system application. The upshot is that it is significantly easier to build a system where we know that only one part suffers from this subjective problem - the analysis/decoding part can be attempted with complete assuredness. The second advantage comes when we consider language dependence. It may not be im- mediately apparent, but the text 3:30 is nearly universally recognisable no matter which natural language surrounds it so we can have (56) THREE THIRTY in English and (57) TROIS HEURES TRENTE in French. If we consider 3:30 as simple a strange encoding of natural language, we would have to build a separate analysis system for every language. With the semiotic model, one classifier and decoder suffices for all languages; all that is required is a language specific verbalisation component 1 We need to comment on the shift in emphasis in text analysis between older systems which concentrated on syntactic, morphological and phonological analysis and paid little attention to text decoding issues, and modern systems which perform little linguistic analysis but devote con- siderable attention to ambiguous text forms. The are three separate factors influencing this shift. Firstly, with the commercialisation of TTS, we have found that in real use, non-canonical word encodings and non-natural language text is rife and any TTS system must deal with this compre- hensively and robustly if it is to be judged usable. Secondly, the simple length of time we have had to collect data and particularly lexicons over the years means that we have very comprehensive 1 It is perhaps worth mentioning at this stage that books on teaching languages to non-native speakers are often particularly good at explaining the rules as to how numbers, times and dates are verbalised in particular languages.
- 133. Section 5.6. Discussion 109 lists of words, their pronunciations and their text forms. This means that morphological analysis for instance, is much less necessary than before and that data driven methods are quite practical in terms of training (we have the data) and run time (we can perform searches or run algorithms with large numbers of parameters/rules). Thirdly, and perhaps most profoundly, pure linguistic analysis such as syntactic parsing has fallen out of favour for the simple reason that it has failed “to deliver”. Few doubt that a full, accurate and true description of a sentences linguistic structure would be very useful, but what form exactly should this take? Modern linguistics has more theo- ries of syntax than even specialists can keep up with, and disagreements run to the very nature of what syntax is and what it represents (and there are now plenty who see it as an epiphenomenon). From a more practical perspective, despite the use of tree-banks, there simply is no such thing as a theory-independent model of syntax which we can use for engineering purposes. All use a grammar and all such grammars make strong claims about the nature of language (such as whether it is context free). Next, and maybe because of these problems, we find that even the best parsers are really quite inaccurate and relying on their output can cause quite noticeable errors in other parts of the system. Finally, we find that even if we had a good model of syntax and an accurate parser, we still may not find syntactic analysis helpful. As we shall see in the next chapter, the main uses for syntactic structure are in determining prominence, intonation and phrasing; but even if we knew the syntax it is far from certain how to derive these other sorts of information. The relationship between syntactic and prosodic structure is a great void in linguistics. Many regard these problems with syntax and other types of “deep” linguistic analysis as ones of expediency; with a more work in these fields we will gradually improve the theories, grammars and parsers to the extent that accurate syntactic structures can be assigned to arbitrary text. An alternative viewpoint is that these difficulties are genuinely profound, and given the enormous work that has been put into the area of theoretical and applied syntax with so little solid progress, that perhaps the nature of the problem itself is incorrectly formulated and that a new approach to the organisational properties of language is required. 5.6.1 Further reading In commercial TTS, text analysis is seen as vitally important as any error can cause an enour- mous negative impression with a listener. Despite this, the problem has only received sporadic attention amongst academic researchers. Notable exceptions include various works by Richard Sproat [410], [411], [412], [413], who tackles nearly all the problems in text analysis (and many in prosody, explained in the next chapter). The Bell Labs system that Sproat worked on is partic- ularly well known for its very high accuracy with regard to text processing. In a separate line of work Sproat et al investigated a number of machine learning approaches to semiotic classification and verbalisation. David Yarowsky [506], [507], introduced an approach which showed the use of the WSD approach to homograph resolution. A more traditional view is given in the MITalk book [10]. Outside the field of TTS we find that a lot of the work in the field of natural language
- 134. 110 Chapter 5. Text Decoding: Finding the words from the text processing (NLP) is directly applicable to these problems. While researchers in those fields are often attempting somewhat different tasks (e.g. homonym disambiguation rather than homograph disambiguation), the approaches are often essentially the same. A particularly good book in this area is by Manning by Schutze, who give a thorough overview of all aspects of statistical language processing, including WSD, POS-tagging and parsing. 5.6.2 Summary Classifiers • Many tasks in text processing involved assigning a label to a unit; this is called classification. • Usually this is done by making use of a set of features that are found elsewhere in the utterance. • A classifier can be therefore be defined as a function which maps from a feature space to a label space. • Treating evidence from features independently is often too crude and inaccurate. Examining every possible combination of features is nearly always impossible due to lack of data. • All sophisticated classifiers therefore adopt an in between solution which gives an answer for every feature combination, but does not necessarily have to have seen all of these during training. The mapping for unobserved feature combinations is inferred from observed ones. • A basic distinction can be made between taggers which assign a label to every token drawn from a single set, and word sense disambiguators (WSD) which assign labels only to am- biguous tokens and do so from multiple label sets. • Popular classifiers include context-sensitive rewrite rules, decision lists, decision trees, naive Bayes classifiers and HMM taggers. Non-natural language text • Text is often full of tokens which encode non-natural language systems such as dates, times, email addresses and so on. • The first stage in text analysis is therefore to perform semiotic classification which assigns a class to each token. This is performed by one of the above mentioned classification algo- rithms • Once a token’s class is found, it is decoded into its underlying form and then translated into natural language by a process of verbalisation Natural language text • The main problem here is to resolve homographs which occur when two words share the same text form. These occur because of accident, part-of-speech changes or abbreviations. • POS-taggers can be used to resolve POS ambiguity
- 135. Section 5.6. Discussion 111 • In general WSD approaches are more focused and produce better results. • In addition some systems use syntactic parsers which generate a syntactic tree for the utter- ance.
- 136. 6 PROSODY PREDICTION FROM TEXT Informally we can describe prosody as the part of human communication which expresses emo- tion, emphasises words, shows speaker attitude, breaks a sentence into phrases, governs sentence rhythm and controls the intonation, pitch or tune of the utterance. This chapter describes how to predict prosodic form from the text while Chapter 9 goes on to describe how to synthesize the acoustics of prosodic from these form representations. In this chapter we first we introduce the various manifestations of prosody in terms of phrasing, prominence and intonation. Next we go on to describe how prosody is used in communication, and in particular explain why this has a much more direct affect on the final speech patterns than with verbal communication. Finally we describe techniques for predicting what prosody should be generated from a text input. 6.1 PROSODIC FORM In our discussion of the verbal component of language, we saw that while there were many dif- ficulties in pining down the exact nature of words and phonemes, broadly speaking words and phonemes were fairly easy to find, identify and demarcate. Furthermore, people can do this read- ily without much specialist linguistic training - given a simple sentence, most people can say which words were spoken, and with some guidance people have little difficulty in identifying the basic sounds in that sentence. The situation is nowhere near as clear for prosody, and it may amaze newcomers to this topic to discover that there are no widely agreed description or representation systems for any aspect of prosody, be it to do with emotion, intonation, phrasing or rhythm. This is not to say that description systems do not exist; the literature is full of them, but rather that none match the simplicity and robustness of the simple systems we use to describe the verbal component of language. There are many schools of prosodic theories, and within one school, everything may have the illusion of simplicity, but the very fact that there are so many different theories often with completely conflicting accounts, leads us to the position of extreme uncertainty with regard to how to describe, model, analyse and synthesise prosody. We mention this here, because firstly these difficulties prevent us from giving a single clean and simple explanation of prosodic phenomena, 112
- 137. Section 6.2. Phrasing 113 and secondly we wish to alert the reader to the very tenuous nature of the literature regarding prosody. To some extent prosody can be seen as a parallel communication mechanism to verbal lan- guage. There we had meaning, form and signal, and broadly speaking we can use these terms when talking about prosody. Hence we can have a meaning (say anger), which is manifested in a form (say raised pitch levels), which is common to all speakers of a group, which is then encoded acoustically in a joint process with verbal encoding. At the semantic (meaning) level, prosody is no more difficult to describe than the semantics of the verbal component (i.e. with great difficulty). It is when we get to prosodic form that we start to see more significant differences between prosody and verbal language. With verbal language, our basic units are words and phonemes, and while these units have some difficulties, they are certainly coherent enough for the engineering purposes of text-to-speech. With prosody though, the situation is much more complex as it is very difficult to determine the basics of what prosodic form should be. We believe on the one hand that abstract prosodic form does exist (we all seem to use similar patterns when asking questions in particular ways), but on the other find it difficult to describe how this should operate in practice. In the next sections then we attempt to give a theory neutral1 account of some prosodic phenomena before describing some of the more well known models. 6.2 PHRASING We use the term prosodic phrasing or just phrasing to describe the phenomenon of where speak- ers groups certain words within the utterances they speak. As will all aspects of prosody, there is acknowledgment that this phenomenon exists, but no agreement as to how best to describe it. 6.2.1 Phrasing Phenomena Many cases are clear cut, so that when we say (58) IN TERMS OF THE NECESSARY POLITICAL EXPEDIENCY | THE RESULT WAS CLEARLY A GOOD ONE FOR SHAW there is a clear break in grouping between the words up to and including EXPEDIENCY and the words that follow (here phrasing is marked with a | symbol). The above case may be unproblematic but there are theories which state that many finer groupings exist, so that we would have: (59) IN TERMS | OF THE NECESSARY | POLITICAL EXPEDIENCY | | THE RESULT | WAS CLEARLY | A GOOD ONE | FOR SHAW (where || indicates a strong phrase break) and furthermore, some theories express this as a brack- eting to show that some groupings are stronger than others: 1 In reality, there is no such thing as theory neutral, we always make some assumptions when we describe anything.
- 138. 114 Chapter 6. Prosody Prediction from Text (60) ((IN TERMS) ((OF THE NECESSARY) (POLITICAL EXPEDIENCY))) (( THE RESULT) (( WAS CLEARLY) (A GOOD ONE)) (FOR SHAW)) A question that has interested linguists for some time concerns the relationship between syn- tactic phrases (ie. the syntax tree) and prosodic phrasing. While there is clearly some relationship (easily seen in the above examples), finding the general nature of this is extremely difficult. Factors which complicate the relationship include: 1. Prosodic phrasing seems “flatter” in that while syntactic phrasing is inherently recursive and can exhibit a large degree of nesting, if levels in prosodic phrasing exist there are only a few, and there is no sense in which the phrases within an utterance seem to embed so strongly. 2. Prosodic phrasing is to some extent governed by purely phonological, phonetic or acoustic factors which can override the syntax. In the classic example, Chomsky [89] commented on the fact that in (61) THIS IS THE DOG THAT WORRIED THE CAT THAT KILLED THE RAT THAT ATE THE MALT THAT LAY IN THE HOUSE THAT JACK BUILT nearly every speaker says this with a flat prosodic phrasing like: (62) THIS IS THE DOG | THAT WORRIED THE CAT | THAT KILLED THE RAT | THAT ATE THE MALT | THAT LAY IN THE HOUSE | THAT JACK BUILT whereas the syntactic phrasing is deeply recursive and embedded (63) (THIS IS (THE DOG ( THAT WORRIED THE CAT (THAT KILLED THE RAT (THAT ATE THE MALT (THAT LAY IN THE HOUSE (THAT JACK BUILT))))))) The speech patterns of the utterance seem to override the syntactic patterns. 3. Phrasing is particularly prone to speaker choice, and while “rules” exist for where phrase breaks may or may not occur, quite often it appears optional whether a phrase break will be found in a particular utterance. Hence all of the following are possible: (64) JOHN AND SARAH WERE RUNNING VERY QUICKLY (65) JOHN AND SARAH | WERE RUNNING VERY QUICKLY (66) JOHN AND SARAH WERE RUNNING | VERY QUICKLY So as we can see the relationship between syntactic and prosodic structure is complicated and a general theory which links syntax and prosody is yet to be developed. Some have even argued that the difficulty really lies with our models of syntax, and that if we developed syntactic models which took more account of the ways sentences were spoken some of these problems might be resolved [1]. 6.2.2 Models of Phrasing While there is no single accepted theoretical model of phrasing, all models can be encompassed within the idea that a sentence has a prosodic phrase structure. The main point of disagreement
- 139. Section 6.2. Phrasing 115 revolves around the details of this phrase structure, and in particular how nested or deep it should be. Some have proposed that phrasing should be represented by a tree and that this should act as a parallel tree to the syntax tree. As such the prosodic tree can be generated by a CFG, and so can have recursion and arbitrary depth. Perhaps the most elegant model of this type is the model of metrical phonology from Liberman [284], [286]. The central idea is that any sentence, or part of sentence has a metrical tree which specifies its phrase and stress patterns. The lowest unit in the tree is the syllable, and so metrical phonology provides a unified theory of syllable, word and phrasal patterns. Ignoring the syllable level for the moment we can see the general form of a metrical trees example below: W W was S clearly S W W a S S good W one S W for S shaw The S and W labels will be explained later, but for now simply note that the metrical tree groups words together. This model seems particularly good at explaining the patterns within com- pound noun phrases, which we will re-examine in Section 6.3.1. A major criticism though is that the theory implies unlimited depth in trees, which implies that a) there are many graded subtleties in the actual production or perception of phrase breaks, or b) that this representation maps to another surface representation that is flatter and has fewer levels. Many solutions have been pro- posed, both as amendments to the original theory2 and as alternatives. The theory of compound prosodic domains proposed by Ladd [268] allows a more limited recursion with arbitrary depth but not full recursion. Another variant, called the strict layer hypothesis [398] is much more constrained and states that every word has a fixed number of prosodic constituents above it. This and many other theories can be seen as forming a model called the prosodic hierarchy. In this, we have a tree as before, but with a fixed number of levels, each of which is its own type. A rich version of this might include the levels utterance, intonational phrase, phonological phrase, word, foot, syllable and segment. In this intonational phrase means a large phrase with a clear intona- tional tune (see below), a phonological phrase is a smaller phrase that spans only a few words, 2 The field of metrical phonology still survives today, but unfortunately bears little resemblance to the elegance and power of the original theory.
- 140. 116 Chapter 6. Prosody Prediction from Text and a foot is a unit which starts with a stressed syllable and covers all the unstressed syllables until the next stressed syllable. It is reasonable to say that nearly all other models are of this general type, and differ only in that they use differing numbers of levels or names for those levels. One of the most popular, especially for engineering purposes has 2 levels of sentence internal phrase, giving 4 levels over all: 1. Utterance 2. Intonational phrase/ Major phrase 3. Phonological phrase/ Minor phrase 4. Word in this scheme, an utterance might look like Major Minor the result Minor was clearly Minor a good one Minor for shaw A slightly different view is to not concentrate on the constituent phrases, but rather the breaks between the phrases. This view has been around for some time, but was formalised most clearly in the ToBI model in the concept of break index [406]. The break index is a number which exists between every pair of words, for example: (67) IN 1 TERMS 3 OF 1 THE 1 NECESSARY 3 POLITICAL1 EXPEDIENCY 4 THE 1 RESULT 3 WAS 1 CLEARLY 3 A 1 GOOD 2 ONE 3 FOR 1 SHAW 6 The numbers give the strength of the break such that between two normal words in running speech we have a 1, and for words which are deemed more separate we have 2 and so on, extending up to 6. A simplified version of this is frequently used, where 3 is given to a minor phrase break and 4 to a major phrase break. Break index systems and strict depth tress are isomorphic, and so the distinction as to which is used is often more of expediency rather than deep rooted theory. A major problem in resolving issues with prosodic phrasing is that there always seems to be a tension between what we can describe as a top down approach, where we use the linguistic structure of the sentence to dictate the phrasing, and a bottom up approach, where we listen to real speech and use clues from pausing, timing patterns and pitch to show where the phrases should be. No model yet proposed has come up with a satisfactory method of doing both of these and reconciling the results.
- 141. Section 6.3. Prominence 117 6.3 PROMINENCE Emphasis, prominence, accent and stress are all terms used to indicate the relative strength of a unit in speech. These terms are used with a variety of definitions in the literature, so to avoid confusion we will use these these terms as follows. Stress indicates lexical stress, which is an inherent property of words, and indicates for instance that the first syllable in TABLE is stronger than the second, while the second syllable in MACHINE is stronger than the first. We discuss this type of stress more fully in the section on phonology in Chapter 7. Prominence is used to indicate the strength of a word, syllable or phrase when it is used in a sentence. We will use the term accent solely to indicate intonational phenomena associated with pitch, and emphasis to indicate a specific use of prominence in discourse. There are many different ways to formally describe prominence and stress, but for exposition purposes we will use the term prominent to indicate a unit has more prominence than normal, and reduced to indicate that it has less prominence than normal. 6.3.1 Syntactic prominence patterns Prominence is really a property of words, such that one word is given extra strength compared to its neighbours. This however is manifested in the lexically stressed syllable of that word receiving the extra strength, rather than all the syllables of the word. It is widely accepted that some, but not all, of the prominence pattern in a sentence is governed by the syntax within that sentence. A first type of effect is that of nuclear prominence3, which describes the particular promi- nence that often falls on the last content word in sentence. In the following examples, we have marked this word in bold: (68) THE MAN WALKED DOWN THE STREET. (69) TELL ME WHERE YOU ARE GOING TO. These examples are what are often termed discourse neutral renderings, indicating the “normal” or canonical way these sentences are phrases are spoken, in contrast to some of the ways in which they may be spoken in particular discourse contexts. Prominence patterns are particularly noticeable in sequences of adjectives and nouns which are collectively called compound noun phrases. For example (70) PARKING LOT (71) CITY HALL (72) THE UNION STRIKE COMMITTEE The final main type of syntactic prominence concerns function words. A very simple rule is that content words (nouns, verbs etc) can have normal or extra levels of prominence but function words are reduced. This certainly covers many simple cases so that in 3 This is often termed a nuclear accent in the literature.
- 142. 118 Chapter 6. Prosody Prediction from Text (73) THE MAN MADE THE OFFER ON THE BACK AN ENVELOPE. we feel that the two instances of THE, ON and AN, carry the least prominence and are reduced. Function words are not always reduced however, and in cases like (74) A BOOK LIKE THAT IS ALWAYS WORTH HOLDING ON TO we see that THAT and TO are not reduced, and in some situations the THAT would even be the most prominent word in the sentence. It can be argued that linguistic items such as THAT are really several distinct words that happen to carry the same form. This item can be used as a pronoun, an adjective a conjunction, or an adverb as in the following examples (75) THAT IS MY HOUSE (pronoun) (76) I WANTED THAT BOOK (adjective) (77) HE SAID THAT HE WAS AFRAID (conjunction) (78) HE DIDN’T TAKE IT THAT SERIOUSLY (adverb) where we see different degrees of prominence depending on use. Getting the prominence right in these situations may therefore not be so much an issue of prominence prediction per se, but rather of making the correct word definitions and finding these from the text. A second type of prominence effect occurs when function words occur in sequences. When we have sentences such as (79) I WAS WALKING THE DOG it seems quite natural to have more prominence on WALKING than on WAS, which follows our sim- ple content word prominent/function word unstressed rule. But when we have longer sequences such as (80) WHAT WOULD YOU HAVE DONE IF IT HADN’T BEEN FOR ME ? we see a sentence made entirely of function words. If these are all spoken in reduced form the speech sounds terrible, and so stresses must be placed on some words. A reasonable rendition of this with prominence might then be: (81) WHAT WOULD YOU HAVE DONE IF IT HADN’T BEEN FOR ME? though many other patterns are possible. In fact the overall degree of prominence in sentences like this is no less than in “normal” sentences where there is a mixture of content and function words. This situation can be likened to the situation in phrasing where speakers feel the need to place phrase breaks at regular intervals somewhat independently of the top down linguistic structure. Each type of prominence can be “overridden” by the next type, so while the normal promi- nence pattern for THIRTEEN is (82) THIRTEEN in many compound noun phrases, the word undergoes prominence shift and to give: (83) THIRTEEN BOYS
- 143. Section 6.3. Prominence 119 This is often explained by the notion that the main stress on BOYS “repels” the stress on -TEEN so that stressed syllables do not directly occur as neighbours. 6.3.2 Discourse prominence patterns When sentences are spoken in discourses or dialogues, additional patterns can occur. One com- monly cited affect is the difference in prominence levels between given and new information. This idea says that when a word is first used in a discourse, it will often be quite strong, but on second and subsequent mentions it is actually weaker, so we might have: (84) FIRST MAKE SURE THAT PAUL ANSWERS THE PHONE. (85) THEN TELL PAUL TO BRING THE BOOK OVER TONIGHT. in which the first mention of PAUL is prominent and the second is not. While this given/new distinction has traditionally received considerable attention in the literature, many recent empirical studies on real speech have shown that this affect is not so common as traditionally though [328], [329], [23], [330]. First, it is far more common that the second sentence would have a pronoun instead of the original name (86) THEN TELL HIM TO BRING THE BOOK OVER TONIGHT. and secondly, as Aylett [24] points out, even when words are repeated in full form in dialogues, there is very little evidence of prominence patterns being governed by the given/new distinction. What does occur, is that the degree of prominence is heavily influenced by the redundancy in the dialogue, such that if the speaker feels the listener has a high probability of understanding then he or she may well give a word less prominence. Crucially, the word receiving less prominence does not have to be the same as the original word, hence (87) THEN TELL THE IRISH GIT TO BRING THE BOOK OVER TONIGHT. here as IRISH GIT is obviously related to PAUL it may receive less prominence. Prominence patterns of these sort are clearly augmentative and will be explained further in Section 6.5.3. We can identify a second type of discourse prominence effect which we shall term emphasis. This occurs when a speaker wants to draw attention to a particular word: (88) IT WAS REALLY HUGE (89) I DIDN’T BELIEVE A WORD HE SAID (90) I’VE HEARD IT ALL BEFORE (91) THATS JUST TOO BAD (92) ITS IMPORTANT TO DO THAT BEFORE YOU ENTER THE ROOM (93) NOT THAT ONE and so on. In general, emphatic prominence is used for augmentation purposes (see below in Section 6.5.3) where the speaker wants to make particularly sure that the listener understands what is being said. A particularly salient example is contrastive prominence of the type:
- 144. 120 Chapter 6. Prosody Prediction from Text (94) THE PROFIT WAS $10M BEFORE TAX, NOT AFTER TAX This however also extends into affective purposes, where for example the speaker knows that the listener will understand, but still uses extra emphasis. This could occur for example when the speaker wishes to express frustration or imply that the listener is stupid and therefore has to have everything spoken slowly or clearly. 6.3.3 Prominence systems, data and labelling As with phrasing, we have a number of models and theories which attempt to explain prominence. Again as with phrasing, the theories range from simple models to ones which posit a rich and embedded prominence structure. Broadly speaking, the phenomenon of lexical stress is fairly uncontentious with good agreement between labellers as to where stress values lie. Prominence and emphasis is more difficult to mark, as labellers often find it hard to distinguish between what is “normal”, discourse neutral prominence and marked emphasis. The distinction here is related to the issue of augmentative prosody, discussed in Section 6.5.3. As with phrasing, the basic distinction is between systems which have a fixed number of prominence values or levels, and ones in which nesting and recursion can occur. And again, we see that the issue is the tradeoff between a system which seems to match linguistic structure, and one that is more bottom up and perceptually verifiable. We have already examined Liberman’s model of metrical phonology for phrasing, but actually this was primarily developed for explaining prominence patterns. In this, we have binary trees in which each node is marked either strong (S) or weak (W). Using this, we can then describe prominence patterns as follows: W city S Hall S James W street S parking W lot The “clever bit” is that this model allows us to elegantly combine smaller groups of words into bigger groups by further use of the metrical trees: W S James W street S S parking W lot in which case there is one and only one node that is dominated only by S, and this is how we specify the most prominent or nuclear prominence of the sentence or phrase. Metrical phonology
- 145. Section 6.3. Prominence 121 also neatly deals with the issue of prominence shift, which occurs in the following example for many speakers: W S city W hall S S parking W lot (the prominence has shifted from HALL to CITY). Just as with phrasing however, it seems difficult to believe that when it comes to a whole sentence such as: W W W I S left S W my S car S W in S W the S W S city W hall S S parking W lot ] that speakers can make and listeners discern all these different degrees of phrasing. This has lead to researchers proposing a fixed number of types or levels of prominence, but as with phrasing, there is virtually no agreement as to how many there should be. Probably the most common models for practical engineering purposes are: 1. a metrical tree in which the values on the high level nodes are ignored 2. a system with three levels of prominence (a) prominent/stressed (b) un-prominent/unstressed (c) reduced 3. as above, but with an extra level for emphatic prominence.
- 146. 122 Chapter 6. Prosody Prediction from Text Figure 6.1 Example of F0 contour with pitch accents and boundary tones marked. 6.4 INTONATION AND TUNE Intonation is sometimes used as a synonym for prosody, and as such includes prominence and phrasing as just described. Here we use a more narrow definition in which intonation is the sys- tematic use of pitch for communication. Often we describe intonation with reference to the pitch contour or the F0 contour, which is the pattern of pitch/F0 variation over an utterance4 Most traditional accounts of intonation speak of intonational tunes which span the length of the sentence. In these accounts, we might have a question tune, an exclamation tune and so on. More modern accounts advocate a more compositional approach, where the global intonation tune is composed of a number of discrete elements. Most important of these is the pitch accent, which can be defined as a significant excursion (either positive or negative) in the F0 contour that lends emphasis to a syllable. Pitch accents only occur on prominent syllables, and hence prominence can be seen as defining the possible places where pitch accents can occur. Figure 6.1 shows a sentence with two clearly visible pitch accents. In addition to this, it is common to have boundary accents or boundary tones which are pitch movements associated with the ends of phrases rather than syllables. A classic example of a boundary tone is where we having rising pitch at the end of a sentence to indicate a question, shown in Figure 6.1. Beyond this, we again find that there is little agreement about how to describe pitch accents and boundary tones. Some theories state there are a fixed inventory of these, while some describe them with continuous parameters. The nature of pitch accents and boundary tones is disputed, with some theories describing them as tones or levels while others say their characteristic property is pitch movement. One prominent theory states that we have an intonational phonology that parallels normal phonology, and as such we have inventories of contrasting units (sometimes called 4 F0 is shorthand for fundamental frequency. Informally this can be thought of as being the same as pitch, the exact relationship is described in Section 9.1.1.
- 147. Section 6.5. Prosodic Meaning and Function 123 tonemes) and grammars which state what can follow what. As with phrasing, we find there is a considerable degree of speaker choice with regard to intonation, such that while it may be common to speak a question such as (95) IS THIS THE WAY TO THE BANK? quite often speakers speak this with the same intonation as a normal statement. Just why speakers do this is addressed in Section 6.5.3. It is difficult to discuss intonation further without reference to the phonetics and acoustics of this part of language, and for that reason we forgo a more thorough account of intonation until Chapter 9. 6.5 PROSODIC MEANING AND FUNCTION One of the reasons why good models of prosody have proved so hard to develop is that researchers have often tried to study prosody without reference to its communicative function. In verbal linguistics it is quite common to study a particular sentence without reference to why a speaker said this, and by and large this separation provides a useful modularity for research purposes. So when we consider a sentence such as (96) THE MAN WALKED DOWN THE STREET We can say that THE MAN is the subject and WALKED is a verb and so on, and then perhaps go on to study the phonetics of the sentence, without having to know why the speaker actually said this sentence. It is common to follow a similar approach in prosody, but we believe this is mistaken, and that in fact, it is more or less impossible to study the nature of prosody without taking into account the dialogue, speaker situation and other functional factors. Hence we now give an account of prosodic function which states the reasons why we use prosody, and as we shall see, this significantly simplifies the practical issues concerning how to generate prosody in a text-to-speech system. 6.5.1 Affective Prosody Affective prosody is the most basic type of prosody and in a sense can be regarded as pure prosody. Affective prosody includes the expression of meaning related to emotion, mental state and speaker attitude. Significantly, prosody is the primary channel for the communication of these meanings: if one is angry and wishes to express this, it is best done with a loud, shouting, growling voice, rather than a simple, natural statement I AM ANGRY. In fact simply saying that one is in an emotional state with words without any matching prosody is usually taken as a sign of insincerity (e.g. YES YOU LOOK NICE IN THAT DRESS), and if the words and prosody seem to contradict, the listener usually takes prosody to be the true expression of the speaker’s meaning. (For example,
- 148. 124 Chapter 6. Prosody Prediction from Text if you ask a child whether they ate all the cookies, they find it easy to say NO but less hard to say this convincingly). Affective prosody has a certain degree of language universality. In all languages, raising one’s voice and shouting are seen as expressions of angry or hostile states, while speaking in a softer voice is seen as conciliatory or calming. In fact the relationship is so direct that often we say “she spoke in a calm voice” etc. Affective prosody is not completely universal however, and in terms of primary emotion, there are significant differences across languages. Even within English it is common to find that one group of speakers are regarded as brash and another meek, when in fact all that is different are the “settings” and “thresholds” with which the basic emotions are expressed. In general, affective prosody has a fairly direct meaning-to-form relationship and hence the more extreme the form the more extreme the emotion. This implies that, unlike verbal language, the affective prosody system is not arbitrary with regard to meaning/form. A number of emotion categorisation systems have been proposed. One common theory holds that there are six basic emotions (anger, happiness, sadness, disgust, fear and surprise) from which follow a number of secondary emotions such as puzzlement or intrigue [110], [113], [153]. Some however argue that there is no fixed set of emotions and rather emotion is best expressed in a multi-dimensional system. One such proposal advocates three independent axes of emotion called activation, evalu- ation and power [393], [126]. In addition to emotion, affective prosody includes speaker attitude. The difference between emotion and attitude is that only emotion can exist independent of communication. As an illustra- tion; you can be happy on your own, but you need someone else to be sarcastic with. Apart from this distinction, attitude functions similarly to emotion in that it is fairly direct, has a significant degree of universality and does not have a arbitrary or dual nature. One point of debate in the literature on both emotion and affective prosody concerns whether there is such a state as “no emotion” or “no affective prosody”. The argument is that if we are able to increase the level of intensity of anger, then we should be able to decrease this to the point where there is no anger. Similarly with the other emotions. This issue is important as it seems that the total affective content of an utterance can vary greatly and can be contrasted with the total propositional content of the verbal part of the utterance. These two effects seem to trade-off such that we can have sentences that are nearly purely affective “thats amazing” and others which are purely propositional “the sum of the square of the opposite and adjacent sides...”. This later type of sentence is what we have previously termed discourse neutral. 6.5.2 Suprasegmental If a sentence is spoke with little or no affective content, ie. in a discourse neutral manner, we still see characteristic patterns in the phrasing, rhythm, pitch, voice quality and timing. Typical effects include phones at the ends of sentences or phrases being lengthened, syntactically salient words (e.g. heads) having more emphasis, F0 levels being higher at the starts of sentences and so on.
- 149. Section 6.5. Prosodic Meaning and Function 125 These effects are often called “prosodic” but according to the model we use here, these are not real prosodic phenomena but just another aspect of the verbal component. We can see this because speakers do not impart extra information by using these effects and they are as highly formulaic as the use of say nasalization, voicing and tongue position. These effects are often called prosodic as they use much the same acoustic means as expression - anger is expressed through pitch and voice quality, which is also used to distinguish the middle and ends of a phrase. In the model used here, we do not include these effects under the umbrella of prosody, but instead term these the suprasegmental part of verbal phonetics. The term suprasegmental here is used to show that while these effects may be verbal in nature, they are operate over the span of several phones (segments). It is best to think of these as a by product in the production of the phonetic content. While suprasegmental effects are rightly seen as the natural effect of phonetic production on F0, timing and voice quality, it is important to realise that the speaker still has conscious control over these. That is, while it is normal for a particular voice quality to be used at the end of a sentence, the speaker can override this if desired. However, while overriding default behaviour is possible, it is highly noticeable if this is done, and so a speaker who deviates from default behaviour will be perceived as wanting to draw attention to the aspect of the utterance which has changed. Another way of looking at suprasegmental is in terms of defaults, and deviation from defaults. The key to this is to realise that the default for a sentence is not connected to mathematical simplicity where we might have a constant duration for each phone, a flat F0 contour and a constant state of vocal effort. Rather, the default behaviour exhibits considerable variation from these fixed values in ways dictated by verbal production. Deviations from the suprasegmental default are indications of real prosody, whether they be for affective or augmentative purposes (explained below). In tone languages like Mandarin, pitch can be used to identify words. For example, liu spoken with one tone pattern means “flow”, but when spoken with a different tone pattern means “six”. In our model, this is a purely suprasegmental effect and treated in just the same way as nasalization is used to distinguish words in English. Mandarin of course has intonation also, which is used for all the same affects as English. The fact that intonation and parts of word identity are expressed in the same acoustic variable, does of course complicate analysis and synthesis somewhat, but with modern synthesis techniques we can model both without difficulty. 6.5.3 Augmentative Prosody Augmentative prosody is used as an aid to ensure successful communication of the verbal com- ponent of a message. Unlike affective prosody, augmentative prosody does not contain or convey any extra information, it is merely a means for the speaker to ensure that a message is decoded and understood more clearly. Augmentative prosody can only be understood with reference to a dialogue model of the type presented in Section 2.3. There we saw that participants in a dialogue have a mental image of each other and are constantly balancing the desire to be efficient in com-
- 150. 126 Chapter 6. Prosody Prediction from Text munication (whereby they would use the least effort in generating their utterances) with a desire to be effective, whereby they wish to ensure the message is correctly decoded and understood. In an ideal communicative setting, where the participants know each other, can see each other and are able to converse without background noise or other distractions, it is nearly always the case that the verbal message can be communicated without need for any augmentative prosody. Whether any prosody is used therefore is purely a matter of the affective nature of the utterance. Successful communication is possible because the verbal channel is highly redundant, to the extent that the speaker should always be able to decode and understand the message. In other conditions we face a different situation, so for example in noisy conditions we might see that the speaker speaks with greater vocal effort and more slowly. Likewise we may see the same effect if the speakers don’t know each other so well, or in more polite situations in which the speaker wants to be seen to be making a clear effort to communicate effectively. While it may be fairly obvious that increased articulation effort will lead to a higher prob- ability of a phoneme string being correctly recognised, it is with other aspects of augmentative prosody that the situation becomes particularly interesting. For example, it is frequently asserted that yes/no questions such as (97) IS THIS THE WAY TO THE BANK? having rising intonation in English. In fact empirical studies of data show that the most common intonation for such sentences is exactly the same pattern used for statements. While speakers can use rising intonation, they often don’t. This behaviour is best explained with reference to the communicative model. In examples such as 97, it is usually obvious that a yes/no question is being asked, there really is little need for a speaker to use the additional cues that a rising intonation pattern would provide. However if the speaker believes she may be misunderstood, or if she is trying to sound polite, the rising intonation pattern may appear. By comparison, it is common to ask “yes/no questions” that are have the same syntax as statements, but in which the speaker does now in fact use a rising intonation pattern to make sure these are interpreted correctly: (98) THIS IS THE WAY TO THE BANK? We see this effect in all aspects of prosody. The given/new distinction, where it is often claimed that subsequent mentions of an entity are less prominent than first mentions, is rarely observed in real speech, partly because it is quite clear from the verbal content which in fact is intended [328], [24]. In classic example sentences such as (99) BILL DOESN’T DRINK BECAUSE HE’S UNHAPPY which are spoken without phrase breaks may potentially be ambiguous but in general the listener will know enough of the dialogue context to correctly decode this. If however the speaker is less sure of a correct decoding, phrase breaks may be inserted to aid communication, and likewise with every other aspect of prosody.
- 151. Section 6.5. Prosodic Meaning and Function 127 In summary, augmentative prosody works by changing the default suprasegmental content of an utterance in particular ways. It is used by the speaker solely to increase the chances of the verbal message being correctly decoded and understood. Significantly, unlike affective prosody, it imparts no extra information into the message. 6.5.4 Symbolic communication and prosodic style It is useful at this stage to compare and contrast the verbal and prosodic components of language. Verbal language was defined in Section 2.2.1, and there we saw its defining properties as being discrete and dual, in the sense that every message is composed of a sequence of discrete word entities which in turn are composed of a sequence of phonemes. Verbal language is arbitrary in the sense that one cannot infer the form of a word from its meaning or vice-versa. Verbal language is productive in that completely new sentences can be created all the time. What if any of these properties does prosody have? First of all, let us realise that in verbal language, to construct a message we must always select a unit (e.g. word) and place these in sequence. A word is either present or absent in an utterance and every utterance is composed of a sequence of words. With prosody the situation is different. Firstly, we need not always have a prosodic unit - it is completely acceptable and common to have an entire sentence with no particular emphasis, intonation or phrasing effects (recall, suprasegmental effects are considered part of verbal language, not prosody). Furthermore, if a prosodic effect is present, it can be there to a greater or lesser extent. So we can have a word that is somewhat emphasised to one which is heavily emphasised - this is unlike verbal language where something is either present or not. In verbal language, speakers really have very little choice in how they speak a particular word. So while speakers when saying the word CAT may vary somewhat in their tongue positions or degree of aspiration, only so much variation is allowed, otherwise a different word may be per- ceived. Hence verbal language is inherently restricted and formulaic. With prosody, the fact that an utterance is not required to be “full” of prosody gives the speaker much more choice. Augmen- tative prosody may be used to aid communication but often it doesn’t have to be; conversely in certain polite or formal situations a speaker may choose to use considerably more augmentative prosody than is strictly required for successful communication. The degree of affective prosody is again very variable - some conversations or messages (e.g. speech from a news reader) are nearly completely lacking in affective content while some are full (e.g. a chat between two friends). An important consequence of the degree of choice in prosody is that it allows the speaker to deviate from the default patterns in certain ways without running the risk of changing the meaning of the sentence entirely. This then allows for the phenomenon termed prosodic style. This is most marked in situations where particular speakers, often actors or comedians, use prosodic style in a quite unconventional way. This use of prosody is not related to the communication of the message or even emotion, but is intended to draw attention to the speaker himself. Informally, we can talk about the speaker “sounding cool” by the way they can deviate from the normal patterns
- 152. 128 Chapter 6. Prosody Prediction from Text of prosody and be creative in their usage. Stylistic use of prosodic is one significant factor in explaining why some speakers seem more interesting to listen too than others. 6.6 DETERMINING PROSODY FROM THE TEXT 6.6.1 Prosody and human reading In text-to-speech, our interest is of course to generate prosody from text. This is problematic in that text mostly encodes the verbal component of language; prosody is by and large ignored. Given that prosody is such a vital part of spoken language, how then is it possible that written communication can even work? We examined this in some detail in Section 3.2, and can summarise the situation as follows: • Quite often text encodes sentences which are rich in propositional content rather than af- fective content. Text genres like newspapers, non-fiction books, histories, company reports, scientific articles and so on simply do not have much affective content in the first place. • The writer knows that prosody is absent from the text that is being written, and therefore compensates for this by using words in different ways that would be the case when speaking. For example in literature, we see phrases such as: (100) ‘thats right’, he said angrily where the he said angrily is obviously written to help the reader understand how the ‘thats right’ section should have sounded. • The awareness that prosody is lacking often leads to the writer writing in a style that is inherently less ambiguous; in such cases we say the writer is writing “clearly” and so on. • Quite often, people who are reading aloud add their own prosody into the speech. By this we mean that when they read the words, they keep accurately to the word, but as the prosody is not encoded in the text the reader “makes up” what they think is appropriate for this situation. It is if the reader is saying “if I were saying this myself, what prosody would I use?” and then adding this into the speech. • This requires considerable skill and is often based on an understanding of the text, rather than just a decoding as is the case for words. The reader is in effect adding in affective and augmentative prosody in places she believes are appropriate. From this we can conclude that in situations where the text genre is quite factual, it is usually sufficient to generate speech from the verbal message only, and so all that is required is the gen- eration of the suprasegmental part of the signal; the affective part of prosody is ignored. In other text genres the situation is significantly more difficult, and if say a dialogue from a play is read in this fashion (or more likely) responses from a computer dialogue system, the generated speech can sound somewhat dull and removed. Finally, we see that mimicking a genuinely good human reader is very difficult indeed, as they will be performing an actual comprehension of the text, and
- 153. Section 6.6. Determining Prosody from the Text 129 then generating their own prosody. In no way are they simply decoding the prosody from the text and then speaking it aloud, as they can do with the words. To date, no satisfactory solution has been found to this problem, and so current text-to-speech systems are, unfortunately, lacking in this regard. 6.6.2 Controlling the degree of augmentative prosody We face a different situation with regard to augmentative prosody. When a person speaks, she uses a mental image of the likelihood of a listener decoding and understanding the message to decide how much augmentative prosody to use and where to use it. We therefore face two problems in TTS: guessing the state of the listener and knowing which parts of the utterance may need disambiguating with prosody. Doing this properly is very difficult - even if we have a good idea about the likelihood of a message being misunderstood, knowing exactly where to place phrase breaks and additional emphasis is extremely hard as often this can require a full and accurate syntactic and sometimes semantic analysis of the sentence. But luckily, as with affective prosody, we have a “get-out” which enables us in many cases to bypass this problem. Firstly, writers often use punctuation and fonts when they feel that the words in some way need disambiguation. So when we see a comma or semi-colon, it is often a safe bet to place a phrase boundary here as we know the author felt that the words before and after the break should be separated. If a word is in italics or bold, it is again probably a safe bet to give this word ex- tra emphasis. A second strategy arises from the fact that augmentative prosody is indeed only there to help in the decoding and understanding of an utterance. If we simply don’t include any augmentative prosody, there is a good chance the utterance will still be decoded and understood, but perhaps just not with so much certainty. We can compensate for this specific lack of disam- biguation information by ensuring that other parts of the signal do aid decoding; for instance if we synthesise the speech more clearly and slowly than normal the listener will have a better chance of understanding. 6.6.3 Prosody and synthesis techniques In the remainder of this chapter we shall concentrate on prosody prediction, the term given to the task of generating the prosodic form from the text. In Chapter 9 we complete the task by consider- ing the issue of prosody synthesis, the job of generating phonetic or acoustic information from the prosodic form. As we shall see, the degree to which prosody needs to be specified depends very much on the synthesis technique employed in waveform synthesis. Roughly speaking, in modern systems suprasegmental prosody can be generated at the same time as the verbal content, whereas in first and second generation synthesis systems, the suprasegmental content has to be generated explicitly. Most of the following sections can be read without too much emphasis being placed on this distinction, it will however become more of an issue in Chapter 9.
- 154. 130 Chapter 6. Prosody Prediction from Text 6.7 PHRASING PREDICTION First we will examine the issue of phrasing prediction which is how to generate a prosodic phrase structure for a sentence from the text. For purposes of illustration we will adopt the model described in Section 6.2.2 where we have a major phrase and a minor phrase and this is probably the most commonly used in TTS. 6.7.1 Experimental formulation To help in our exposition, we will formalise the problem in the manner suggested by Taylor and Black [443]. In this, our input is a list of tokens: T = t1,t2,t3,...,tn, and between every token is a juncture, so that a full sentence is: T = t1, j1,t2, j2,t3, j3,..., jn−1,wn, The job of the phrasing algorithm is to assign values to every ji drawn from an inventory of juncture types, which we will define to be { non-break, minor-break, major-break, sentence}. About 1 in 5 junctures are of type non-break (that is the gap between normal words) and so we have to be careful in assessing the accuracy of any algorithm - if we use one which always assigns non-break we will have achieved 80% accuracy on a simple junctures correct measure. Because of this, Taylor and Black proposed a number of measures, later extended by Busser et al [77]. These are calculated from the following basic quantities: Breaks B number of breaks in the corpus Insertions I a non-break has been incorrectly marked as a break Deletion D a break has been incorrectly marked as a non-break Substitution S a break of one type has been incorrectly marked as a break of another type Number of junctures N total tokens in the corpus -1 These are then used to calculate the following scores:
- 155. Section 6.7. Phrasing prediction 131 Breaks correct = BC = B−D−S B ×100% (6.1) Non-breaks correct = NC = N −I −S N ×100% (6.2) Junctures correct = JC = N −D−I −S N ×100% (6.3) Juncture insertions = JI = I N ×100% (6.4) Juncture deletions = BI = I B ×100% (6.5) (6.6) An alternative scheme, more in line with that used in natural language processing is based on precision and recall, which are specifically designed to give accuracy figures for problems where the number of occurrences is much less than the number of non-occurrences. Precision = P = number of breaks correct number of breaks predicted = BC −I BC (6.7) Recall = P = number of breaks correct number of breaks in the test set = BC −I B (6.8) These are often combined in a figure of merit, F: Fβ = (β2 +1)PR β2P+R (6.9) where β is used to govern the relative importance of whether it is a worse error to falsely insert a break or falsely miss one. If these are considered equal, then β = 1 and F = 2PR P+R (6.10) This is a commonly used measure of overall quality, but as we shall see, it is almost certainly not true that human perception believes an insertion and deletion cause the same gravity of error. 6.7.2 Deterministic approaches There are two popular and very simple phrasing algorithms: 1. deterministic punctuation (DP): Place a phrase break at every punctuation mark
- 156. 132 Chapter 6. Prosody Prediction from Text 2. deterministic content, function (DCF): Place a phrase break every time a function word follows a content word. The first algorithm, DP, works very well in sentences of the type: (101) The message is clear: if you want to go, you have to go now For DP algorithm to work, we must have an accurate means of determining underlying punctua- tion. This is so that sentences like (102) But really!? You can go -- but it won’t do you any good. have the punctuation sequence !? and --mapped to a single underlying form because we don’t want to have an additional phrase break between the ! and the ?. The DCF algorithm works well in sentences such as (103) It was taking a long time to gather all the information where we place a phrase break after every function word that follows a content word: (104) it was taking | a long time | to gather |all the information These can be combined into a single algorithm, deterministic content function punctuation (DCFP), which places a break after every content word that precedes a function word and also at punctuation. The algorithms were tested on a large corpus of read speech which had phrase breaks labelled by hand. The results for these algorithms given in Taylor and Black [443] are: Algorithm breaks correct junctures correct juncture insertions Deterministic Punctuation 54.274 90.758 0.052 Deterministic Content Function Punctuation 84.40 70.288 31.728 These figures basically show that that the DP algorithm massively under predicts and the DCFP algorithm massively over predicts. The DP algorithm result is worth examining further because it shows that placing a phrase break at a punctuation mark is in fact a very safe thing to do; this nearly always coincides with a real break. The only problem is that many other breaks are missed. The DPCF algorithm over predicts far too much, which lead to the idea that if finer distinctions could be made than just content word and function word, a more accurate algorithm would be possible. A number of more sophisticated deterministic systems have been proposed [26], [186], [10] which make of rules for specific cases. For example, the verb balancing rule of Bachenko and Fitzpatrick [26] works through a sentence left to right and compares the number of words in a potential phrase formed with the verb and the syntactic constituents to the left, and the number of words in the constituent to the right. The potential phrase with the shortest number of words is chosen as the correct one.
- 157. Section 6.7. Phrasing prediction 133 6.7.3 Classifier approaches The general historical trend in phrase break prediction is the same as with most problems we ex- amine: as more data has become available, researchers have moved away from rule based systems towards trainable systems. This is partly because the new data showed how errorful rule based systems were, and partly because it allowed algorithms to have their parameters and rules auto- matically inferred from the data. We now turn to data-driven techniques, where interestingly we see a similar pattern of algorithm type to that used in text classification. In this section we examine local classifier algorithms, which share many features in common with WSD algorithms. In the next section we examine HMM approaches, which are very similar to POS taggers. Wang and Hirschberg [363] introduced the idea of using decision trees for phrase break prediction. Decision trees allow a wide variety of heterogeneous features, examples of which are given below: • total seconds in the utterance • total words in the utterance • speaking rate • time from start of sentence to current word • time to end of sentence from current word • is the word before the juncture accented? • is the word after the juncture accented? • Part-of-speech • syntactic category of constituent immediately dominating the word before the juncture • syntactic category of constituent immediately dominating the word after the juncture These features (and others) were used as question in the tree, which was grown in the normal way. Wang and Hirschberg trained and tested their system on 298 utterances from the ATIS corpus, giving the following results (the results aren’t available in the formats we described above): Boundaries correct 88.5% Non-Boundaries correct 93.5% It should be pointed out that this approach can’t strictly be used for TTS purposes as acoustic features (e.g. time in seconds) measured from the corpus waveforms were used in addition to features that would be available at run time. Following this initial work, a number of studies have used decision trees [264] [418], and a wide variety of other machine learning algorithms have been applied to the problem including memory based learning [77] [402], Bayesian classifiers [516], support vector machines [87] and neural networks [157]. Similar results are reported in most cases, and it seems that the most important factors in the success of a system are the features used and the quality and quantity of data rather than the particular machine learning algorithm used.
- 158. 134 Chapter 6. Prosody Prediction from Text 0 50 100 150 200 250 0 2 4 6 8 10 12 14 16 18 20 Figure 6.2 Histogram of phrase lengths. 6.7.4 HMM approaches The HMM approach to phrase-break prediction was introduced by Taylor and Black [443] and has been extended by others [37], [369]. In their basic system, Taylor and Black proposed that the decision on whether a phrase break should or should not be placed was dependent on two main factors: 1. The parts-of-speech surrounding the juncture 2. The distance between this proposed break and other breaks The first of these factors is simply a restatement of what we have already examined; i.e. that syntactic patterns have a strong influence on phrase break position. The second factor is used in the classification approaches, but is given particular importance in the HMM model. The hypothesis is that in addition to any high level syntactic, semantic or dialogue factors, phrasing operates in a semi-autonomous fashion in that speakers simply like to place phrase breaks at regular intervals. This can be explained to some extent by the fact that speakers need to pause to get breath, but it seems that the phrase breaks occur much more frequently than is needed for breathing purposes and so this alone can not explain the patterns. Regardless of why however, it is clearly the case that speakers do exhibit strong tendencies to place phrase breaks at regular intervals, a fact noted by many researchers in this area [26], [156]. Figure 6.2, which shows a histogram of phrase lengths in words, clearly demonstrates this pattern.
- 159. Section 6.7. Phrasing prediction 135 B N N B 0.2 0.8 0.8 0.2 N B NB NN BB 0.97 0.03 0.24 0.76 0.88 0.03 0.03 0.71 0.12 0.97 0.29 0.97 Figure 6.3 Finite state machine depictions of n-gram phrasing models. The phrase break length histogram shows a preference for phrase lengths of about 3 - 6 words, and shows that very long or very short phrases are highly unlikely. Alternatively, we can view this distribution cumulatively, meaning that the longer we have been since the previous phrase
- 160. 136 Chapter 6. Prosody Prediction from Text break the more likely it is that the next juncture will have a phrase break. While it is easy to add this sort of information into a classifier, it is difficult to do so in an globally optimal way. The basic problem is, that say we are at a juncture that is six words away from the previous juncture. Our distribution may say that it is now quite likely that we will observe a break at this point. If the evidence from the POS sequence seems reasonable, then a break will be placed here. The problem is that the next juncture, seven words away from the last break may be an even better place to put a break, but we can’t do this now as the break at word six has already been assigned. What is required is a way to search all possible sequences of phrase breaks, and this, via the Viterbi algorithm, is exactly what the HMM approach can deliver. In the HMM approach, we have a state for each type of juncture, so ignoring sentence breaks, we either have three states if we distinguish major and minor breaks or two states if we don’t. Taking the second case first, we now have a state which gives: P(C| ji) that is, the probability of a juncture after token i omitting a POS sequence L. Ideally the POS sequence L would be the tag sequence for the entire sentence, but as these distributions are non- parametric models which have to be learned by counting, the size of the window is quite limited in terms of what can be estimated robustly. Hence we approximate the POS sequence of the whole sentence as a window of POS tags around the juncture. This can be seen as a generalisation to the DCFP algorithm where we used one tag before and after the juncture. Taylor and Black however reported that a window of two tags before and one tag after the juncture gave the best results given limited training data. Informally, we can see that we would expect the following to give relatively high probabilities P(ADJ NOUN AUX|break) P(DET NOUN NOUN|non-break) The juncture sequence model is given by an n-gram, which is calculated in the way explained in Section 5.2.2, that is by counting occurrences of sequences. An important difference to the n-gram in a POS tagger is that we now only have two or three types in the n-gram rather than 40 or so. This means that the number of unique sequences is much less, which means that for the same amount of data we can robustly compute longer n-grams. In these models we would see that P(N,N,N,B,N) would give a high probability and that P(B,B,N,B) would give a low probability. Taylor and Black report that the best results are found from which uses a 6-gram language model, where there are two tags before the juncture and one after. This gave figures of 79.24% breaks correct, 91.597% junctures correct and 5.569% juncture insertions.
- 161. Section 6.8. Prominence Prediction 137 6.7.5 Hybrid approaches More recently a number of approaches have been proposed which combine the advantages of decision tree approaches (use of heterogeneous features, robustness to curse of dimensionality) and the HMM approach (statistical, global optimal search of sequences). In addition, there has been somewhat of a re-awakening of use of syntactic features due to the provision of more robust parsers. Rather than attempt an explicit model of prosodic phrasing based on trying to map from the syntax tree, most of these approaches use the syntax information as additional features in a classifier [508], [209], [257]. While the pure HMM approach uses a discrete probability distribution over observations (in our case POS tags), in practice, all we need is a means of finding the likelihood of seeing the observation given the juncture type P(C|J). Many have therefore tried alternative ways of generating this. One of the most popular is actually to use decision trees, as these can be used to given probabilities as wall as simple decisions on classification. As it stands, a decision tree gives a posterior probability, not a likelihood, but a likelihood value can be obtained by either ensuring that the tree is trained on an equal amount of break and non-break data, or by normalizing the posterior by the basic frequency of breaks and non-breaks. The advantage in this combined approach is that we can use any features that require, but also perform the global Viterbi search using n-grams to find the best overall sequence of breaks [403], [369], [423]. 6.8 PROMINENCE PREDICTION Prominence prediction algorithms follow the same basic approaches as phrase break prediction where we have simple deterministic algorithms, sophisticated deterministic algorithms and data driven algorithms. 6.8.1 Compound noun phrases The simplest prominence prediction algorithm simply uses the concatenation of the lexical promi- nence patterns of the words. So for: (105) IN TERMS OF THE NECESSARY POLITICAL EXPEDIENCY TO ENSURE SURVIVAL — THE RESULT WAS A CLEARLY A GOOD ONE FOR SHAW we might generate a prominence pattern of: (106) IN TERMS OF THE NECESSARY POLITICAL EXPEDIENCY TO ENSURE SURVIVAL THE RESULT WAS A CLEARLY A GOOD ONE FOR SHAW The main ways in which this rule fails (again ignoring affective and augmentative effects) are in compound noun phrases and certain function words uses. The prominence patterns in com- pound noun phrases in English are complex and can seem quite baffling. When we consider the prominence patterns in street names such as
- 162. 138 Chapter 6. Prosody Prediction from Text (107) WINDSOR AVENUE (108) BEAUFORT DRIVE (109) PARK LANE it seems quite obvious what the pattern is, until we consider examples with the word STREET (110) WINDSOR STREET (111) BEAUFORT STREET (112) PARK STREET in which the prominence pattern is reversed. Sproat [410] conducted a very thorough study into compound noun phrasing patterns, and concluded that in general, there is no simple syntactic pattern which governs prominence. Instead he proposed an extensive set of rules based partly on synactic and partly on semantic patterns. The most basic rule says that compound noun phrases with two words assign their prominence according to whether the two words are taken as a single compound “word” or a phrase. In general in English, phrases receive their main prominence on the last word in the phrase. English words, and particularly nouns, by contrast tend to have their prominence on the first syllable. This can be seen in the difference between (113) BLACKBIRD meaning the type of bird (114) BLACKBIRD a bird which just happens to be black This principle can then be generalised to cope with all two word cases, with the idea being that only conventionalised cases are treated as compound words which receive their prominence on the first word. In addition to this, there are a number of semantically driven rules , which range from the general to the more specific: (115) Furniture + Room → RIGHT, e.g. kitchen table (116) Proper-name + street → LEFT, e.g. Park street For more compound noun phrases involving three or more words, a CYK parser (see Section 5.5) is used to assign the internal structure to the noun phrase, after which prominence values can be assigned. To resolve ambiguity, a set of heuristics is used to find the best parse. These are: 1. Prefer parses determined by semantic rather than syntactic means. (This partly works be- cause semantic parses are much more specific) 2. Choose phrase based parses over compound word bases parses 3. Choose parses with right branching structure. Additional rules are used to shift prominence in cases where the parses would produce two prominent syllables side by side.We also find that many types of non-natural language entities have characteristic patterns. For example, it is common place to find telephone numbers read with a particular stress pattern (and indeed phrasing). Producing the correct stress patterns for these is often fairly straight forward and the creation of these patterns can be integrated into the verbalisation process.
- 163. Section 6.8. Prominence Prediction 139 6.8.2 Function word prominence Less work as been carried out on the second main problem of determining the prominence patterns in function word sequences. The most we can say about this is that not all function words are equal when it comes to propensity for reduction. At one extreme, we have determiners (THE and A) which in their default form are always reduced, and at the other extreme closed class words such as HOWEVER, WELL and SO can receive the most prominence in a sentence. Given a sequence of content words, a reasonable approach is then to give each a score according to where it comes in the strength list, and starting with the strongest, add prominence at regular intervals. A possible hierarchy, from weakest to strongest, might be: articles THE, A, AN auxiliary verbs WAS, HAS, IS prepositions TO, FROM, BY particles IF, THEN, WELL conjunctions THAT, AND, BUT pronouns HE, HIM, SHE, ME, THEM modals WOULD, COULD, SHOULD Note that auxiliary verbs (e.g. HAVE) can function as main verbs also. We find a fairly clear pattern that when used as an auxiliary the word is often reduced, when used as a main verb it is often unreduced or prominent. With this table, we can then use a rule which states that prominent syllables should not occur side by side, and so in a sequence of function words we first assign prominence to the strongest word given in the table, ensure its neighbours and reduced and then look for the next strongest word. This then can give the stress pattern for the sentence we saw in Section 6.3.1: (117) WHAT WOULD YOU HAVE DONE IF IT HADN’T BEEN FOR ME? 6.8.3 Data driven approaches Prominence prediction by deterministic means is actually one of the most successful uses of non- statistical methods in speech synthesis. This can be attributed to a number of factors, for example the fact that the rules often don’t interact or the fact that many of the rules are base on semantic features (such that even if we did use a data driven technique we would still have to come up with the semantic taxonomy by hand). Sproat notes [410] that statistical approaches have had only limited success as the issue (especially in compound noun phrases) is really one of breadth and not modelling; regardless of how the prominence algorithm actually works, what it requires is a broad and exhaustive list of examples of compound nouns. Few complex generalisations are present (what machine learning algorithms are good at) and once presented with an example, the rules are not difficult to write by hand. That said, with the provision of larger, labelled corpora, the natural progression has still been towards using machine learning techniques to predict prominence. We should probably not
- 164. 140 Chapter 6. Prosody Prediction from Text be surprised to find that these techniques follow the basic paradigm used for phrase break predic- tion, where we obtain a corpus of speech, label the prosodic content of this by hand (in this case prominence levels) and train an algorithm to predict these values from features available to the system at run time. Approaches tried have included decision tress [207], [378], [147], memory based learning [303], transformation based learning [200] and neural networks [72], [323], [389]. 6.9 INTONATIONAL TUNE PREDICTION Here we have explicitly separated prominence from the tune part of intonation, and so for our pur- poses intonation prediction is specifically the prediction of intonational tune from text, rather than the broader definition of this problem that encompass predicting tune, accentuation and sometimes phrasing. Of all the prosodic phenomena we have examined intonational tune is the most heavily re- lated to augmentative and particularly affective content. In situations where these effects are ab- sent, we can say to a first approximation that all utterances have in fact the same intonational tune; the only differences occur as to where the pitch accents and boundary tones which make up this tune are positioned. Hence we can almost argue that for discourse neutral synthesis, there simply isn’t any intonational tune prediction to be done. In other words, the real task is to predict a suitable F0 contour that expresses the prominence and phrasing patterns and encodes the suprasegmental, rather than true prosodic patterns of the utterance. While we can describe prominence and phrasing in quite abstract high level terms, this is significantly harder with intonation as all theories to a greater or lesser extent make explicit refer- ences to F0 patterns, levels and dynamics. Given this, and the fact that for the most part we are generating discourse neutral suprasegmental intonation, we will leave the entire topic of intonation until Chapter 9. 6.10 DISCUSSION At first glance it may seem that predicting prosodic form from text is an impossible task. Two significant barriers stand in our way; first, the text is greatly under specified for the type of in- formation we require as textual encoding of prosody is more or less non-existent. Secondly, the uncertainty surrounding what prosodic form should be makes it hard to label data, train algorithms and basically even know if we are headed in the right direction. 6.10.1 Labelling schemes and labelling accuracy Taking the second issue first we see that in every area of prosody there is considerable disagree- ment as to how to represent prosodic form. The problem isn’t so much that researchers disagree
- 165. Section 6.10. Discussion 141 completely; there is wide spread agreement about how to label many utterances; in many cases it is quite clear that a particular utterance has a strongly prominent syllable, has a clear phrase break and has a distinctive intonational tune. There are many “black and white” cases where all theories and all labellers agree. The problem is rather that there is a significant grey area, where we can’t tell if this really sounds prominent, where we aren’t sure that there is a phrase break between two words and where we can’t really decide on the intonational tune for an utterance. There are of course grey areas in other aspects of linguistics. As we saw in Section 4.2, pin- ning down the exact definition of a word can be tricky, and we will come across similar difficulties in Section 7.3.2 when we consider the definition of the phoneme. But in general the agreement among researchers is much much higher with regard to these phenomena. Furthermore, so long as we are aware of the difficulties, the definitions of word and phoneme which we use in this book are good enough, meaning that, as engineers, using these models does not seem to result in signif- icant loss of quality in synthesis. Studies that have examined inter-labeller reliability of prosodic schemes have shown the true extent of the problem; for instance in the original publication of the ToBI system, Silverman et al reported that the agreement on phrase breaks was only 69% for four labellers, and 86% for whether a syllable is prominent (bears a pitch accent in their paper). These figures might not seem too bad, but are about an order of magnitude worse than the equivalent results for verbal, phonetic transcription, and up to two or three orders of magnitude worse for the transcription used for text analysis in Chapter 5. Furthermore, when we consider that the number of choices to be made in prosodic labelling is often small (say choose one break index value from 5) we see just how difficult labellers find this task. These results for ToBI are not particularly un- usual, consistently similar figures have been reported for many other schemes [426], [488], [487]. Furthermore it is misleading to blame poor labelling agreement on the labellers not being “expert” enough - non-experts can readily label many of the types of data we require in TTS, and so in comparing the figures for other types of labelling we are in fact comparing like with like. Why does these problems arise?. Essentially we face two related issues;: which labelling scheme, model or theory to use; and how to assign the labels arising from this to a corpus of speech. Most researchers involved in both the scientific investigation and engineering implemen- tation are aware that there are a huge number of differing and often incompatible theories and mod- els of prosody. The temptation then is therefore to seek a “theory-neutral”, “common-standard” or “engineering” system which we can use to label the data and so avoid having to nail our colours to any one particular theoretical mast. We have to emphasise at this point that such an approach is folly; there is simply no such thing as a theory neutral scheme. If we take the break index scheme for instance, this explicitly states that there are a fixed number of types of breaks, that no recursion or limited recursion exists, that phrasing is not explicitly linked to prominence patterns and so on. Furthermore, there is no sense in which we can just “label what we hear” in the data; all labelling schemes are based on a model, the fundamentals of that model are implied when we label speech, and if our scheme is wrong, inaccurate or at fault in some other way, we run a high risk of enforcing inappropriate labels on our data. If we then use these labels as ground truth, we are running a severe risk of enforcing our training and run-time algorithms to make choices that
- 166. 142 Chapter 6. Prosody Prediction from Text are often meaningless. 6.10.2 Linguistic theories and prosody Given these difficulties with labelling schemes and models, it is worth asking whether we need a model of prosodic form at all; perhaps it would be possible to simply define a mapping from the text or semantics to the acoustics. This viewpoint certainly has some merit, but we should be wary of dismissing the idea of prosodic form entirely. A brief digression into the theoretical history of prosody may help explain why. It is probably fair to say that the field of linguistics has always had a troubled relationship with prosody. In fact many accounts of language completely ignore prosody (e.g. prosody is hardly ever mentioned in any work by Chomsky): whether this is because it is deemed to be quite separate from verbal language or simply too tricky to handle is hard to say. If we however want a full account of language and communication we have to include prosody. Older linguistic accounts of prosody tended to treat it as quite a different system, and this to a degree was unproblematic. But with the rise of more formal theories of language, new theories of prosody were put forward that used the new techniques of formal language theory, feature structures and so on. The problem facing these researchers was how closely should a formal theory of prosody follow the formal theories of verbal language. Unfortunately (in my opinion) too much emphasis was placed on an insistence that prosodic phenomena could be explained with the same mechanisms for verbal language, and hence we saw grammars of intonation, intonational phonology and schemes for pitch accent description which were meant to be the direct equivalent of normal phonology. To be fair, external pressures pushed research in this direction; there was a strong desire to move from the very impressionistic and informal models of the past towards a more rigorous model and in addition there was the idea that if prosody could be explained with the same mechanisms as verbal language this would somehow be more parsimonious and would also ensure prosody was taken seriously as a linguistic area of study (instead of just being ignored as mentioned above). Ladd [269] points out that a major justification for the development of a phonology for prosody was to counter attempts in linguistics to correlate say F0 values with syntax patterns. While correlations may exist, building such a model ignores the fact that different speakers have different “parameters” in their F0 encoding model, for example that high pitched voices will pro- duce different patterns to low pitched ones. A better model is to posit a system of abstract prosodic form, where all speakers of an language/accent would share the same representations for similar semantic and functional purposes. The actual acoustic encodings of these might vary with speaker physiology, but the prosodic form would be the same for all speakers. In formal terms, all we are saying is that there are a number of levels, or interacting systems of prosody. In summary the phonological school of prosody is nearly certainly right in advocating abstract description systems for prosodic form, but is perhaps less right in advocating that these should be formulated with the same mechanisms as verbal language. Prosody has an abstract structure, but it operates in a
- 167. Section 6.10. Discussion 143 fundamentally different way to verbal language. 6.10.3 Synthesising suprasegmental and true prosody Turning to the problem of predicting prosody from under-specified text, we find the situation is less bleak that it might appear. Let us consider the issue of prosody not being encoded in the text. In many applications, we are not synthesising text with any particular emotive or attitudinal content, and so are therefore going for a discourse neutral reading. This simple fact saves us, and in such cases allows us to ignore the affective side of prosody. Many non-TTS synthesis systems are of course required to synthesize this type of text, and if emotive or attitude information is provided in the input there is no reason that these effects can not be successfully synthesized. When we consider augmentative prosody, we can again find ways around the problem. Firstly, as we discussed in Section 3.1.4, authors do often use punctuation and sometimes use italics and so on, and these are very good indicators that augmentative disambiguation is appro- priate at this point. As we saw from Table 104, placing a phrase break at a punctuation mark is rarely wrong; human readers nearly always do this. However, recall that the whole point of augmentative prosody is that it acts as an aid to decoding and understanding and in many cases isn’t strictly necessary. Hence if we under predict phrase breaks, the only risk is that we may increase the chance of mis-comprehension by the listener. Significantly, speech generated with less augmentative prosody may sound completely natural - its just the case that the listener may have to pay more attention. It is also important to realise that the most crucial phrase breaks are often just those marked by commas in the text as it is here that the authors wants to be most cure of correct comprehension. While not ideal, under prediction is a safe strategy. Conversely how- ever, inserting a phrase break in a completely inappropriate position can sound terrible and often completely throws the listener, resulting in loss of naturalness and intelligibility. Hence when de- veloping algorithms for phrase break prediction it is probably not sensible to balance insertions and deletions as equally bad errors; insertions are in general far worse than deletions. In practical terms, this means that the figure of merit scores which combine precision and recall should be weighted strongly in favour of precision rather than recall. One area where augmentative prosody does however hurt us is in the data. While we can bypass the augmentative prosody problem at run time by under predicting, when we come to consider real data we effectively a priori have no idea what augmentative prosody the speaker will choose to use. While we can measure the acoustic encoding of this (pause lengths and positions and so on), this does not help us in determining the high level features required to predict these. We can imagine the speaker “inventing” these features and using them when speaking; they are not available to us. This to a large extent explains the widely reported phenomenon of “speaker choice” or “speaker variability” in prosody. Basically, when speaking or reading, because of the redundancy in the verbal component, the speaker has considerable choice where to place the phrase breaks; it all depends on their mental image of the listener. For practical purposes this has the effect of adding noise into the data; the features we have access to (the ones we can derive
- 168. 144 Chapter 6. Prosody Prediction from Text from the text) explain the presence of phrasing to a certain degree. The other features explain the rest of the behaviour, but as we don’t have access to these, they effectively appear as noise in our data and variance in our models. The consequence of this, taken in conjunction wit the issue of very low level of transcribed agreement, may well mean that many prediction algorithms are as accurate as theoretically possi- ble; no amount of extra refinement in the techniques or extraction of alternative features will make much difference. The reason why the algorithms can not improve is due to noise in the data, from either or both or these sources. 6.10.4 Prosody in real dialogues In Sections 6.8 and 6.7 we have seen how to predict prominence and phrasing patterns from text. The techniques described though in general only operate on single sentences, which are usually declarative in nature. Much of interest in prosody takes part at a different level, where many of the main phenomena occur in conversational, discourse, emotive or spontaneous communication. Hence most TTS systems only model a small range of the possible prosodic phenomena. We can however imagine systems which go beyond current capabilities. If speech synthesis were used in more conversational applications, the discourse effects of prosody it would become important to model the discourse effects of prosody. In many current dialogue systems, the interaction feels very artificial because the system is only capable of simple questions and declarative sentences. The study of prosody in discourse in a rich topic of research in linguistics, but little has be done in terms of well adopted practices in TTS systems. Hirschberg [208] gives a review of many discourse and functional aspects of intonation and describes much of the work that has been done in discourse prosody in TTS. We need to reiterate some points about the augmentative function of prosody and how this affects studies into prosodic phenomena. While it is widely accepted that prosody does undergo more variation, or undergoes more speaker choice than other parts of language, it is our belief that the degree to which this occurs is vastly underestimated in most studies. If we compare prosodic form to verbal form, we see that while speakers can indeed vary the articulation of a word such as HAT, they are very constrained in doing, and too much deviation from the canonical pronunciation runs the risk of the word being decoded incorrectly. A key reason for the difference in degree of variation between verbal and prosodic language is to do with their functional use in communication. To a large extent, we can study the verbal part of language without reference as to why a particular utterance is used in a discourse but the same situation simply does not hold for prosody. The way in which prosody is used is primarily driven by the discourse, and taking sentences out of their discourse and studying them without reference to this, leads to problematic analyses. In verbal phonetics a common practice is to carefully elicit sentences spoken under laboratory conditions. These conditions are artificial and can lead to dis- crepancies between the phonetic/acoustic patterns observed between “canonical” speech and real, spontaneous speech. In general though, a considerable amount can be learned from controlled
- 169. Section 6.10. Discussion 145 experiments in this way. The situation is radically different in prosody however, and eliciting indi- vidual sentences in laboratory conditions is extremely risky in that these bear virtually no relation to how prosody is used in real, spontaneous speech. Many studies unfortunately impose a strong discourse bias, so that when we examine utterances elicited for the purposes of studying say the given/new distinction, what we are doing is asking speakers to produce sentences in a very artifi- cial discourse context. The effect of this is to grossly exaggerate the way the prosody is generated, such that effects such as given/new or contrastive prominence can seem quite clear, distinct and or- derly. In real, spontaneous speech by contrast, the discourse situation and the speaker and listeners mental models of each other dictate how prosody is used. In these situations, the discourse context and information content is often so obvious and redundant that prosody is simply not required for further disambiguation. This helps explain while earlier laboratory based studies of prosody claim clear patterns and effects, but when studies of real data are performed [328], [24], these effects are often far less clear. 6.10.5 Conclusion The take home message from this discussion is that the theoretical difficulties with prosodic mod- els and the under specification of prosody in text mean that this is an inherently difficult problem and should be approached with a degree of caution. While it is possible to continually drive down the error rate with regard to semiotic classification, homograph disambiguation and so on, it is unrealistic to assume the same can be done with prosody. 6.10.6 Summary Prosodic form • It is usual to consider three basic components of prosodic form: phrasing, prominence and intonation. • Phrasing is used to group words together in the utterance, and is influenced by top down factors such as syntax, and bottom up factors which place phrase breaks at roughly equal intervals. • Prominence is used to give extra strength to certain words. It is partly determined by syntax, where in particular it governs stress patterns in compound noun phrases and function word sequences. • Intonation is the use of pitch to convey prosodic information. • Very little agreement exists on how to represent prosodic form, which makes the design of practical algorithms more difficult than in other areas of TTS. Prosodic function • Affective prosody conveys emotion and attitude in communication.
- 170. 146 Chapter 6. Prosody Prediction from Text • Augmentative prosody is used to help disambiguate the verbal component of communica- tion. • Suprasegmental prosody arises from the natural patterns of how words are joined together in utterances. Prosody prediction from text • This is inherently difficult as prosody is largely absent from text. • In practice though, the genres that we often use in TTS are not high in affective content. • Phrasing and prominence prediction can be performed by a variety of algorithms, including simple rules, sophisticated rules, data driven techniques and statistical techniques. • Most popular machine learning algorithms have been applied to prosodic prediction. The results are often similar, and difficult to improve upon because of the inherent difficulties in representing prosodic form and under specification in the text. Prosodic labels and data • In general prosodic labelling by hand has low levels of inter-labeller agreement compared to other labelling tasks in TTS. • This may stem from the fact that we have yet to develop a robust system for describing prosody. • Speakers change their prosody depending on the discourse context and so it is difficult to associate one prosody with a particular sentence. • As speakers naturally vary their prosody, and labels are often unreliable, the labelling of prosodic databases is often very noisy.
- 171. 7 PHONETICS AND PHONOLOGY This chapter gives an outline of the related fields of phonetics and phonology. A good knowledge of these subjects is essential in speech synthesis as they help bridge the gap between the discrete, linguistic, word-based message and the continuous speech signal. More traditional synthesis tech- niques relied heavily on phonetic and phonological knowledge, and often implemented theories and modules directly from these fields. Even in the more modern heavily data driven synthesis systems, we still find that phonetics and phonology have a vital role to play in determining how best to implement representations and algorithms. 7.1 ARTICULATORY PHONETICS AND SPEECH PRODUCTION The topic of speech production examines the processes by which humans convert linguistic mes- sages into speech. The converse process, whereby humans determine the message from the speech is called speech perception. Together these form the backbone of the field know as phonetics. In general, Regarding speech production, we have what we can describe as a complete but approximate model of this process. That is, in general we know how people use their articulators to produce the various sounds of speech. We emphasise however that our knowledge is very approximate; no model as yet can predict with any degree of accuracy how a speech waveform from a particular speaker would look like given some pronunciation input. The reason for this lack of precision is that the specifics of the production process are in fact incredibly complex; while we can make solid generalisations about how a [s] sound is pronounced for instance, it is another matter en- tirely to explain how this varies from speaker to speaker, how [s] is affected by speech rate, how surrounding sounds influence it and how this is affected by prosodic interaction. The fact that this knowledge is incomplete should of course not be a surprise to readers of this book; if we fully understood this process we could simply implement it in a computer and thereby solve the speech synthesis problem. In fact, as we shall see, in the field of text-to-speech, our lack of knowledge has led us to abandon (for the time being at least) the path of trying to mimic the human production process directly. However, what knowledge we have can be put to good use, and hence a solid 147
- 172. 148 Chapter 7. Phonetics and Phonology grounding in what speech production and perception knowledge we do have is an essential part of constructing a high quality synthesiser. Probably the main difficulty in discovering the mechanisms of human speech production is that of measurement and data collection. Of course this is a problem common to all science; before telescopes only the simplest models of astronomy were possible; afterwards detailed and accurate observations could be made which allowed the rapid development of theories. We see the same in the scientific study of phonetics; in the early days there were no instruments available for examining speech (phonetics even predates the invention of sound recording). With the invention first of sound recording, then spectrographic analysis, x-ray films and other techniques, speech scientists were able to examine the processes in more detail. Such instruments and more elemen- tary techniques (such as looking directly into a speaker’s mouth or simply trying to determine the position of one’s own tongue) have led to the level of understanding we have today and which we shall briefly describe next. Progress in speech perception has been significantly slower. The lack of progress can be explained simply by the difficulty in extracting information about the process. While some tech- niques exist for examining the cochlea and analysing brain patterns when listening, in general it is very difficult to collect concrete data on the problem, and as a result scientists have little to go on when developing theories. What ever progress has been made is typically conducted within within the framework of experimental psychology (e.g. [58]]. This significant imbalance between our knowledge of production and perception helps explain why phonetics is in general approached from a speech production point of view; that is simply the area of the speech process that we best understand. 7.1.1 The vocal organs We generate speech by the coordinated use of various anatomical articulators known col- lectively as the vocal organs. Figure 7.1 shows a mid-sagittal section of the head showing the vocal organs. If we consider just the tongue in isolation for a moment, it should be clear that it can move in all three dimensions, and can create a complex variety of movements and trajecto- ries. The movement and position of the tongue directly affects the speech produced and even a movement of a few millimeters can dramatically alter the speech sound produced. While the other organs are somewhat simpler, it should be clear that a substantial number of vocal organ configu- rations are possible, and each gives rise to a different sound. This alone gives some insight into the complexities of speech. 7.1.2 Sound sources Nearly all sounds in English are created by by air moving from the lungs through the vocal organs to the lips and then outwards. This flow is called an egressive pulmonic air stream; egressive, because the air is flowing outwards, and pulmonic because the source is the lungs. During the
- 173. Section 7.1. Articulatory phonetics and speech production 149 nasal cavity oral cavity tongue pharynx glottis larynx vocal folds hard palate soft palate lips nostril alveolar ridge teeth velum uvula Figure 7.1 Diagram of the vocal organs or articulators passage of the air flow, one or more constrictions is applied, the effect of which is to generate a sound. We call constriction that causes the sound the source and the sound produced the source sound. We shall consider the types of source in turn. The vocal folds are two folds of tissue which stretch across the larynx. A speaker can control the tension in his or her vocal folds (Figure 7.2) so that they can by fully closed, narrow or open. The gap between the vocal folds is called the glottis, and we usually refer to this type of sound production as a glottal source. When the vocal folds form a narrow opening, the air stream moving through them causes them to vibrate, giving rise to a periodic sound. We term the rate of vibration of the vocal folds the fundamental frequency, denoted F0. The term pitch is used for the rate of vibration that is perceived by the listener, and in general the pitch and fundamental frequency can be taken as the same thing. By varying the tension in the vocal folds, a speaker can change the fundamental frequency of the sound being produced. When the vocal folds operate in this way, they are said to be generating a voiced sound. A typical male speaker can vibrate his vocal folds between 80 and 250 times a second, so using the standard terminology we say that his fundamental frequency varies from 80 Hertz (Hz) to 250 Hertz. By comparison a female speaker might have a fundamental frequency range of 120Hz to 400 Hz. All vowels are voiced sounds -
- 174. 150 Chapter 7. Phonetics and Phonology (a) The vocal folds (b) Vocal fold vibration. If the vocal folds are tensed in a certain way they will close. If air is pushed from the lungs, it will build un pressure under the closed folds until it forces an opening. The released air will cause the pressure to drop, and the tension in the vocal folds will cause them to close again. This process repeats itself, leading to a periodic noise being emitted from the folds. The amount of tension controls the rate at which this opening and closing cy- cle will happen. Figure 7.2 Vocal fold operation you can experience the effect of the vocal folds by placing your finger on your larynx and speaking a constant “aaah”. You should be able to feel the vocal folds vibrating. If you then whisper “aaah” the vibration will stop. In addition to the fundamental frequency, a periodic signal usually has energy at other fre-
- 175. Section 7.1. Articulatory phonetics and speech production 151 quencies known as harmonics. These are found at multiples of the fundamental, so if we have a glottal sound with a fundamental of 100Hz, we will find harmonics at 200Hz, 300Hz, 400Hz and so on. In general the harmonics are weaker than the fundamental, but nonetheless are heard and can be thought of as giving the sound its basic timbre. A musical analogy might help here. One can create a sound of the same fundamental frequency by either bowing or plucking a violin string. In both cases the fundamental frequency is the same, however the different in action cre- ates a different pattern of harmonics and this is why the timbre of bowing and plucking sounds different. If the glottis is opened slightly further, periodic vibration will cease. Instead a non-periodic turbulent air flow will be created, which generates a different type of sound, termed noise. Here the term “noise” is used in the technical sense (as opposed to a general definition which is a synonym for “sound”) to mean the random sound created, somewhat like an untuned analogue radio. This is the natural way to generate the [h] sound in HEDGE, but is also the way in which whispering works. If the vocal folds are brought together completely, the air stream stops and no sound is produced. By carefully controlling the timing of this opening and closing, it is possible to produce a glottal stop, which is the normal way to realise the [t] in words such as BUTTER in some accents of English (e.g. Glaswegian or Cockney). Sound can be created from sources other than the glottis. If the glottis is open, it will not generate any sound, but will allow the air stream to flow through. It is then possible to use com- binations of the tongue, lips and teeth to form a constriction, and thereby generate a non-glottal sound source. The [s] sound in SIT is generated this way - by speaking this, and extending the length of time the [s] is spoken for, you can feel that the glottis is not vibrating and that the tongue is very close to the roof of the mouth. Noisy sounds can also be created by holding the teeth close to the lips (as in the [f] sound in FALL) or by placing the tongue near the teeth as in the [th] sound in THANK. Because there is no glottal periodic vibration in these sounds, they are called unvoiced. It is possible to have sounds which are both voiced and noisy, such as the [z] sound in ZOO. In these cases, the glottis operates as with vowels, but a further narrow constriction is created in the mouth to generate a secondary, noisy, sound source. Sounds such as [s], [f] and all vowels can be spoken as a continuous sound, and are therefore called continuants. By contrast stops are sounds which have a relatively short duration which can only be spoken as “events” and can not be spoke continuously. The [p] sound in PEN or [t] sound in TIN are stops. Stops are produced by creating a closure somewhere in the vocal tract so that all air flow is blocked. This causes a build up of pressure, followed by a release where the air suddenly escapes. Because the sound is produced in this way, the sound must have a finite, relatively short duration and hence these sounds cannot be produced continuously. 7.1.3 Sound output For most speech sounds, the acoustic signal leaves via the mouth - such sounds are termed oral. For a few however, the sound leaves via the nose; such sounds are called nasal. In English the first
- 176. 152 Chapter 7. Phonetics and Phonology (a) oral output (b) nasal output (c) nasal and oral output Figure 7.3 Oral sounds are when the velum blocks the nasal cavity, and the sound escapes from the open mouth. Nasal sounds are caused, first by blocking sound escape from the mouth and secondly by lowering the velum so that sound can escape from the nose. Nasalised sounds have sound escape from both the mouth and nose. sounds in NO and ME are nasal, as are the last sounds in THEM, THIN and THING. As evidence that these sounds leave via the nose and not the mouth, say “mmmmm” and then pinch your nose so as to close it - the sound will stop. Sometimes sound exists through the both the nose and mouth. There are no canonical sounds in English which do this, but in French for instance for the last sound in ONT exits through both the nose and mouth. Such sounds are called nasalised. The velum is a piece of skin at the back of the mouth. When it is raised, the nasal cavity is blocked off and the speech is completely oral. When the velum is lowered, the sound will be nasal or nasalised, depending on if the mouth is closed or open. It is important to note that in an oral
- 177. Section 7.1. Articulatory phonetics and speech production 153 sound, the velum blocks the entrance to the nasal cavity, and so this plays no part in determining the sound. In nasal sounds however, while the oral cavity is blocked, it is not blocked at the entrance, rather it is blocked at some other point (at the lips in an [m] sound and near the alveolar ridge in an [n] sound). During the production of a nasal, air enters the oral cavity, is reflected and eventual escapes from the nose. The shape of the oral cavity plays an important part in determining the sound of the nasal, even though no air escapes from it directly. Figure 7.3 shows vocal tract configurations for these output types. 7.1.4 The vocal tract filter The diversity of sound from the source and output are further enriched by the operation of the vocal tract. The vocal tract is collective term given to the pharynx, the oral cavity and the nasal cavity. These articulators can be used to modify the basic sound source and in doing so create a wider variety of sounds than would be possible by the source alone. Recall that all voiced sounds from the glottis comprise a fundamental frequency and its harmonics. The vocal tract functions by modifying these harmonics which has the effect of changing the timbre of the sound. That is, it does not alter the fundamental frequency, or even the frequency of the harmonics, but it does alter the relative strengths of the harmonics. In general it is the oral cavity which is responsible for the variation in sound. The pharynx and nasal cavity are relatively fixed, but the tongue, lips and jaw can all be used to change the shape of the oral cavity and hence modify the sound. The vocal tract can modify sounds from other sources as well by the same principle. This model, whereby we see speech as being generated by a basic sound source, and then further modified by the vocal tract is known as the source/filter model of speech. The separation into source and filter not only adequately represents the mechanics of production, it also repre- sents a reasonable model of perception in that it is known that listeners separate their perception of the source in terms of its fundamental frequency from the modified pattern of its harmonics. Furthermore, we know that the main acoustic dimension of prosody is fundamental frequency, whereas the main dimensions of verbal distinction are made from a combination of the type of sound source (but not its frequency), and the modification by the vocal tract. The mathematics of both the source and the filter will be fully described in Chapter 10. 7.1.5 Vowels All vowel sounds are voiced and hence have the same sound source (the glottis). What distin- guishes one vowel sound from another is the shape of the oral cavity and to a lesser extent the shape of the lips and duration with which the sound is spoken. The oral cavity and lips operate to create different shaped cavities through which the source sound must pass. By moving the jaw, lips and tongue, different filter effects can be created, which serve to modify the harmonics of the glottal sound source and produce the wide variety of vowel sounds. It is important to note that
- 178. 154 Chapter 7. Phonetics and Phonology (a) HEAT vowel (b) HOOT vowel (c) HOT vowel Figure 7.4 Vocal tract configurations for three vowels the pitch of the speech is controlled entirely by the glottis, whereas the type of vowel is controlled entirely by the tongue, jaw and lips. As evidence of this, generate any vowel and keep the pitch constant. Now vary the shape of your mouth in any way you wish - it is clear that the pitch remains the same, but the mouth alone is determining the quality of the vowel. Conversely, say an “aaah” sound and convince yourself that you can change the pitch without changing the vowel itself. It helps to describe vowels if we define a number of dimensions of oral cavity and lip config- uration. First, the height of the tongue combined with the position of the jaw defines a dimension called height (sometimes called open/close). If one compares the position of the mouth when speaking HEAT and HOT one should be able to tell that the jaw is raised and the tongue is high in HEAT, whereas for HOT the jaw and tongue are lower. Secondly, the position of the raised part of
- 179. Section 7.1. Articulatory phonetics and speech production 155 the tongue with in relation to the front of the mouth defines another dimension, which we can call front/back. Consider the difference in tongue position in the vowel in HEAT and HOOT - while this is slightly more difficult to feel or measure than in the case of height, it should at least be clear that the tongue is in a different position and that the jaw remains relatively fixed. Figure 7.4 shows the articulator positions for HEAT, HOOT and HOT. A third dimension is called lip rounding or simply rounding. This effect is caused when the lips are protruded from their normal position. Some vowels (e.g. HOOT) have noticeably rounded vowels, while others are noticeably unrounded (e.g. HEAT). A final basic dimension of vowel classification is length. This feature is very accent specific, but in many accents the vowel in HEAT is considerably longer than the one in HIT. In languages such as English, there is a particular vowel sound known as schwa. This is often called the neutral vowel in that its position is neither front nor back, high nor low, rounded nor unrounded. It is first vowel in ABOUT, SEVILLE, COLLECT, and the second vowel in SARAH, CURRENCY, and DATA. Most other sounds are simply called by how they sound; schwa is given its special name because it is difficult to pronounce in isolation: if one attempts this, the sound that comes out is normally too strong sounding. Schwa is spoken with a tongue position which is close to that when the tongue is in its rest position, and as such the tongue doesn’t create any particular cavities or shapes in the vocal tract; rather the vocal tract is in the shape of a fairly uniform tube. Some vowels are characterised by a movement from one vocal tract position to another. In HEIGHT for instance, the vowel starts with a low front mouth shape and moves to a high position. Such vowels are termed diphthongs. All other vowels are called monophthongs. As with length, whether or not a vowel can properly be considered a diphthong is heavily accent dependent. In particular the vowels in words such as HATE and SHOW vary considerably in their vowel quality from accent to accent. 7.1.6 Consonants Consonant sounds are more heterogeneous than vowels in that they are produced by a wider variety of sound sources. While consonants share some inherent characteristics, it is not inappropriate to define consonants as the set of speech sounds that are not vowels. Consonants can be classified in terms of voicing, manner and place of articulation. As we have mentioned, voicing describes whether the sound source is produced by the vibration of the glottis or by another means. Consonants such as [f] and [s] are unvoiced while [v] and [z] are voiced. This can be seen by speaking a [v] continuously and turning it into an [f] - it should be possible to do this without altering the shape of the vocal tract and simply changing the vibration mode of the glottis. The first important category of sounds are known as oral stops. In English, stops occur in voiced/unvoiced pairs, so that [b] and [p] use the same manner and place and of articulation, and differ only as to whether the glottis is vibrating or not. The three nasals (the last sounds in THEM, THIN and THING) are often also described as nasal stops. While it is possible to continuously utter these three phones, they usually follow the pattern of the stop in normal speech, with a closure and
- 180. 156 Chapter 7. Phonetics and Phonology (a) Alveolar constriction (b) Labio-dental constric- tion Figure 7.5 Place of articulation for alveolar and labio-dental fricatives release. During the production of nasals, the mouth is closed, so the sound only escapes from the nose. The position of the tongue in the mouth is important however - the nasal tract is fixed and so does not distinguish the three nasals. The production of nasals can be thought of as the glottis sending a periodic sound into the mouth which is modified (filtered) by the oral cavity and then passed out through the nose, which performs further modification. Fricatives are generated by creating a narrow constriction in the vocal tract, by placing (say) the tongue near to the roof of the oral cavity, thus creating a turbulent, noisy sound. Fricatives can be voiced (e.g. [v]) or unvoiced (e.g. [f]). In voiced fricatives, there are two sound sources, the glottis, which produces a periodic sound source, and the constriction which adds turbulence to this. To demonstrate that this is indeed the case, try saying [v] and [f] with different pitches. This is possible with [v] but not with [f]. The final class of true consonants is the affricates, such as the first and last sound in CHURCH. These are spoken as a stop followed by further frication. What then distinguishes the unvoiced fricatives such as [f] and [s]? The answer lies in where the constriction occurs. For [f] the teeth combines with the lips to form the constriction - this is termed labiodental. In [s] the sound in generated by the tongue coming into close proximity with the roof of the mouth near the alveolar ridge and hence this sound is called alveolar. There are many other possible positions, and each will produce a slightly different sound. While in reality a continuum of constriction positions can occur, for convenience it is useful to identify and name a few key positions for constrictions. These include bilabial in which both lips combine to form the restriction, and dental in which the tongue combines with the teeth. The alveolar position is where the tongue approaches the roof of the mouth in a forward position. The tongue can create a constriction at positions further back, which are termed post-alveolar, palatal, velar and uvular. The place of articulation can be described on a scale starting at the lips and moving to the near the
- 181. Section 7.1. Articulatory phonetics and speech production 157 back of the vocal tract. The important point is not so much which articulators (teeth, lips, tongue) are involved, but rather where the constriction occurs, as this governs the shape of the oral cavity. One important point worth noting about fricative sounds is that the constriction as a sound source does not define the difference between an [f] sound and a [s] sound. In fact the sound source produced by all fricatives is the more or less the same. Rather, it is the fact that the con- striction gives rise to a vocal tract configuration, and it is this acting as a filter which modifies the sound source and thereby produces the difference in sound. As we shall see in Chapter 11, it is not the case that sound always moves forward in sound production. In dental fricatives, while some portion of the sound source is immediately propagated from the mouth, the remainder is propa- gated back into the vocal tract, where it is modified by the filter, after which it is then reflected and propagated forward. It is by this mechanism that the vocal tract generates different sounds even if the sound source is near the mouth opening. The term obstruent is used as a collective term for fricatives, affricates and oral stops to indicate their noisy character. The approximants are a class of sounds which are interesting in that they share many of the properties of both vowels and consonants. All approximants are voiced and all are produced in roughly the same manner as vowels, that is by varying tongue, lips and jaw position. [j] and [w] are know as glides and have similar sounds and means of articulation to the vowels in HEAT and HOOT. In some analyses, diphthongs are described not as compound vowels but as a combination of a vowel and a glide, as the glide comes after the vowel it is termed an off-glide. It is also possible to have on-glides, so CURE is described as a [y] sound followed by a vowel. Glides also function as proper consonants, for example as the first consonant in YELL , WIN and so on. If these functioned as vowels rather than consonants, the perception would be that these words have two syllables rather than one. The approximants [r] and [l] are known as liquids and act more like normal consonants, but still have many unusual properties (see Section 8.2.4). 7.1.7 Examining speech production In traditional phonetics, the tools available to the phonetician were extremely limited and much of what we have so far described was developed with virtually no equipment. Progress was made by simply “feeling” the position of the articulators when speaking, or by looking down someone’s throat as they spoke (which is not very helpful for examining nasal stops). The ability of these phoneticians to construct a model of speech production with such limited means is quite astonish- ing, and by and large, they got the basics of the model correct (or correct in terms of our current understanding). One aspect which was not however realised at this time was the complexities of the syntagmatic dimension, meaning the influence that sounds had with one another when spoken as a sequence. Furthermore the role that “top-down” knowledge was playing in the phoneticians’ examination was also largely unrecognised. As recording and examination apparatus improved it became clear that what had seemed like clearly identifiable distinct sounds of speech where in fact quite hard to classify based on the articulatory evidence alone. We will address this question fully in Section 7.3.2, but for now let
- 182. 158 Chapter 7. Phonetics and Phonology us simply note that it was found that what were thought of as single sounds (e.g. [p]) in fact had a huge range of possible articulatory realisations. Over the years a number of analysis tools have been developed to examine the production process. The first was the use of x-ray still photography, and then x-ray motion pictures. These techniques, especially the x-rays movies clearly showed the various articulators in motion. Unfor- tunately, the availability of x-ray movies is extremely limited due to the now understood harmful effects of x-ray exposure and hence this technique can not be used much. More recently a number of techniques have been developed. These include: electropalatography This takes the form of a plate fitted to the roof of the mouth which measures contact between the tongue and palate. electoglottograph or laryngograph This is a device which fits around the neck and measures the impedance across the glottis. From this, a signal measuring glottal activity can be found. airflow measurement A mask and measuring apparatus can be used for measuring air flow from the mouth and separately from the nose. This is particularly useful in that we can measure the output from the nose and mouth separately, and we can do this in terms of air flow signals rather that the pressure signals which microphones record. magnetic resonance imaging (MRI) This technique, originally developed for medical use, can generate 3-D images of the head and so measure the position of the articulators. electromagnetic articulograhpy (EMA) This is a device which works by attaching small elec- tronic coils to the tongue and other articulators. During speech, the movement of the coils causes fluctuations in a surrounding magnetic field, which allows articulator movement to be tracked. A description of the operation of many of these devices can be found in Ladefoged [272]. Examples of studies using EMA and MRI can be found in [498], [380], [332], [13], [331]. 7.2 ACOUSTICS PHONETICS AND SPEECH PERCEPTION We term output of the speech production process the speech signal. This acoustic signal travels through air to the listener, who can then decode this to uncover the message that the speaker uttered. Speech signals can be recorded by means of a microphone and other equipment, such that with a little care, we can store a perfect representation of what the speaker said and the listener will hear. The field of acoustic phonetics studies speech by analysis of acoustic signals. The idea here is that since this is what the speaker receives when listening, all the required information for perception should be present, and so by suitable analysis we can discern the processes of human speech perception itself. That said, it must be emphasised that to date we have only a very limited understanding of speech perception. We know some of the initial processes (which for instance separate pitch from other aspects of the signal), but virtually nothing about how people actually distinguish one sound from another. Hence speech analysis should be thought of as an independent
- 183. Section 7.2. Acoustics phonetics and speech perception 159 0 2 4 6 8 10 12 14 0 1000 2000 3000 4000 5000 6000 7000 8000 amplitude normalise frequency Figure 7.6 The log magnitude spectrum shows the pattern of harmonics as a series of evenly spaced spikes. The pattern of the amplitudes of the harmonics is called the spectral envelope, and an approx- imation of this is drawn on top of the spectrum. study of the signal, in which we are trying to find the dimensions of variability, contrast and distinctiveness from the signal itself, rather than as an actual study of perception. The main tool for use in phonetics is the spectrogram, shown in Figure 7.7. The vertical axis of the spectrogram represents frequency, the horizontal axis time and the level of darkness amplitude, such that a dark portion states that there is significant energy at that frequency at that time. For illustration, the phones for this example utterance are written along the bottom. Some patterns can be seen immediately. Firstly the fricatives and vowels contrast strongly - the vowels have characteristic vertical streaks, and dark bands which move with time. To a certain extent the fricatives also have bands, but the important difference is that the vertical streaks are completely absent. The vertical streaks in the spectrogram are caused by the vibration of the glottis - each streak represents one pitch period and by measuring the distance between streaks one can estimate the pitch: if for example the glottis is vibrating at 100 Hz (100 vibrations per second) streaks will occur once every 0.01seconds. With the fricative sounds, there is no voicing, hence no periodicity and no streaks. 7.2.1 Acoustic representations The acoustic waveform itself is rarely studied directly. This is because phase differences, which significantly affect the shape of the waveform, are in fact not relevant for speech perception. We will deal with phase properly in Chapter 10, but for now let us take for granted that unless we “normalise” the speech with respect to phase we will find it very difficult to discern the necessary patterns. Luckily we have a well established technique for removing phase, known as spectral
- 184. 160 Chapter 7. Phonetics and Phonology Figure 7.7 Wide band spectrogram analysis or frequency domain analysis. We can use analysis software to transform the signal to the frequency domain and when this is done, it is a simple matter to remove the phase. Figure 7.6 shows what is known as a log magnitude spectrum of a short section (about 20ms) of speech. In this figure we can see the harmonics quite clearly; they are shown as the vertical spikes which occur at even intervals. In addition to this, we can discern a spectral envelope, which is the pattern of amplitude of the harmonics. From our previous sections, we know that the position of the harmonics is dependent on the fundamental frequency and the glottis, whereas the spectral envelope is controlled by the vocal tract and hence contains the information required for vowel and consonant identity. By various other techniques, it is possible to further separate the harmon- ics from the envelope, so that we can determine the fundamental frequency (useful for prosodic analysis) and envelope shape. While useful, these spectral representations only show the speech characteristics at one point in time; a further representation called the spectrogram is commonly used to show how the (phase-free) speech representation evolves over time. Figure 7.7 shows an example, in which the vertical axis of the spectrogram represents frequency, the horizontal axis time and the level of darkness amplitude, such that a dark portion states that there is significant energy at that fre- quency at that time. For illustration, the identities of the sounds for this example utterance are written along the bottom. Some patterns can be seen immediately. Firstly the fricatives and vow- els contrast strongly - the vowels have characteristic vertical streaks of the harmonics, and dark bands which move with time. To a certain extent the fricatives also have bands, but the important difference is that the vertical streaks are completely absent. It is important to realise that the spectrogram is an artificial representation of the speech sig- nal that has been produced by software so as to highlight the salient features that a phonetician is interested in. The software for generating spectrograms can be configured in different ways, for ex- ample to vary the contrast between the levels of darkness. A second way to vary the spectrogram is to emphasise either the time or frequency resolution of the speech (it is generally not possible to do both at the same time). A spectrogram with high frequency resolution is known as a narrow-band spectrogram (shown later in Figure 12.9), one with high time resolution a wide-band spectrogram.
- 185. Section 7.2. Acoustics phonetics and speech perception 161 Wide-band spectrograms are actually more useful for examining the frequency patterns in speech because the lack of frequency resolution somewhat blurs the frequency information and makes the patterns within more visible. Finally, it is often useful to “zoom” with respect to the time domain, so that sometimes a full sentence is displayed, while at others only a few phones. These points illustrate some of the artifacts of spectrogram display so as to inform that two spectrograms of the same utterance can look quite different because of the particular settings being used. In most modern spectrogram software, these settings can easily be varied. With these and other acoustic representations it is possible to study speech from an acoustic (and pseudo-perceptual) point of view. 7.2.2 Acoustic characteristics At integer multiples of the fundamental frequency we have the harmonics. Speech with a low fundamental frequency (say 100Hz) will have closely spaced harmonics (occurring at 200Hz, 300Hz, 400Hz ...) while speech with a higher fundamental frequency (e.g. 200Hz) will have widely spaced harmonics (400Hz, 600Hz, 800Hz etc). The tongue, jaw and lip positions create different shaped cavities, the effect of which is to amplify certain harmonics, while attenuating others. This gives some clue as to why we call this a vocal tract filter; here the vocal tract filters the harmonics by changing the amplitudes of each. An amplification caused by a filter is called a resonance, and in speech these resonances are known as formants. The frequencies at which resonances occur are determined solely by the position of the vocal tract: they are independent of the glottis. So no matter how the harmonics are spaced, for a certain vocal tract position the resonances will always occur at the same frequencies. Different mouth shapes give rise to different patterns of formants, and in this way, the production mechanisms of height and loudness give rise to different characteristic acoustic patterns. As each vowel has a different vocal tract shape, it will have different formant pattern, and it is these that the listener uses as the main cue to vowel identity. The relationship between mouth shapes and formant patterns is complicated, and is fully examined in Chapter 11. By convention, formants are named F1, F2, F3 and so on. Somewhat confusingly, funda- mental frequency is often called F0. Note that fundamental frequency/F0 is not a formant, and has nothing to do with formants - it is determined by the glottis alone. Studies have shown that not all formants are of equal perceptual importance; in fact, the identity of a vowel is nearly completely governed by the frequency of the first two formants (F1 and F2). Figure 7.8 shows a formant chart in which the axes represent F1 and F2 values. Typical positions of each vowel (determined experimentally) are shown on the graph. and from this we can see that each vowel occupies a different position on the graph, giving evidence to the idea that it is in fact the first two formants that distinguish vowels. Other speech sounds have characteristic spectrogram patterns also. Nasals are generally weak and have a wide formant at around about 300Hz, caused by resonance in the nasal cavity. As the nasal cavity is fixed, the resonance will always occur at the same position. Each nasal has
- 186. 162 Chapter 7. Phonetics and Phonology i I ae u First Formant 400 800 1200 Second Formant 500 1000 2000 3000 4000 Figure 7.8 Chart of measured mean positions of vowels plotted in terms of first and second formant positions. The chart shows that to a large extent, the two formant positions separate most vowels. This has led to the general assumption that F1 and F2 positions are used to discriminate vowels. This chart is shown in the form standard in phonetics; note that the axes have been specifically set up to make this plot easier to interpret. its own oral cavity shape however, and the resonances in this are the main distinguishing feature between [m] and [n]. In stops, it is often possible to see the distinct phases of closure and the subsequent burst. While subtle, it is possible to tell one stop from another from the resonant patterns in the burst and in the immediately neighbouring vowel. Approximants look like weak vowels, which is what we would expect. 7.3 THE COMMUNICATIVE USE OF SPEECH We have shown in the previous sections how the vocal organs can organise into a rich variety of configurations which in turn can produce a rich variety of speech sounds. We have also seen that the differences in vocal organ configurations can produce patterns which are discernible in the acoustic signal and representations derived from it. We now turn to the question of how this capability can be used to communicate. 7.3.1 Communicating discrete information with a continuous channel Recall that natural language is a discrete symbolic semiotic system. In this system we combine words in different ways to produce a near limitless number of sentences, each with a separate
- 187. Section 7.3. The communicative use of speech 163 meaning. One might imagine a system of speech communication whereby each word was rep- resented by a single unique speech sound, so that DOG would have a high central vowel, CAT a alveolar fricative and so on. The problem with such a system would be one of distinctiveness: if we had 10,000 separate words we would need 10,000 different speech sounds. It is unlikely that such a system would work simply because it would not be possible for listeners to accurately dis- tinguish all the separate sounds, nor would it be likely that speakers would have enough precision to utter each sound exactly. To get round this problem, we can make use of a much smaller set of sounds, but use multiple sounds in sequence for each word. As the set of fundamental sounds is small it should be easier to distinguish one from the other; and so long as we can “segment” a word into a sequence of separate sounds, we should be able to create a large number of distinct words with a relatively short sequence for each. This is of course how speech communication actually works; we have a small number of units which we can distinguish from one another and we use these in different sequences to create the form of a large number of words. We call these units phonemes. The number of phonemes varies from language to language, but all languages use a roughly similar set of units ranging in size from about 15 to about 50. There is clearly a tradeoff between the number of phonemes we use and the length of words; languages with small numbers of phonemes will need on average to use longer sequences of these phonemes to produce distinct forms. The question now is, out of all the possible vocal organ configurations, which do we use for this process? First, let us say that the evidence shows that we do no simply pick a arbitrary set of 40 particular sounds from the entire set and make use of these alone. We know from our principle of semiotic contractiveness that it makes sense to pick a set of sounds that are readily distinguishable from one another; hence we don’t find languages which only have fricatives or only have front vowels. Secondly we want sounds, or more precisely a system of sound sequencing, that produces sequences that are relatively easy to speak and relatively easy to “segment” when listening. (For example, we would not expect a language to have sequences of velar stops followed by bilabial nasals followed by oral stops, as these sequences are difficult to say). The choice of phonemes in a language is therefore a tradeoff between picking a large number of sounds (short words) and a short number (easier to identify each), sounds which are distinctive from one another, sounds which are easy to speak in sequence and sounds which are easy to segment when heard in sequence. Even given a set of sounds that obey these criteria, it is not the case that each phoneme has a unique articulatory position or acoustic pattern. To take an example, when speaking the word SHOE, most speakers utter the first sound (represented by the symbol [M]) sound with a mouth position in which the lips are pursed or rounded. Compare this with the lip position in SHINE and we see that the lips are in another position, which is more spread and in a position somewhat similar to a smile. However, despite the fact that the acoustic and articulatory differences between the initial sounds in these two words is quite big, they are perceived cognitively as a single sound. The reason why this variation occurs is quite obvious; SHOE has a rounded vowel, and so all the speaker is doing is anticipating speaking the rounded vowel by already having the mouth in this
- 188. 164 Chapter 7. Phonetics and Phonology position before hand. This is therefore a consequence of wanting to produce a sequence which requires less effort to speak. This does not produce confusion because while the rounded [M] is different from the normal [M], the rounded [M] is not confusable with any other phoneme, and hence the listener should not have difficulty in identifying the correct phoneme. Note that it is quite possible that the rounded and unrounded versions of [M] could be different phonemes in another language; it is just a property of English that they are grouped together. Hence the language system makes use of a set of contrasting sounds as its basic units, but within each unit a considerable degree of variability is allowed. These properties are entirely to be expected from the semiotic principles of contrast and sequencing, and the communicative principles of ease of production and ease of perception. 7.3.2 Phonemes, phones and allophones We will now define four terms commonly used in phonetics. The term phoneme is used for the members of the relatively small (between 15-50) set of units which can be combined to produce distinct word forms. Phonemes can have a range of articulatory configurations, but in general the range for a single phoneme is relatively limited and occupies a contiguous region in the articulatory or acoustic space. The term phone is used to describe a single speech sound, spoken with a single articulatory configuration. The term allophone is used to link the phoneme and phone: different ways of realising a single phoneme are called allophones. In our example, we identified a single phoneme at the beginning of the two words SHOE and SHINE. As each of these has a different articulatory position, each is a separate phone, but as they both “belong” to the same phoneme they are termed allophones of this phoneme. Finally, we will use the term segment as a general term to refer to phonemes, phones and allophones. Phones can be thought of as units of basic speech production and perception whereas phonemes are specific to a particular language. The way phonemes and phones are connected varies from language to language and is not necessarily a simple relationship. The easiest way to demonstrate this for speakers of English involves demonstrating a case where two sounds are phonemes in English but allophones in a different language. The classic case of this is the [r] vs [l] distinction in South East Asian languages; all English speakers readily tell these apart and usually have no idea they are even related phonetically. In Japanese, there is no such distinction, and they are allo- phones of a single phoneme (by convention usually represented [r]), such that Japanese speakers speaking English often seem to get [l] and [r] “wrong”. To many speakers of English, this seems incredible that such a “basic” distinction can be missed, and this unfortunately becomes the sub- ject of many a lame comedy parody. In fact mastering the allophonic and phonemic distinctions is one of the most difficult tasks when learning a new language as our whole perceptual mechanism seems geared to treating all allophones in our native language as the same. The above in fact shows the most accepted way of determining the phoneme inventory for a language. This is based on the principle of the minimal pair, and in effect says that if one can find two distinct words (such as RUSH and LUSH) that differ only in one sound, then those sounds
- 189. Section 7.3. The communicative use of speech 165 are cognitively distinct in the language, and are classed as phonemes. If on the other we take a word and change the manner of articulation of one of the sounds, but in doing do don’t change its meaning, or the cognitive impression of that word, then the original and modified sound do not form a minimal pair and are not distinct phonemes. Theoretically (and only theoretically) one could randomly combine all the phones in various ways, perform tests on every change, and in so doing cluster all phones into sets of allophones and sets of sounds which form phoneme contrasts for the language in question. In traditional phonetics, phones were seen as a set of distinct, discrete, separable units. This notion is no longer defensible as we now know that in reality the phonetic space is a multi- dimensional continuum (describable either with articulatory or acoustic dimensions), and there are no “bottom up” divisions within it. The concept of using phones should now be seen as an idealised division of phonetic space and this can be justifiable as a handy notational device; it is cumbersome to describe every sound in terms of continuous dimensions (of say tongue height). It is important to realise however, that phones really are a discrete abstraction of a continuous space; no procedure or process can be given for objectively defining these units. One of the advantages of using phones is that we can easily compare sounds from different speakers, accents and languages with some degree of standardisation. In doing this, it is common to make use of the standard International Phonetic Association symbol set (often call the IPA alphabet) to name each phone; having done so we can then transcribe a word to the extent that another person trained in this system can make a reasonable guess at determining the phonetics of the speech from this transcription alone. There are over 100 basic IPA phone symbols and this set is further enhanced by a system of diacritics which adds detail, such as nasalisation, rounding and so on. The standard IPA vowel chart and consonant chart are shown in figures 7.9 and 7.10. In the IPA vowel chart, the vowels are arranged in a pattern that reflects the position of each vowel in terms of height and front/back. For comparison, we also show in Figure 7.11 a revised version of the formant chart of Figure 7.8. We see that the two charts bears a striking resemblance. We have of course arranged the axes help show this (e.g.Ḟ1 has its values going downwards), but even so the diagrams do show a clear similarity. The usefulness of this correspondence is arguable - while on the one hand it demonstrates the distinctness of the vowels, it is important to realise that a vowel being high does not in itself cause a high F2. The relationship between mouth shape and formant patterns is very complicated - and is not a one to one relationship. In fact, as we show in Chapter 10 and 11, many different mouth positions can give rise to the same formant shapes. Although it is rarely described this way, the IPA symbol set can be thought of defining a set of possible sounds based on the union of phonemic definitions. If we imagine each language as being a set of phonemes, each described by boundaries in phonetic space, then the set of phones is the union of the sets of phonemes for all languages. This is of course how the IPA symbol set evolved historically; phoneticians started with a basic set of symbols, and as new languages were discovered or analysed, new symbols so as to further divide the phonetic space. Hence the set of phones should be thought of as a set of named regions in phonetic space, which are defined from
- 190. 166 Chapter 7. Phonetics and Phonology æ V *℄ • • • • T • '• u • 5• o • • = • • • • Œ • a• œ • • ø • e• y • i• Figure 7.9 IPA Vowel chart. The position of each vowel indicates the position of the tongue used to produce that vowel. Front vowels are to the left, back vowels to the right. Where a unrounded/rounded pair occurs, the unrounded version is on the left. Bilabial Labiodental Dental Alveolar Post-alveolar Retroflex Palatal Velar Uvular Pharyngeal glottal Plosive p b t d P c , k g q b Nasal m 4 n 9 7 8 : Trill B r R Tap/Flap D F Fricative C f v S s z M ` L _ CX jX X [ K # h $ Lateral fricative / 1 Approximant Y G H j 6 Lateral approximant l 0 L Figure 7.10 IPA Consonant chart. Where there is a unvoiced/voiced contrast, the unvoiced symbol is shown on the left. the union of all known phonemic boundaries. Yet another way of thinking of phones and phonemes is in terms of cognitive and physical spaces. Cognitively, we represent all the instantiations of a particular phoneme as being the same; in our SHOE/SHINE example, the first sound is clearly the same in the sense that most people are completely unaware that there is even any difference until the differences are explicitly pointed out. By the same measure, although the physical difference between say [p] and [t] may be very slight, people readily agree that they are different sounds. It is important to realise that we don’t group allophones together because we can’t tell the difference, but rather because they group via the cognitive map of the phonetic space. To see this, we can again repeat the experiment where we use an unrounded Mat the start of SHOE and a rounded Mat the start of SHINE. Most people can tell the difference here, and while maybe not being able to attribute this difference to the articulation itself, can none the less hear the difference. So it is not the case that somehow the ear or low level perceptual hearing filters these changes, as we can hear them if we try. It is now timely to comment on the practice of phonemic notation. Unlike the IPA there is
- 191. Section 7.3. The communicative use of speech 167 4000 3000 2000 1000 500 1200 400 800 First Formant Second Formant ae I i u Figure 7.11 Rearranged version of the formant chart of Figure 7.9. The similarity between the F1/F2 positions as shown here, and the height/front positions as shown on the IPA vowel chart can clearly be seen. no universally agreed method of phonemic notation, but a few conventions are common. The first is to simply pick the IPA symbol for one of the allophones of that phoneme. Hence we would represent the first phoneme in SHONE as M. This has the advantage that we are drawing from an established symbol set, and also helps a newcomer to that language quickly grasp the basics of the phonemic system. This method does have its drawbacks however. If a phoneme has two or more allophones only one symbol can be chosen, and this can be seen as unfairly biasing one allophone over another. Secondly, when using a phonemic transcription using IPA symbols, it is sometimes tempting to step outside the defined phoneme set and pick a symbol representing an allophone or other phone. An alternative is to define a completely different set of symbols. A common approach in speech technology is to use one or more ascii characters for each phoneme, so that the consonant at the start of SHOE is /sh/, the vowel is /uw/ and so on. Note that the use of ascii characters is relatively easy when our task is to represent 50 or less phonemes; coming up with a unique ascii representation for each hundreds phones is much more difficult. For now we will use the modified timit ascii character set, defined for general American English: a full description and justification of this will be given in Chapter 8. In addition to any practical, computer benefit, the use of a
- 192. 168 Chapter 7. Phonetics and Phonology different notation system for phonemes than phones helps emphasise the difference between the two. This is helpful for our purposes as it makes it clear that phonemes are a set of contrastive linguistic units, and not entities which are expected to have well defined or invariant physical manifestations. Notationally, it is conventional to use square brackets [ ] when describing phones and segments, and slash “brackets” / / when describing phonemes. Hence for the word HELLO we have a phone sequence [h l oV] and a phoneme sequence /h ax l ow/. 7.3.3 Allophonic variation and phonetic context As has already been mentioned, phonemes do not represent a single point in phonetic space; if they did then each phoneme would have a unique acoustic pattern and the process of synthesis and recognition would largely be trivial. We have already touched on the subject of allophonic variation, which we define as the realisation of a phoneme as a range of possible phones. We have already discussed anticipatory rounding for M, but other anticipatory affects are quite common. For example in the [k] of {COT, /k aa t/} and {KIT, /k ih k/ } we see that in COT the tongue position for [k] is in a back position, whereas in KIT it is much further forward. This can be explained by the position that the tongue will be in during the following vowel. The allophonic possibilities for a phoneme can depend on its position in relation to other phonemes. For example, the [n] in {NIT, /n ih t/} needs to sound different from the [m] in {MIT, /m ih t/}, but the [n] in {HAND, /h ae n d/} does not need to contrast with [m] as there is no word /h ae m d/ that it could be confused with. It does need to be distinct from say [r], [sh] and [v] otherwise it could be confused with HARD, HASHED and HALVED. This shows that the [n] needs to still be a nasal in that it still needs to be identified with respect to some other consonants; it doesn’t need to be identified with respect to other nasals and so its distinctiveness with respect to [m] and [ng] need not be so precise as in other positions. As there are constraints on what phoneme sequences occur; only distinctions between possible phonemes at a given point need to be maintained. The variability mainly arises from the fact that it is physically easier for a speaker to produce a given phoneme one way in one position and a slightly different way in a different position. It turns out that a great degree of verbal variability can be explained once we take phonetic context into account. By this we mean that if we consider the identity of the phonemes surround- ing the one we are interested in; the identity of those surrounding phonemes helps explain why we see the variation in the speech. For example, we find in general that nasals before stops often share the place of articulation with the stop; so we naturally get words such as {HUNT, / h ah n t/} and {HUMP, / h ah m p/}, where the [n] and [t] are both alveolar, and the [m] and [p] both bilabial; but we don’t get words /h ah m d/ and / h ah n p/, where the nasal has one place of articulation and the following stop has another. This phenomenon is called assimilation, reflecting the idea that the place of articulation of one phone assimilates to the other. Another common effect is called colouring, where the identity of the phoneme in question isn’t affected or limited, but where the phoneme takes on some characteristics of the a neighbour-
- 193. Section 7.3. The communicative use of speech 169 ing phoneme. A common type of colouring is nasalisation, in which a vowel followed by a nasal itself becomes somewhat nasalised (that is the velum is not completely closed during the produc- tion of the vowel). Colouring in vowels frequently happens when they are followed by nasals and also to some extent [l] and [r] (in rhotic accents). In English, this is again usually anticipatory. Other effects include elision or deletion in which a sound is “deleted” and epenthesis or insertion is where an extra sound is inserted. For these phenomena, consider the words MINCE and MINTS; they sound indistinguishable and clearly a process of elision or epenthesis has occurred. Interest- ingly, it is not clear which has happened; either we could have elision, in which both would have the phoneme sequence /m ih n s/, or epenthesis in which the sequence for both would be /m ih n t s/. Basic phonetic context on its own is not enough to explain all the variation. Consider the behaviour of [t] in the words TOP, LAST and STOP. In the first two, the [t] is normally aspirated, that is, there is a clear burst of energy after the release, which is heard of as a short section of noise. In STOP there is no burst; the vowel starts immediately after the release. This cannot be explained simply by context alone; after all the [t] is aspirated when preceded by a [s] is LAST and is also aspirated when followed by a vowel in TOP. These and other examples require reference to other structural factors such as the position in the syllable and word. 7.3.4 Coarticulation, targets and transients If contextual allophonic variation was all we had to contend with, the phonemic to acoustic process would still be quite simple; all we would need is a model whereby each phoneme had a range of possible realisations; the process of conversion from one domain to the other would still be relatively trivial. In fact, even if we take into account any allophonic variation that may occur, we find that the articulatory and acoustic patterns for a single phoneme are quite complex. This is because it is not the case that a single allophone is spoken with a single articulatory position; rather the articulators are constantly in motion. This is readily discernible from any of the acoustic or articulatory rep- resentations already encountered. The articulators require time to move from one position to the next; and while this can be done quite quickly, it is usually the case that by the time the articulators have reached the position required to speak one phoneme the speaker is already moving them to a new position to speak the next phoneme. This means that the articulators are constantly moving, and that the pattern of movement for a particular phoneme is heavily dependent on the phonemes preceding and following. This phenomenon is known as coarticulation indicating that to a certain extent, two neighbouring phonemes have a joint articulation. As the number of possible contexts a phoneme can occur in is very large, this massively increases the possible variability in phonetic patterns. A impressionistic model of the process can be thought of as follows. Each allophone has a set of canonical articulator positions. When each phoneme is “input” to the production apparatus, the articulators start to move to those canonical positions and because of this “aiming” effect,
- 194. 170 Chapter 7. Phonetics and Phonology these positions are often referred to as targets. At some point, the next phoneme will be input, and the articulators will start moving to the target for the required allophone of this new phoneme. Importantly, the target positions of the first phoneme need not have actually been reached, it is possible to start moving to the second phoneme before this happens. Sometimes this process can continue with no target positions ever being reached. Again our model of semiotic contrast helps explain communication is still possible under such circumstances; so long as the positions reached are sufficiently near the canonical target positions that the phoneme can be distinguished from other possible phonemes, then the listener should be able to decode the speech. Again, while it would be possible for the speaker to reach the target position and hold this for some time before moving on, to do so would substantially lengthen the time taken to say the message, which could be disadvantageous to both speaker and listener. The degree of “undershoot” with regard to targets is significantly affected by the rate of speaking; when speaking slowly the targets are often met; when speaking rapidly, the speech tends to “blur” and can become less distinct. To some extent, speakers can control this effect; we notice that some speakers “pronounce” their words distinctly; we say they “enunciate” clearly or are “articulate”. It should also be clear that the degree of undershoot does affect the chance of being understood; slow, “careful” speech is in general easier to recognise. It is important to realise that coarticulation and allophonic variation are distinct phenomena. Coarticulation is largely a physiological process outwith the speaker’s control; while it may be possible to speak more slowly or with more effort, it is very difficult to explicitly change the trajectory of the articulators. By contrast, while allophonic variation may have originally arisen by reasons of minimal effort, it is a property of the particular language and we find speakers in one language can easily control and differentiate an allophonic effect which seems automatic to speakers in another language. 7.3.5 The continuous nature of speech One of the most startling facts about speech is that there are no specific acoustic cues in the signal to indicate boundaries. That is, the speech waveform is in general a completely continuous evolv- ing signal, with no gaps or other markers that show us where the boundaries between phonemes, syllables or words lie. In general sentences and often long phrases are delimited by silence, but this is about all we get from the signal in terms of specific markers. This is of course in sharp contrast this with writing systems such as that used for English, where the letters are distinct, and words usually are delimited by spaces or punctuation. Note that this fact of speech should be less surprising to speakers of languages which do not have gaps between the written word, or which use little or no punctuation. It is important not to confuse the notion of phoneme with the speech that is encoded by these units. Strictly speaking, when we look at a waveform or spectrogram, we are not looking at a sequence of phonemes, but rather the output of the speech production process that has encoded these phonemes. Hence strictly speaking, we can not identify phonemes in acoustic representa-
- 195. Section 7.3. The communicative use of speech 171 tions; rather we identify a section of speech that we believe is the output of the speech production process when that phoneme was spoken. Making this distinction is not mere pedantry. Of course informally we often fail to distinguish a representation of an object from the object itself; if I show a photograph of a car and ask what it is, and someone answers “a porsche”, it would seem extremely pedantic if I then said, “no, its a photograph of a porsche”: usually it is clear what is intended and no-one really mistakes a photograph for a real object. In speech analysis however, this mistake is sometimes indeed made, which does lead to genuine confusions about the differ- ence between the cognitive unit of contrast (the phoneme) and a particular example of speech that encodes that phoneme. This has practical implications in speech analysis because the coarticulation effects can be so strong that the linearity of the phoneme sequence is lost in the speech. The targets generated by the phonemes may interact to strongly that they do not encode in sequence and therefore it is impossible to point to a particular section of speech and say “that is a representation of a /m/ phoneme”. There is a limit to these interaction effects; in broad terms one can locate a section of speech in an utterance and say what word, syllable or phoneme that this represents, but it is not always possible to do this with a high degree of precision. Examples of this effect are shown in Figure XXX. Here we can see that some phonemes have speech representations that are clearly identifiable; furthermore the “boundary” between the speech encoding one phoneme and the speech encoding the next can clearly be seen. In other cases however, this is certainly not so; the spectrogram shows XXX and XXX and XXX. This phenomenon questions the concept of whether we can in fact describe instances of speech using a linear sequence of symbols at all. While we can describe the canonical pronunci- ation of each word in terms of a sequence of phonemes, it does not follow that we can describe speech generated from these linear representations in a linear way. As the effects of the phonemes overlap when uttered, there is no simple or fully legitimate way of dividing the speech into a se- quence of discrete units. That said, contiguous sequence descriptions are highly attractive from a practical point of view; people find them easy to use (after all, normal writing is done in this style) and furthermore, sequence representations are highly amenable to computational analysis as it allows us to use many of the standard tools in signal processing and finite state processing methods. 7.3.6 Transcription We are often confronted with the problem where we have to determine the correct sequence of words, phonemes or phones from a given a waveform of speech. The general name given to this process is transcription. Transcription is involved in virtually every aspect of analysing real speech data, as we nearly always want to relate the data we have to the linguistic message that it encodes. In considering this problem, let us first try and determine just what exactly we should tran- scribe. This depends on our purpose, and we find many different transcription systems, notations
- 196. 172 Chapter 7. Phonetics and Phonology and devices that have been employed for various tasks. We can however describe most systems as lying on a scale between ones which attempt an “abstract” or “close to the message” transcription, and ones which attempt a “literal” or “close to the signal” transcription. Along these lines, in tra- ditional phonetics it is common to make reference to two main types of transcription; called broad and narrow transcription. For our purposes we find it useful to further divide broad transcriptions into canonical transcriptions and phonemic transcriptions. Each will now be explained. Most words have a single canonical pronunciation; it is this that we think of as being stored in the lexicon; it is this which is the “definition” of how that word sounds. Hence a canonical transcription aims to transcribe the speech in terms of the canonical pronunciations which are given for each word in the lexicon. In such a system, the task can be described as one where we first identify the words, and then transcribe the speech with the phonemes for each word and finally mark the boundaries between the phonemes. Complications lie in the fact that the speaker may say a filled pause (i.e. an “um” or “err”), or that they may “mispronounce” a word (i.e. they might say /n uw k y uw l er/ instead of /n uw k l iy er/ for NUCLEAR). In such cases, one must decide whether to take account of these effects and label them, or whether to stick to the canonical pronunciation. A second type of broad transcription, known as phonemic transcription takes a slightly more “literal” approach, whereby the transcriber marks the sounds as he or she thinks they occur, but in doing so only draws from the inventory of defined phonemes. Thus in such a system, the transcriber would be able to describe the differences between the two renditions of NUCLEAR. Narrow transcriptions, also simply known as phonetic transcriptions, take this process further, and attempt to transcribe the speech in terms of a set of phone symbols (such as the IPA set) rather than phoneme symbols. This type of transcription is more tricky as of course the range of labels to choose from is far greater and it is usual in narrow transcriptions to find vowels marker with nasalisation diacritics, stops with aspiration and so on. Narrow transcriptions have a somewhat tenuous existence which follow from our description of the problems of segmentation and classification described above. In traditional phonetics, narrow transcriptions were meant to be an indication of “phonetic reality”, a sort of high level recording of the speech uttered. The problem is of course that we now know that such a task is impossible as the speech is simply too greatly under-specified for a pure bottom up determination of the phones to be found. What traditional phoneticians were implicitly doing was to listen and understand the speech, so that for each word the canonical or general phonemic sequence was known. From that, detail could then be added to describe the specifics of nasalisation, rounding, deletion of consonants and so on. While some phoneticians did in fact attempt to perform narrow transcriptions for languages they did not speak or understand, this task was considerably harder and many results fell short of the accuracy achieved with known languages. One benefit of narrow transcription is that it is somewhat free from the problems of enforcing a fixed pronunciation (canonical transcription) or a fixed sound set (phonemic transcription) on speech which clearly is saying something different. For instance, if a speaker of English says a word in French with a good French accent, we would find it impossible to express this accurately with a broad system.
- 197. Section 7.4. Phonology: the linguistic organisation of speech 173 As a final example, consider the example of MINTS and MINCE that we mentioned before. While the canonical pronunciation of these words can be different (/m ih n t s/ and /m ih n s/), when transcribing speech, we are in the position of having two words which sound the same. We could use a single canonical pronunciation for both, but this would mean that in a process of word formation (described below in Section 7.4.2) we would need some process or rule which would delete the /t/ of MINT when the plural morpheme was added. Or we could use different transcriptions for each, meaning that we would have to acknowledge that two transcriptions can sound the same. Or, two words could have the same narrow transcription but different broad transcriptions; but again we would have to decide what this was and whether the [t] should be deleted. The main point is that there is no single sure way to determine the transcription for real speech. While useful in various ways, all transcriptions are an enforced symbolic descriptions of what was said and none reflect the “full reality” of the message and signal from every perspective. 7.3.7 The distinctiveness of speech in communication Given these two problems of segmentation and classification, one can rightly ask, how then is it possible that humans can perform these tasks and thereby decode the speech when listening? The answer again lies in our model of semiotic communication. We find that people make heavy use of top down information when listening to speech; such that, using their internal grammar and knowledge of the conversation and context, they form strong hypotheses of the words that they expect to hear. This limits the choice of phoneme sequences which could agree with these words, and by effectively narrowing the search in this way they make the problem tractable. This is also how modern speech recognisers work. They are composed of two main parts; an acoustic model which models the acoustic patterns for each phoneme, and a language model which dictates the probability that a word (and hence phoneme) will occur in a given context. We find that the degree of “clarity” or distinctiveness between phonemes in speech therefore varies depending on the conversational context. In situations where the listener has a strong idea of what will be said next, we find that speakers tend to be less distinct as the chance of confusion is low. Hence we find that speech in conversations between friends and family can exhibit quite indistinct acoustic patterns. Conversely, we see that in situations where the listeners cannot easily guess what will be said next, we find the level of distinctiveness increases. Lectures or news broadcasts are good examples of this type of speech. 7.4 PHONOLOGY: THE LINGUISTIC ORGANISATION OF SPEECH Now that we have described the production process and the means by which discrete information can be encoded in a continuous signal, we turn to the issue of how speech sounds are organised linguistically. So far we have simply said that words are made from sequences of phonemes, but a closer examination will show however that it is far from the case that any sequence of phonemes
- 198. 174 Chapter 7. Phonetics and Phonology can make up a word and in fact, we find that there is considerable structure to the relationship between words and phonemes. This area of linguistics is called phonology. We can describe phonology as having a dual character because it is related to both the syntactic part and phonetic part of the communication process. Firstly, it shares many properties with the field of syntax or grammar, in that it studies the patterns of sounds that occur within a word, just as syntax studies the patterns of words that occur within a sentence. Phonology is seen as being distinct from phonetics in that it is part of the discrete symbolic language system, and not part of the encoding process. That said, it is clearly the case that the primitives in phonology are firmly grounding in the means of speech production and perception; as we shall see, the rules, grammars and processes described in phonology make use of the sort of phonetic information and descriptions we have previously introduced. 7.4.1 Phonotactics It is generally accepted that just as sentences have a grammar that defines what words can consti- tute a legal sentence, so words in turn have a grammar that defines what phonemes can constitute a legal word. The study of this is known as Phonotactics. As with syntax, it is common to use the mechanisms of formal language theory to define a phonotactic grammar. In fact, the use of formal grammars for this purpose is less controversial than their use for syntax, as the phenomena in question are simpler, and many adequate grammars exist as proofs of concept. Let us first consider some sequences of phonemes which occur at the starts of words. In English, a word can start with a vowel, or one, two or at most, three consonants. Even a casual glance at some English words with three initial consonants show consistent patterns: STRING: / s t r ih ng / SPLENDOUR: / s p l eh n d er / SPRING: / s p r ih ng / SQUASH: /[ s k w aa sh / In fact, it turns out that if a word has three initial consonants, the first must be an /s/. Also, the second consonant can only be a /p/, /t/ or /k/, and the third must be an approximant. There are some restrictions within this; the sequence /s p l/ occurs, but the sequence /s p w/ does not. A simple finite state grammar that defines sequences with three consonants is given in below
- 199. Section 7.4. Phonology: the linguistic organisation of speech 175 0 → s 1 1 → p 2 1 → t 2 1 → k 2 2 → r 5 1 → p 3 3 → l 5 1 → k 4 4 → w 5 From examination of phone sequences, we can gradually build a grammar which generates all and only those sequences which occur. Table 7.12 defines one such set of rules for a grammar of the legal phoneme sequences for most single syllable English words. The set of terminal symbols of this grammar is the set of phonemes, and the set of non-terminals is defined as a set V = {0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16}. A state and transition diagram of this is shown in Figure 7.14. Legal words can be created by starting at state 0, and following any path until an end state is encountered. This grammar can be extended to polysyllabic words by the addition of arcs from the terminal states back to the start states. Some care is needed however to ensure that phenomena that occur only at the starts and ends of words are excluded from the grammar which defines consonant sequences in the middles of words. The grammar in Table 7.12 captures many facts that we have already mentioned, such as the fact that /ng/ cannot occur word initially and so on. While our finite state grammar may accurately describe what sequences can occur, its form is very “verbose” in that there are many regularities which the grammar does not explicitly express. For instance after a /s/, only the unvoiced stops occur, but we have no way of expressing this directly, we have to list these stops [/p/, /t/ and /k/] explicitly. We can look at this problem by taking a different approach as to how we define segments (which in the following can be either phonemes or phones). In our account of segments and their properties, we made use of such terms as “nasal”, “high”, “stop” etc, as these helped us categorise the different ways in which sounds are generated. One of the central ideas in modern phonology is that these terms are not simply labels for groups of somewhat similar segments but act as a set of more basic building blocks. Hence we define a set of phonological primitives known as distinctive features, and from arrangements of these we form segments. This system is sometimes view as an additional level in the sentences/words/phonemes hierarchy, and indeed it follows on from this in that there are fewer distinctive features that phonemes, and arrangements of these few distinctive features give rise to a greater range of phonemes. One significant difference though is that distinctive features do not arrange in sequences to form segments; rather their values are all realised at the same time instant. In this way we can see that the process of phoneme formation is one of “mixing” as opposed to “sequencing”. One analogy therefore is the way that the full spectrum of colour can be formed by mixing the primary colours. Another comes from atomic
- 200. 176 Chapter 7. Phonetics and Phonology 0 → p 1 0 → k 1 0 → b 1 0 → g 1 0 → f 1 1 → l 7 1 → r 7 0 → t 2 0 → th 2 0 → d 2 2 → w 7 2 → r 7 0 → p 7 0 → t 7 0 → k 7 0 → b 7 0 → d 7 0 → g 7 0 → ch 7 0 → dz 7 0 → m 7 0 → n 7 0 → f 7 0 → th 7 0 → s 7 0 → sh 7 0 → v 7 0 → dh 7 0 → z 7 0 → zh 7 0 → l 7 0 → r 7 0 → w 7 0 → y 7 0 → h 7 0 → E 7 0 → s 3 3 → p 4 3 → t 4 3 → k 4 4 → r 7 3 → p 5 5 → l 7 0 → s 6 6 → p 7 6 → t 7 6 → k 7 6 → m 7 6 → n 7 6 → l 7 6 → r 7 6 → w 7 7 → a 9 7 → e 9 7 → i 9 7 → o 9 7 → u 9 7 → uh 9 7 → ii 8 7 → ei 8 7 → aa 8 7 → uu 8 7 → ou 8 7 → oo 8 7 → ai 8 7 → au 8 7 → oi 8 7 → i 8@ 7 → e 8@ 7 → u 8@ 7 → 7 @@ 8 → 8 @ 8 → E 9 9 → t 14 9 → b 14 9 → d 14 9 → g 14 9 → ch 14 9 → dz 14 9 → ng 14 9 → f 14 9 → th 14 9 → sh 14 9 → v 14 9 → dh 14 9 → z 14 9 → zh 14 9 → l 10 9 → s 10 10 → p 15 10 → t 15 10 → k 15 9 → l 11 11 → m 15 9 → n 12 9 → l 12 12 → d 15 9 → f 16 9 → p 16 9 → k 16 9 → m 13 13 → p 15 Figure 7.12 Complete grammar for syllable structure using classical finite state grammar rewrite rules physics; once atoms where thought the primary and hence “indivisible” building blocks of nature; but subsequently electrons, protons and neutrons were discovered. Various arrangements of these three types of particle gives rise to more than the 100 known elements. Likewise, phonemes were thought of as the “atoms of speech”, but now we believe that a more parsimonious model can be built with reference to more basic units. When using distinctive features, a common approach is to define each segment as a feature structure, of the type we introduced in Section 4.5. So we can define a segment [t] as VOICED false PLACE alveolar MANNER stop From this, “t” simply becomes a label which we attach to this particular set of values. This notion has many attractive qualities. Firstly we can argue that it is somewhat more scientifically economical to have a basic set of say 12 primitives, and then to form a larger set of 40 segments from these, rather than have just a flat list of segments. Using our physics analogy we can say that regarding classification of the elements, instead of just saying that we have hydrogen, helium, etc, we instead say that these are made from protons, neutrons and electrons, and that the number of protons and electrons fully defines the difference between each element.
- 201. Section 7.4. Phonology: the linguistic organisation of speech 177 0 → VOICED f alse MANNER f ricative PLACE alveolar 1 1 → VOICED f alse MANNER stop 2 1 → VOICED f alse MANNER stop PLACE labial 3 2 7→ MANNER liquid PLACE alveolar 7 3 → MANNER liquid PLACE lateral 7 0 → VOICED f alse MANNER f ricative PLACE alveolar 4 4 → MANNER stop VOICED f alse 7 4 → OUTPUT nasal # 7 4 → MANNER liquid # 7 4 → MANNER glide PLACE labial 7 0 → MANNER stop PLACE non−alveolar 5 0 → MANNER f ricative VOICED f alse PLACE labio−dental 5 5 → MANNER liquid # 7 0 → MANNER stop PLACE alveolar 6 0 → MANNER f ricative VOICED f alse PLACE dental 6 6 → MANNER glide PLACE labial 7 6 → MANNER liquid PLACE alveolar 7 0 → TYPE consonant OUTPUT oral 7 0 → TYPE consonant OUTPUT nasal MANNER non−velar 7 7 → TYPE vowel LENGTH long 8 7 → TYPE vowel LENGTH short HEIGHT mid FRONT mid 8 7 → TYPE vowel LENGTH short 9 Figure 7.13 Partial grammar for syllable structure using feature structures and finite state grammar rewrite rules. Using this formalism the generalisation that we can have /s/ followed by /p/ /t/ /k/ but not /b/ /d/ /g/ can be written: 1 → VOICED false MANNER stop 2 the idea being that regardless of any other feature values if a segment has features which match this definition then it is allowed by the rule. Hence this rule generates all unvoiced stops in this position. This single rule replaces three rules in the symbol grammar. Even more powerfully we can state:
- 202. 178 Chapter 7. Phonetics and Phonology 0 1 p k b g f 2 t th d 7 k b d g ch dz m n f th s sh v dh z zh l r w y h E p t 3 s 6 s l r w r 9 a e i o u uh 8 ii ei aa uu ou oo ai au oi i@ e@ u@ @@ @ 4 p t k 5 p p t k m n l r w r l 14 t b d g ch dz ng f th sh v dh z zh 10 l s 11 l 12 n l 16 f p k 13 m E 15 p t k m d p Figure 7.14 Diagram of finite state machine for the phonotactics of single syllable words.
- 203. Section 7.4. Phonology: the linguistic organisation of speech 179 0 1 s 4 s 5 non-alveolar-stop f 6 alveolar-stop th 7 non-ng-cons E 2 unvoiced-stop 3 p unvoiced-stop nasal liquid w liquid w r 9 short-vowel 8 long-vowel schwa r l 14 non-h-cons 11 l 12 n l 16 f p k 13 m E 15 m d p 10 unvoiced-stop Figure 7.15 Diagram of phonotactics using features rather than atomic phoneme symbols. 0 → TYPE consonant MANNER non−nasal 7 0 → TYPE consonant MANNER nasal PLACE non−velar 7 which states that a syllable may start with any consonant except /ng/, using two rules, rather than the 14 in the symbol grammar. In this way, we can build a finite state grammar based on feature definitions and in so doing arrive at a much more succinct set of rules. Table 7.13 defines a set of rules which use feature structures rather than symbols as the terminals. A finite-state automata of this grammar is shown in Figure 7.15. While the rules may take a little skill to read and interpret, it should be clear that far fewer rules are needed than before. It is important to note the difference between the formal devices we use and the sequences they define: the use of features does not itself gives us a grammar which more accurately describes which phoneme sequences occur and which do not. Different types of rules, feature systems and operators do allow us to express some things more succinctly than others, and it can be argued that a system that allows us to succinctly generate correct sequences with a small number of rules are more natural. The grammar defined in Table 7.13 is not without its problems and we shall now turn to these. Firstly, one sequence which this grammar does not generate is /s f V/ where V is a vowel. This sequence is found in only a very few words, such as SPHERE and SPHINX. Therefore it seems that we have to add a rule to account for these few words. Furthermore, this rule would be the only one to have two fricatives in the same consonant cluster. Another, somewhat more contrived example is VROOM, which while being an onomatopoeic word, is listed in most dictionaries as being a valid word of English. Again this requires its own rule, which would be the only rule which uses a voiced fricative in a consonant cluster. A further sequence this grammar does not allow is /y/ in a pre-vowel position, as in MULE /m y uw l/. The existing grammar is composed of
- 204. 180 Chapter 7. Phonetics and Phonology three distinct parts; a set of initial consonant sequences (all the sequences leading from state 0 to state 7), a set of vowels (states 7 to 9) and a set of final consonant sequences (states 9 to 16). In the grammar, each is fairly independent, so that any initial consonant sequence can be attached to any vowel. The inclusion of /y/ in the initial consonant sequence part would destroy this simplicity because /y/ can only be followed by a /uw/ vowel thereby ending the separation. This phenomena is even stranger in that it accounts for one of the main accent differences between American and British English, namely that in American, the standard pronunciation of DUKE is /d uw k/, whereas in British English it is /d y uw k/. This generalisation covers all words with this pattern (with the possible exception of NEWS in which many Americans say the /y/). Hence there is some important generalisation going on here. We will return to thorny problems such as these in the discussion, but the real point is that within a rule and feature system, there will always be a trade-off between accounting for all the facts, and doing so in a neat and succinct way. 7.4.2 Word formation One of the central topics in phonology is the study of what is called word formation or lexical phonology. Simply put, this concerns the processes by which individual morphemes combine into whole words. To demonstrate some aspects of word formation, let us look at what the case of plural formation in English. In spelling, the “regular” plural in English is the character s, so that the plural of DOG and CAT is represented as dogs and cats. In speech however, the phonemic form differs for these words, and so we get /d ao g z/ and /k ae t s/; that is, in one case the phoneme /z/ forms the plural, while in the other it is the phoneme /s/. What explains why one form is chosen rather than another? If we look at /d ao g z/ and /k ae t s/ we see that the last phoneme before the plural suffix is voiced in one case and unvoiced in the other; this in fact turns out to be the basis of the distinction, and we see that all words which end in an unvoiced consonant take a /s/ plural, while those that end with a voiced consonant take a /z/ plural. As before, we could express this in purely phonemic terms as: g + PLURAL → g + z b + PLURAL → b + z v + PLURAL → v + z k + PLURAL → k + z t + PLURAL → t + z f + PLURAL → f + s . . . and so on. But as with the case of the phonotactic grammar, this entails a large number of rules which are somehow missing the generalisation that this is simply a matter of voicing. If we use distinctive features, we can succinctly write the rule as:
- 205. Section 7.4. Phonology: the linguistic organisation of speech 181 VOICING true # + PLURAL → VOICING true # z VOICING false # + PLURAL → VOICING false # s In fact, we can do better than this, by making use of some further devices common in phonological theory. One is the principle of markedness which is a linguistic term for the concept of default. The idea is that given a binary or other distinction, one form of behaviour is normal, unremark- able and therefore unmarked, whereas the other form is noticeable, stands out and is therefore marked. In our example, we could say that the unmarked plural morpheme is simple /s/, but some special behaviour is invoked if this is added to a word that ends in a voiced consonant. Hence we would not need a rule for joining /s/ to many words; this would be the default rule that happens when any morphemes are joined (i.e. no changes occur). Hence we only need a rule for voiced cases, which would be: VOICING true # + s → VOICING true # z We might well ask why it should be that /s/ is chosen as the default rather than /z/ (note that we shouldn’t be misguided by the spelling convention here) and this is in fact a prime concern in developing phonological theories, namely that we choose “natural” rules rather than simply ones that describe the date. At this stage it is worth examining why we have chosen to describe the plural formation effect as a phonological one rather than a phonetic one. Wouldn’t it be possible to simply say that this is a matter of allophonic variation or coarticulation? While the rule is clearly based on articulatory principles, it is included as part of the phonological because it is seen as a rule of English rather than a general process of speech. It is quite possible to overrule this effect and say /d ao g s/ and /k ae t z/. More importantly however, we see that there are many non-plural English words that end in /s/ which don’t undergo this rule, for example {BASIS, /b ey s ax s/} and {SENSE, /s eh n s/}. It no more difficult or natural to say PENS as /p eh n z/ than PENCE as /p eh n s/. This demonstrates that this plural formation rule is one of grammar and phonological organisation and not some speech encoding effect. English is rich with such word formation rules. For example, one common process, known as palatalisation, describes the process whereby consonants move from their original place of articulation to a palatal place of articulation, usually resulting in a [sh] phoneme. So if we take examples such as:
- 206. 182 Chapter 7. Phonetics and Phonology PRESS PRESSURE CREATE CREATION DIVIDE DIVISION we see that in the process of forming PRESSURE from PRESS the /s/ has palatalised to a /sh/, and this process has happened with CREATE and DIVIDE also. We can still observe the historical effects in the way the words are spelled; in pressure and creation the original consonant is preserved (even though the pronunciation has changed) whereas in division the spelling has changed to reflect the sound more accurately (but not completely accurately). Another common process is that of vowel tensing in derived forms that we see in words such as: DIVINE DIVINITY PROFANE PROFANITY SERENE SERENITY Phonological rules can extend beyond purely phonemic distinctions; stress shift (explained fully below) also happens in pairs such as: PHOTOGRAPH PHOTOGRAPHY TELESCOPE TELESCOPIC ANGEL ANGELIC In studying word formation in this way, it becomes nearly inevitably to adopt a paradigm where we have a basic representation of a word or morpheme, which when combined with other words or morphemes interacts in some way, and then produces a final version. So for example, we have two words {CAT, /k aa t/} and {DOG, /d ao g/} and a plural morpheme +S/s/. We combine CAT and +S to form CATS/k ae t s/ and this is simple enough. In the case of DOGS however, the pronunciation is /d ao g z/, and we observe that the /s/ has “changed” into a /z/. In this paradigm, we start with a deep representation, and by rules, convert this into a surface representation. Some- times the difference between the two is relatively simple, as in the plural example, but in others it is more complicated. If we consider the word SIGN, we see that it has derived forms SIGNAL, SIGNIFY and so on. What is noteworthy here is that these derived forms contain a /g/ whereas the root form does not. Rather than saying a /g/ is inserted, it has been argued rather that the deep form of SIGN is really /s ay g n/ (or even /s ih g n/) and that by process of rule, this is converted to surface form /s ay n/. Thus the rules can operate even when there is no word formation per se. Such solutions have always attracted controversy in phonology. Firstly, it can be argued that the statement of such rules is heavily influenced by the orthography, and as often the exact nature of orthography derives from some whim of a printer, it is a poor basis on which to form rules. Secondly, once such processes are allowed there seems no limit to what devices can be employed.
- 207. Section 7.4. Phonology: the linguistic organisation of speech 183 vocalic consonantal high back low anterior coronal round tense voice continuant nasal strident Figure 7.16 Binary features in the Sound Pattern of English 7.4.3 Distinctive Features and Phonological Theories In defining a feature system properly, one must formally define the attributes and the type of values they can take. So far we have described a somewhat traditional set of features (manner, place etc) each of which can take a multiple values (e.g. dental, alveolar, or palatal) and have made use of descriptive devices, such as “assimilation” to describe how sounds change in particular environments. In phonology researchers attempt to improve on this, by defining features and rule mechanisms in an attempt to explain these sorts of effects in a more elegant and parsimonious manner. The Sound Pattern of English [91] (known as “SPE”) was a seminal landmark in phonology, in that it was the first full account of how to describe English phonology formally, that is, with precise features and rules. SPE uses a binary feature system, in which each phone is described by 13 features each of which takes a +/- value. The SPE binary feature set is shown in Table 7.16. While some of these have the familiar interpretations, a few are new. ANTERIOR is used for consonants and indicates a front constriction, CORONAL, also used for consonants, indicates an arched tongue. Combinations of these are enough to distinguish most English consonants. TENSE is used for vowels in a very similar way to length and STRIDENT is used to indicate a “noisy” phone such as [s] or [f]. SPE showed how many seemingly desperate phenomena could in fact be explained by rules that made direct access to the feature set. Nearly all rules were posited in terms of context- sensitive rules (these rules are used to define a third fundamental type of grammar in addition to finite state rules and context free rules that we saw before.). For example, the tensing rule that describes how word patterns such as DIVINE, DIVINITY and PROFANE, PROFANITY arise is given as:
- 208. 184 Chapter 7. Phonetics and Phonology V → [−tense] / CVCV meaning that the vowel V in question, becomes lax (the opposite of tense) when it occurs before a CVCV sequence (such as the /n ih t iy/ of PROFANITY). Even for different feature sets, models and theories, rules of this type are very common. It is worth noting however that the use of context-sensitive rules here is merely a notational convenience; most phonological rules do not use recursion, and so the above can readily be implemented with finite state rules, if required for practical purposes. Despite its success in many regards, several problems immediately appeared with SPE. One of the main problems is that if there are 13 binary features, this should give rise to 213 = 8192 possible phones. Not only is this considerably greater than most phonologists would agree exist, it is clear that many combinations (e.g. [+tense, +anterior, +high]) simply don’t correspond to any known or even theoretically possible phone. For this and many other reasons, a huge flurry of activity started with the aim of defining more succinct and more “natural” systems of features, and accompanying rules. In the 1970s, two theories broke from the mold of SPE by proposing that sound patterns could better be understood if more elaborate rule and representation mechanisms were used. Au- tosegmental phonology [180] uses two (or more) tiers, in which we have a phoneme tier (roughly similar to SPE) and a tonal tier. The idea is that rules can be written to operate on just one tier or the other, which is much simpler that would be the case if the phonemes and tones had to be repre- sented on a single tier. The other theory is that of Metrical Phonology [284] which was designed to explain phonological phenomena of stress and intonation (described below). Together, these theories and some that follower were often called non-linear phonology (this is a different use of the term “non-linear” to that in mathematics), because their representations were not sequences of segments as in traditional phonetics, phonology and SPE. Since then a bewildering variety of other theories and models have been produced. Among these are Natural Phonology [137] which argued against the sometimes very abstract positions taken in theories such as SPE (for instance the example we gave where the underlying representation of SIGN would require a [g] segment). Despite the large number of different theories, some trends have stood out in how phonology has developed. One strand of work has attempted to move away from the context-sensitive rule formalisms described above. Mainly this follows a general move in this direction in other fields of linguistics such as syntax, and is motivated by a desire to avoid the complications of determining what these rules really stand for cognitively. For instance, it is often difficult to believe that a speaker or hearer has to apply a whole sequence of rules in just the right order to generate the surface form of a word. Instead, a series of constraints are proposed, from which new words can be formed. Theories of this school include Dependency Phonology [145], [144], and Govern- ment Phonology [248]. Many theories use feature systems more elaborate than those expressible naturally in feature structures. The idea here being that features form hierarchies, and in doing so, we can use these hierarchies to naturally limit the number of feature combinations and rules, thus
- 209. Section 7.4. Phonology: the linguistic organisation of speech 185 avoiding the “segment explosion” problem of SPE. The theories of Articulatory Phonology [70] and Feature Geometry [100], [306], [386] are based on this idea. In addition to the issues discussed in this chapter, it should be understood that the field of phonology tackles many issues which aren’t directly relevant to speech synthesis needs, such as say finding the set of language universals. Optimality Theory is one such model which has gained particular attention in recent years [362]. In addition to this, the emphasis on phonol- ogy even within directly relevant topics is often somewhat different from what we require from an engineering perspective. In particular the old issue of “economy” comes up again and again, with phonologists striving for every more natural and economical representations, without usually much interest in assessing the processing requirements incurred. The general trends in phonol- ogy are clear though; a move away from finite state, linear models towards more complex feature organisations, a move from ordered context-sensitive rules to systems which use constraints and dependencies, and a move from procedural systems to declarative systems. Somewhat ironically, while most phonologists trace their work back to SPE, the “formal” side of this approach has largely been forgotten, and few works in phonology since have anywhere near the level of ex- plicitness and comprehensiveness of that work. A notable exception comes from the work of Computational Phonology [43]. 7.4.4 Syllables One of the most important aspects of phonology concerns the structural representation of sound patterns above the phoneme. We have already seen that morpheme boundaries are significant (e.g. in the difference between the realisation of PENS and PENCE) as are word boundaries. In addition to these, we find it is very useful to make use of a third unit, the syllable as this also helps explain many of the effects and patterns in speech. As we have already seen, the syllable provides a solid basis for defining the phonotactic grammars of what constitutes a legal word. If we define a grammar for a syllable, it is an easy matter to define a grammar for any words by simply including more syllables and any possibly special sequences for word beginnings and ends. That said, it is interesting to ask real evidence do we have for syllables as real entities. We seem to instinctively know that SET, FROM and BATH have one syllable; that TABLE, WATER and MACHINE have two, and OPERATE, COMPUTER and ORGANISE have three. These judgments seem intuitive and are not dependent on any expert knowledge. Secondly, if we consider singing we see that in normal singing each note is sung with one syllable. If this principle is broken the singing can sound awkward if syllables are inappropri- ately squeezed in or elongated over several notes. Thirdly, syllables seem to act as the basic unit of many aspects of prosody. Syllables carry stress such that in THE PROJECT the first syllable of PROJECT is stressed, but in PROJECT THAT LIGHT the second gets the stress. There are no other possibilities - it is not possible to stress the /p r/ and not the vowel, or any other individual phone. So what exactly is a syllable? We will take the view that the syllable is a unit of organisation. As such, all words are composed of a whole number of syllables, and all syllables are composed
- 210. 186 Chapter 7. Phonetics and Phonology of a whole number of phonemes. For an initial working definition, we will state that a syllable is a vowel surrounded by consonants. This definition seems to fits the facts fairly well - in all the above cases the numbers of vowels and numbers of syllables were the same, and again from singing we can see that the vowel seems to be the “centre” of the note. There is one commonly raised exception to this, namely the existence of so called syllabic consonants. Some phoneticians give two syllable words such as BOTTLE and BUTTON the transcriptions /b aa t l/ and /b ah t n/, in which the second syllable is comprised only of /l/ and /n/ sounds. Conventions differ as whether to allow this, or use the alternative, which gives these words the transcriptions /b aa t ax l/ and /b ah t ax n/. There is probably no real way to decide from a phonetic point of view, but practical TTS experience has shown no deficiencies in using /ax/ and as this makes the definition of the syllable simpler, we shall use that and not use syllabic consonants. So, taking the vowel as the centre of the syllable, all we have to do is decide which conso- nants belong to which syllables. In words such as HOTEL, we can safely say that the /h/ belongs to the first syllable and the /l/ to the last, but what about the /t/? There are a number of positions to take on this. One position is the so-called maximal onset principle where consonants that potentially could be in the onset or coda of a syllable are taken as being in the onset. Using the symbol /./ to denote a syllable boundary, HOTEL is then represented as /h ow . t eh l/. This will provide a satisfactory account for many words, and it can be argued that this has some cognitive reality because, again from singing, we find that consonants tend to follow this pattern. There are a number of problems however. Firstly, in non-word-initial syllables which have /s t r/ and other such sequences, {INSTRUCT, /ih n s t r ah k t/} it can be argued that the /s/ attaches to the first syllable and not the second. Secondly, consider such words as { BOOKEND, /b uh k eh n d/} - here it definitely seems that the /k/ attaches to the first syllable - in fact a syllable final /k/ and an syllable initial one sound quite different and so /b uh k . eh n d/ sounds different from /b uh . k eh n d/. There is an obvious reason to this, namely that BOOKEND is a word formed by compounding BOOK and END, and it seems that the word/morpheme boundary has been preserved as a syllable boundary. One of the main reasons we are interested in syllables is that pattern of syllables within a word, and the pattern of phonemes within a syllable helps determine much of the phonemic variation due to allophones and coarticulation. Hence it is often useful to assign a basic constituent structure to describe the structure of phonemes internal to a syllable. We will explain one way of doing this, in which four basic constituents are used, called the onset, rhyme, nucleus and coda. The syllable structure for the single syllable word STRENGTH in these terms is: Onset s t r Rhyme Nucleus eh Coda ng th
- 211. Section 7.4. Phonology: the linguistic organisation of speech 187 The first distinction is between onset and rhyme. Naturally enough, words which have the same rhyme do in fact rhyme, for example LEND and BEND: Onset l Rhyme Nucleus eh Coda n d Onset b Rhyme Nucleus eh Coda n d Within the rhyme, the nucleus is the vowel, and the coda is the post vocalic consonants. If extra- metrical phones are included in the syllable, these will be included in the coda. So for a word such as CONSISTS, the syllable structure would be one of: Coda k Rhyme Nucleus ax Coda n Onset s Rhyme Nucleus ih Coda s t s Coda k Rhyme Nucleus ax Coda n Onset s Rhyme Nucleus ih Coda s t EM s 7.4.5 Lexical Stress Some syllables sound “stronger” or “heavier” than others. For instance, taking a word such as TABLE, we can say the first syllable of this sounds stronger than the second. This contrasts with a word like MACHINE, where we find the second syllable sounds stronger than the first. Because of these types of distinctions, we view stress as being a property of syllables rather than of phonemes
- 212. 188 Chapter 7. Phonetics and Phonology or words, and hence we always talk about the stress of a syllable, rather than the stress of a word or phoneme. It is important to note that the term stress is used in a number of different ways in the linguis- tics literature. The most fundamental distinction is between lexical stress and sentential stress. Lexical stress is a core property of a word, and as such doesn’t change. In our previous exam- ple, TABLE is stressed on the first syllable while MACHINE is stressed on the second; and these are properties of the words themselves. Sentential stress on the other hand describes the effect whereby one syllable in the sentence receives more stress than the others; and this is normally due to some pragmatic or communication reason, and is not a fundamental property of the word itself (e.g. I SAID I WANTED THAT ONE. We shall ignore this second type of stress for now as this is dealt with in detail in Chapter 8. As usual, there are many different theories and models of lexical stress. Here we will adopt a relatively simple model which describes the stress patterns without necessarily tackling the issue of how these patterns arise in the first place. In this model, each multi-syllable word has exactly one primary stressed syllable. Using the IPA symbol hto indicate primary stress, we can write TABLE and MACHINE as /’t ey b . ax l/ and /m ax .’sh iy n/. It is also common practice to use numbers to indicate stress, so that we have /t ey 1 b . ax l/ and /m ax . sh iy 1 n/, where again by convention, the number immediately follows the vowel. We can indicate stress in longer words such as INFORMATION in the same way as /ih n . f ax . m ey1 . sh ax n/. What of the other syllables? It appears that the first syllable of INFORMATION is stronger than the second and fourth, but less strong than the third, so we call this secondary stress. Some linguists go further and have tertiary stress and so on, but primary and secondary stress patterns are usually taken as sufficient. Secondary stress is denoted by the IPA symbol i , or alternatively by a 2 following the vowel. We can identify a third category called reduced. This is mainly used to indicate the stress level of syllables containing the schwa vowel. Reduced vowels occur in at least three subtly different cases. First, some words naturally, or always, have a schwa in a certain position, for example in the first syllable of MACHINE / m ax sh iy n/. But consider the second syllable of INFORMATION. We have given this a schwa in our annotation and most speakers would agree with this pronunciation. Now consider the word {INFORM, /ih n f ao r m/ }, here the second syllable is a normal full vowel. We can say that INFORMATION is created from INFORM by the application of morphological rules, and in the process, the second syllable has been reduced from a full vowel to a schwa. A third form of reduction occurs in function words, when for instance a word such as {FROM, / f r ah m/} is produced as /f r ax m/ in many instances. Finally, when people speak quickly, they often under-articulate vowels, which leads to a loss of distinction in their identity leaving them sounding like a schwa. This effect is also termed reduction. The number 0 is used to represent a reduced vowel, so the full transcription for INFORMATION is /ih2 f ax0 r m ey2 sh ax0 n/ In English, stress is a fundamental part of the pronunciation of a word, and as such weak- ens our the notion we have been adopting until now that a word’s pronunciation can be repre-
- 213. Section 7.4. Phonology: the linguistic organisation of speech 189 sented only by its phonemes. There are main related noun/verb pairs in English which are dis- tinguished by stress, such that we have {RECORD NOUN, / r eh1 k ax0 r d/} contrasting with {RECORD VERB, / r ax0 k ao1 r d/} ; and likewise {PROJECT NOUN /p r aa 1 j eh2 k t/} contrast- ing with {PROJECT VERB, / p r ax0 j eh1 k t/ }. That said, there are very few genuine minimal pairs in English. Some have gone to considerable lengths to argue that stress can in fact be derived from just the phonemes (this is advocated in SPE). Valiant though these attempts are, there are simply too many words with unpredictable stress patterns for us to be able to derived stress from phonemes for all words. Stress can be graphically represented in the sort of tree used to describe syllable constituent structure. In our previous example the nodes at the syllable level were unlabelled; we can now show the same tree but with stress labels at the nodes: 2 Rhyme Nucleus ih Coda n 0 Onset f Rhyme Nucleus ax Coda r 1 Onset m Rhyme Nucleus ey 0 Onset sh Rhyme Nucleus ax Coda n An alternative to the use of absolute levels of stress such as primary and secondary, is the theory of metrical stress [284], [286] in which stress is purely relative. In this, a binary tree is used to represent each word, and within this, each syllable is defined as either strong or weak. So INFORMATION would be defined as: W S in W for S S ma W tion The strongest syllable is the one which only has S nodes above it; everything else is defined relative to this. If lexical stress is a relative phenomenon, what of single syllable words? As there are no other syllables to contrast with, the stress status of such words is technically undefinable. However,
- 214. 190 Chapter 7. Phonetics and Phonology by convention, most of these words are given a primary stress or strong stress in their syllable. The exception is content words, which are given a primary stress if in their full form (/f r ah m/) or no stress in their reduced form (/f r ax m/). In order for us to perceive stress differences, stress patterns must of course manifest them- selves in the acoustic signal. How is this achieved? Firstly, taking a contrast between “normal” syllables and their stressed version, we can say that in the case of stress, the syllable is spoken with greater articulation effort, that is, the articulators are more likely to move to their canonical positions, and in so doing, produce a syllable with many of the idealised sounds which we would expect from our discussion of articulatory phonetics. As a by product of this, we can see the that the syllable seems more “distinct” in the spectrogram; the formants of the vowel are clearly ob- servable, and it is often easier to ascertain the identity of the vowel from the formants of stressed syllables than for unstressed ones. Likewise with the consonants, stops often have more distinct closure and burst patterns in stressed syllables and so on. Following from the great articulation effort, stressed syllables are general longer in time than normal versions. The source acts differ- ently in stressed syllables also. The voice quality is likely to be more clearly periodic (not creaky or breathy). Also, the pitch of the syllable is likely to be higher in a stressed syllables than in the syllables immediately surrounding it. This effect is difficult to determine in isolation however, as the intonational patterns of the utterance interact strongly here (described in more detail in Section 9.1. By contrast, in reduced syllables, the opposite effects seem to occur. The articulation effort is diminished to the extent that any distinctiveness in the vowel quality is absent - all vowels are either /ax/ or /ih/. Consonants in reduced syllables tend to be less clear. The duration of reduced syllables can be very short indeed, even to the extent of being non-detectable in faster or more casual speech. Finally, it is unlikely that the pitch in an unstressed syllable with be higher than its surrounding syllables (note however, that the presence of a reduced syllable does not cause a decrease in pitch). 7.5 DISCUSSION The study of phonetics has a long and rich history [97]. Just as the world’s most famous detec- tive is the fictional Sherlock Holmes, so indeed is the world’s most famous phonetician. Today the Henry Higgins of Pygmalion and My Fair Lady would probably appall any self-respecting linguist or phonetician with his dogmatic and prescriptive view of how English should be spoken. Nev- ertheless, Higgins probably does share some genuine features in common with the early British school of phonetics. The study of modern phonology is widely regarded to have started with the “Prague School” of linguists, of which the main figures were Nikolay Trubetzkoy and Roman Jakobson. They established the principle of the formal (or scientific) study of phonology and the notion of the phoneme. Most importantly, they showed the semiotic basis of phonological contrast with the
- 215. Section 7.5. Discussion 191 idea being that the linguistic use of sounds could only be studied as within a system and not in individual, purely phonetically terms. Jakobson collaborated with two other notable figures in speech and linguistics, Morris Halle and Gunnar Fant to produce a coherent system of distinctive features. As already mentioned, the Sound Patter of English was a seminal work, both in its scope and its affect in proposing formal methods for phonological work. 7.5.1 Further reading The most popular work in introductory phonetics is undoubtedly Peter Ladefoged’s classic a course in phonetics [271], and is highly recommended. Another excellent book, which extends the subject and gives more detail on acoustic issues is Clark and Yallop [97]. A more advanced and comprehensive book is Laver’s Principles of Phonetics [275]. There are many excellent introductory text books on phonology including Giegerich [174], Katamba [245], Lass [274] and Anderson [17]. Clark and Yallop [97] is again particularly recom- mended as it gives a thorough introduction to phonology, does so with reference and comparison to phonetics, and discusses the issues involved from the type of traditional scientific perspective familiar to most engineers and scientists. Unfortunately, nearly all these works are quite introduc- tory and few venture beyond SPE type phonology. While there are plenty of books and articles which cover more recent work, they typically focus on one particular theory or model, making it hard to perform the type of comparison that would be necessary for us to adopt any one of these for speech technology purposes. Books and articles on the various schools of phonology are listed above. 7.5.2 Summary Phonetics • Speech sounds can be described with reference to either their articulatory or acoustic fea- tures. • Vowel sounds are described with the primary dimensions of high/low and front/back, and secondary dimensions of rounding and length. • Consonants are described in terms of voicing, manner and place of articulation. The main values manner can take are stop, fricative, nasal, affricate and approximant. • Computer analysis can be used to study the acoustic patterns in speech. The primary repre- sentation is that of the spectrogram. • The resonances of the vocal tract are called formants, and these are thought to be the primary means by which listeners differentiate between different vowel sounds. Communication with speech • A phoneme is a linguistic unit of speech defined on the semiotic principle of contrast.
- 216. 192 Chapter 7. Phonetics and Phonology • We communicate via speech by using the speech apparatus to encode words as strings of phonemes, which are then fed through the speech production apparatus. • In traditional phonetics, phones are seen as concrete objective units, while phonemes are contrastive linguistic units. Phonemes are realised as phones and if a phoneme has more than one phone the set of phones are called allophones. • Even for a single allophone, there is a large amount of variability. This is mostly a conse- quence of coarticulation effects from the continuous movement of the articulators. • In the modern view, there are no objective phones; these are just descriptively useful points in a continuous space. • There are no boundaries or markers in continuous speech; it is not possible to bottom up determine the phoneme or phone boundaries in speech. • We can described the linguistic content of speech using sequences of phonemes or phones; this is known as transcription. Transcription is inherently difficult because we are enforcing a discritisation of a continuous encoding. Phonology • Phonology is the study of the linguistic use of speech • A phonotactic grammar describes which sequences of phonemes are legal for a language • Lexical phonology describes how words are formed from constituent morphemes • Phonemes and phones can be described with distinctive features. More compact and “nat- ural” rules can be created by making use of distinctive features rather that phonemes or phones. • The syllable is a useful organisational unit which sits between the words and phonemes. All words have an integer number one of syllables, each of which has an integer number of phonemes. • Lexical stress is manifested at the syllable level. Exactly one syllable in a word is desig- nated as the primary stressed syllable, all other syllables have either secondary stress or are reduced.
- 217. 8 PRONUNCIATION We now turn to the problem of how to convert the discrete, linguistic, word based representation generated by the text analysis system into into a continuous acoustic waveform. One of the pri- mary difficulties in this task stems from the fact that the two representations are so different in nature. The linguistic description is discrete, the same for each speaker for a given accent, com- pact and minimal. By contrast, the acoustic waveform is continuous, is massively redundant, and varies considerably even between utterances with the same pronunciation from the same speaker. To help with the complexity of this transformation, we break the problem down into a number of components. The first of these components, pronunciation, is the subject of this chapter. While specifics vary, this can be thought of as a system which takes the word-based linguistic represen- tation and generates a phonemic or phonetic description of what is to be spoken by the subsequent waveform synthesis component. In generating this representation, we make use of a lexicon to find the pronunciation of words we know and can store, and a grapheme-to-phoneme (G2P algo- rithm to guess the pronunciation of words we don’t know or can’t store. After doing this we may find that simply concatenating the pronunciations for the words in the lexicon may not be enough; words interact in a number of ways and so a certain amount of post-lexical processing is required. Finally, there is considerable choice in terms of how exactly we should specify the pronunciations for words, and hence rigorously defining a pronunciation representation is in itself a key topic. 8.1 PRONUNCIATION REPRESENTATIONS In this section we shall define just what exactly the pronunciation component should do, what its input should be and what its output should be. 8.1.1 Why bother? First of all, we should ask ourselves - why bother at all with a pronunciation component? Why not just synthesise directly from the words? To answer this, consider an approach where we don’t have any pronunciation component per se, but attempt synthesis directly from the words. In doing so, let us imagine a concatenative synthesiser in which we have a database of recorded samples for 193
- 218. 194 Chapter 8. Pronunciation each word. If we take even a modest size vocabulary, we might have more than 50,000 words, and so this would require 50,000 basic types in our synthesiser. It is not possible for the most part to successfully join words directly; the nature of continuous speech means that the effect of one word influences its neighbours and this must be dealt with in some way. There are two possible solutions to this; either we could adopt an “isolated word” style of synthesiser, where we record each word in isolation and then join the words with a short pause, or we could record one example of each word in the context of each other word. The first approach is taken by many simple concatenated prompt systems, and while quite reliable, will never perform well in terms of naturalness. The problem with the second approach is that for our vocabulary of 50,000, recording each word in the left and right context of each other word would mean 50,0003 = 1.25 × 1014 types in all. Recording this number of words from a single speaker is clearly impossible, but even if some way of reducing this was found, it is clear that the number of required recordings could still be vast. Even if we avoid a concatenative type of approach, we would still have to build a model for every word in the database, and this would still be an enormous undertaking. This is clearly a problem of data sparsity in that we can’t collect or design models for such a large number of types. When we consider that the required vocabulary for a real system may in fact be even larger, and that unknown words will have to be handled somehow, we see that in reality the situation may be even worse. 8.1.2 Phonemic and phonetic input Now consider an approach where we transform each word into its phonemes and use those as the input to the synthesiser. In doing so, we have drastically reduced the number of base types from 50,000 words to the 40 or so phonemes. We still have to deal with continuous speech influence, but if we assume that the scope of this is limited to the immediately neighbouring phonemes, then we will need a left and right context for each type, and this will give us a total of 403 = 64,000 types which is clearly more manageable. The total number is however still quite large, so it is natural to ask if we can reduce this further. The source of the large number is of course having to take the left and right context into account, rather than the number of phonemes themselves. Hence it is reasonable to ask whether we can transform the phonemic representation further into a space which has a further reduction in the number of base types. We can do this by attempting to predict the effect of the context of a phoneme’s neighbours. As we saw in Chapter 7, the traditional model of phonetics states that phonemes are be realised as phones and that in general this realisation process is controlled by context. For example, an /l/ has two allophones, described as clear and dark, where the clear phone is found at the start of a syllable and the dark one at the end; a /t/ is aspirated at the start of a word, but unaspirated when following an /s/ in a consonant cluster. So if we could apply this process to every phoneme, we could generate a sequence of phones for the input, and in doing so free the base types from many of the effects of context. With a phoneme representation, we might have a type /s-t-r/, representing a /t/ in the context of being after an /s/ and before an /r/. With a phonetic representation we replace
- 219. Section 8.1. Pronunciation representations 195 this with just [t] (the IPA symbol for an unaspirated [t]), using [th] for cases where aspirating exists. Context will still play a part, because coarticulation (as opposed to allophonic variation) still operates, but the degree of contextual effect will be lessened. As many phonemes have more than one allophone, the number of base types may well increase (to say 60), but because we can do without a left and right context for each, then the overall number will be considerably lower. 8.1.3 Difficulties in deriving phonetic input There is however a drawback with this approach. The problem with phone representations was discussed previously in Section 7.3.2 where we stated that although phone descriptions could be very useful in certain circumstances, they should be treated with caution as they are a discrete representation enforced on a continuous space; and as such only crudely reflect the actual pro- cesses going on. To give an example, let us consider again an example of place of articulation assimilation that occurs with words such as TEN + BOYS → t e n + b o i z → t e m b o i z where the alveolar place of articulation of the /n/ of the first word assimilates to the bilabial place of articulation of the /b/ of the second word (bilabial). There is no doubt that that something is going on here; the [n] certainly seems somewhat different from a “normal” [n], and this can be verified from both acoustic and articulatory measurements. This is not the issue; rather, it is whether the [n] has turned into an [m]. Such a statement says that there are only two choices; an [n] or an [m] and in certain situations the [n] keeps its identity and in others it changes. Consider the situation of saying TEN BOYS with a pause between the two words; it should be clear that the [n] sounds like a “normal” [n] and that if we said the (made up word) TEM (/t e m/) as TEM BOYS with a short pause, we can tell the difference between the two. Now if we say a sequence of utterances TEN BOYS starting with a long pause and gradually shortening it, until there is no pause, we see that there is a gradual shift in the nature of the /n/, until, when we talk quickly, it seems impossible to make the distinction between TEN and TEM. This shows that assimilation is certainly taking place, but that it is not the case that an underlying [n] is at one point realised by the phone [n] but at a certain points suddenly becomes a different phone [m]. The process is a continuous one of change, and it is quite inappropriate to model it as a discretely. We see exactly the same effects with /r/ colouring, nasalisations and reduction. In a way, we are forcing a discrete level of representation onto a range of effects that have a substantial continuous component; and this is damaging in all sorts of ways. As an example, consider again the phenomena of reduction; whereby we have AND represented underlyingly as [a n d], but where this undergoes reduction, resulting in a range of possible phonetic forms including [a n], [n d], [n], [n] and so on. It may be clear that some level of reduction has occurred, but at what stage do we delete the [d]? Where do we draw the boundary between a full [a] and it being reduced to a schwa? Studies of hand human labelling have shown that even trained phoneticians
- 220. 196 Chapter 8. Pronunciation often give different labels to the same utterance, and much of this disagreement stems from where the draw the boundaries as to whether something is reduced or deleted. 8.1.4 A Structured approach to pronunciation In essence, our situation is this. If we stick with high-level representations such as words, we can have a high degree of confidence about the accuracy of these, but the number of base types is far too large. As we generate representations nearer the signal, we see that the number of base types reduces considerably, but at a cost, in that we are more likely to create a representation that is inaccurate in some way. There is no single solution to this problem: the answer lies in picking a point where the compromise between number of types and accuracy of representation best suits a particular system’s needs. Perhaps unsurprisingly, the consensus as to where this point should be has shifted over the years. When more traditional systems were developed, memory was very tight and hence the num- ber of base types had to be kept low regardless of any errors. In more recent years technological developments have eased the pressure on memory making more abstract representations possible. Given this, there is more choice over where exactly the ideal representation should lie. In fact, as we shall see in Chapter 16, the most successful systems adopt a quite phonemic representation and avoid any rewriting to a phonetic space if at all possible. Because of this, the pronunciation component in modern systems is in fact much simpler than was perhaps the case in older systems, and quite often the input to the synthesiser is simply canonical forms themselves, direct from the lexicon. Hence the solution we advocate for a state of the art unit selection synthesiser can be de- scribed as a structured phonemic representation, where we use the phonemic information more or less as it is specified in the lexicon, but do so in such a way that the synthesiser also has direct access to the context, syllable structure and stress information that governs the allophonic and coarticulation processes. In doing so we are relying on the subsequent synthesis component to model the allophonic and coarticulation effects implicitly. This can be achieved very effectively by representing pronunciation as a phonological tree of the type introduced in Sections. 7.4.4 and 7.4.5, as this allows us to describe the phonemes, syllables and stress information in a single rep- resentation. In such a scheme, the words STRENGTH, TOP, LAST and HAND would for example be represented as follows: Onset s t r Rhyme Nucleus eh Coda ng th Onset h Rhyme Nucleus a Coda n d
- 221. Section 8.1. Pronunciation representations 197 Onset t Rhyme Nucleus aw Coda p Onset l Rhyme Nucleus a Coda s t From these trees, we can access the information we require; the phonemic context can be read directly by accessing a phoneme’s neighbours; the syllable structure is presented as a tree, and from this we can distinguish between /s t/ in a syllable onset in STRENGTH, and offset in LAST. 8.1.5 Abstract phonological representations We have just seen that despite potential issues with data sparsity, in general, higher level phonemic representations are preferred over lower level phonetic ones. In pursuing this approach however, we should ask ourselves whether the phonemic representations we have been using are in fact optimal; might there not be an even more abstract, and less linear representation that would help the dimensionality problem? Progress in modern phonology has in fact moved away from using linear sequences of phonemes or phoneme like units as the “deep” or canonical representation; freed from the constraint that representations must be linear, the more structured “non-linear” representations reviewed in Section 7.4.3 could be considered instead. Ogden et al [334] discusses just this issue and point outs that the problem with specifying phonemic style representations is that this concentrates overly on the paradigmatic dimension; the phonemic system is based on what contrasts occur at a particular point, and the issue of what phonemes follow what other phonemes is left to a different, phonotactic part of the phonology. In fact, we see that it is indeed possible to construct representations in which both the normal notion of phonemic contrast, and a mechanism for governing legal sequences are handled in a unified way. While phonemic representations are less susceptible to borderline decisions than phonetic representations, they are not entirely free of problems. At present, we know of no unit selection techniques which make use of such abstract phono- logical representations, but this possibility is sure to be explored eventually. While there is every chance the field will indeed move in this direction, because of the lack of practical proof we will continue with phonemic representations for the remainder of the book. We should point out that phonemic representations do have one significant advantage over virtually any other type of pho- netic or phonological representation in that they are generally easy for people to work with. Overly phonetic representations can be difficult in that it takes time and effort to fill in the phonetic detail; overly abstract representations require a significant knowledge and understanding of the partic- ular theory which underlies them. With a little training, most people get the hang of phonemic transcription and are able to read and generate transcriptions. This is important for text-to-speech
- 222. 198 Chapter 8. Pronunciation because at some stage, someone must of course actually enter all the lexicon entries by hand. In doing so, we wish to have a system which is both quick (there might be hundreds of thousands of entries) and accurate, and the phonemic system seems to fulfill these criteria reasonably well. That said, we should make the point that it is perfectly possible to enter the pronunciation for a word in the lexicon in terms of phonemes, and then by rule convert this to either a deeper or more superficial representation. Finally, using a linear phonemic representation has the benefit that it makes automatic database labelling considerably easier. In Chapter 17 we shall consider the issue of labelling a speech database with phoneme units, both by hand and by computer. As we shall see, it is easier to per- form this labelling if we assume a linear pronunciation model, as this works well with automatic techniques such as hidden Markov models. 8.2 FORMULATING A PHONOLOGICAL REPRESENTATION SYSTEM Having in principle chosen a system which uses phonemes and syllable structure, we now turn to the issue of how to actually define this for a given speaker, accent or language. It is essential that this is done with care; if we decide say to use a single phoneme /ch/ instead if /t/+/sh/, we have to consistently use this, such that we represent CHURCH as /ch u r ch/ and not /t sh u r ch/ or /ch u r t sh/ and so on. The following sections explain how a phonological system is developed. The first thing that we must decide is which accent we are dealing with from the language in question. In general it is not possible to define a phoneme inventory which can cover multiple accents, so whatever will we develop will be particular to that accent. One solution is to this is to use the actual speaker the TTS system will be based on, and from analysis of his or her phonemic contrasts ensure that the correct pronunciation system is found. This solution can require too much effort, so more general accents often have to be defined. In this chapter, we will conduct an exercise on developing a representation for just two accents of English, namely General American and British English received pronunciation (RP). The principles involved however are the same for most other accents and languages. It does seem somewhat wasteful that separate pronunciation systems and therefore lexicons are required for British English and American English; while differences of course exist, pronun- ciations in the two accents aren’t completely different. Fitt [161], [162], [163] proposed a solution to this where a more abstract pronunciation drawn from a large set of phonemes was used as a base lexicon. From this, filters could be used to generate accents for any accent of English. 8.2.1 Simple consonants and vowels Here we will adopt the standard approach of using the principles of minimal pairs and phonetic similarity to develop a phoneme inventory (see Section 7.3.2). This is not an exact procedure and
- 223. Section 8.2. Formulating a phonological representation system 199 it is possible to come up with two equally valid phoneme inventories for a given accent. There is therefore some discretion in determining a phoneme inventory; the key point is not to find “absolute truth” of what constitutes the set of sound contrasts; but rather to define an inventory that suits our purposes, and then standardise on this; so that exactly the same conventions are used in all parts of the system. Experience has shown that “drift” in phoneme conventions, whereby say we transcribe BOTTLE as /b aa t el/ in one place and as /b aa t ax l/ in another can have a disastrous effect as the data, rules, synthesiser and lexicon can become misaligned. The crucial point is to chose a standard as best we can and stick to just that standard everywhere in the system. Recall that the principle of minimal pairs states that it is not the similarity or dissimilarity of two sounds per se which is the issue, but rather whether they distinguish two separate words. By using this test, it is straightforward to determine a set of contrasts in a particular position. So if we take the form [X i p], we can find a set of contrasts as follows: p PIP m MIP l LIP t TIP n NIP r RIP k KIP s SIP w WHIP b BIP sh SHIP y YIP d DIP z ZIP h HIP The /X i p/ pattern doesn’t give the full list of possible contrast because not every possible legal phoneme sequence actually ends up as a real existent word. But if we take some further patterns, we can elicit further contrasts: X /X eh t/ X /X ih n/ X /X ay n/ p PET sh SHIN v VINE b BET f FIN dh THINE d DEBT th THIN f FINE g GET s SIN s SIGN m MET sh SHINE n NET When we consider one particular pattern, the analysis seems clear: PET and BET are different words, so /p/ and /b/ must be different phonemes. It is reasonable to ask though, how we know that the sound at the start of PET is the same sound as at the start of say PIT? There is no known deterministic or objective way to prove that this is the case, all we can do is rely on judgment and state that in terms of articulation, the two sounds are both produced with the lips, are unvoiced are stops and so on, and that if we examine them acoustically we see they are similar. But they are rarely so similar to be taken as identical; their articulation and acoustic patterns are affected by the following vowel and this does have some effect. To a certain extent, there is a leap of faith required to say that PET and PIT start with the same phoneme. This problem is even more acute when we consider sound contrasts in different positions:
- 224. 200 Chapter 8. Pronunciation p RIP m RIM n RAN t WRIT n RING m RAM b RIB f RIFF n RANG d RID s REECE b BRIT g RIG d WRITHE t STRING So while we have no problem is saying that the last sounds in RIP, WRITE and RIB are different phonemes, it is a harder matter to say that the last sound of RIP should be the same phoneme as the first of PIT. We can again only appeal to some notion of acoustic and articulatory similar- ity. We should bear in mind though that we are attempting a compromise between an accurate representation and a minimal one, and that different possibilities can be tried and tested. If we consider the words SPAN, STAN and SCAN, we see that we have a three way contrast, in which the second sound differs in terms of place of articulation. There is however no voicing contrast possible at these points, so there are only three possible phonemes, unlike in word initial position where we have three place of articulation contrasts in addition to a voicing contrast, giving us six phonemes in all. The question then is, should SPAN be transcribed as /s p ae n/ or /s b ae n/? The phonetic evidence tends to favour /p/, but this is not conclusive, and will vary from speaker to speaker and accent to accent. Clearly though there is a systematic effect going on. There is probably no objective answer to this, so we obey the normal convention (which may have been influenced by the spelling) and represent these words with the unvoiced phonemes /p/, /t/ and /k/. A similar minimal pair test can be used for vowels, in which pairs (or n-tuples) of words can be compared and phoneme identities determined. In the following: iy HEAT uw HOOT ih HIT ah HUT eh HET ae HAT ao HAUGHT aa HOT we can clearly identify a number of distinct vowel phonemes. 8.2.2 Difficult consonants Now we turn to some more problematic cases. In English, the consonants /dh/, /zh/, /ng/ and /h/ are a little more difficult to deal with. There are clearly exactly four unvoiced fricatives; /th/, /f/, /s/ and /sh/ in English and of these, /f/ and /s/ have the voiced equivalents /v/ and /z/, and these are unproblematic. The voiced equivalent of /th/ is /dh/ and while /th/ is a common enough phoneme, /dh/ only occurs in very particular patterns. For a start, the number of minimal pairs that occur between /dh/ and /th/ are few; real examples include TEETH NOUN, / t iy th/ and TEETH VERB, /t iy dh/, while most are “near” minimal pairs such as BATH /b ae th/ and BATHE / b ey dh/ where the quality of the vowel differs. Even if we except /dh/ as a phoneme on this basis, it occurs in strange lexicon patterns. In an analysis of one dictionary, it occurred in only about 100 words out of a total of 25,000. But if we take token count into consideration, we find in fact that it is one of the most common phonemes of all; in that it occurs in some of the most common function words
- 225. Section 8.2. Formulating a phonological representation system 201 such as THE, THEIR, THEM, THEN, THERE, THESE, THEY and THIS. When we take the remainder of the words in which /dh/ is found we see that it is found in many “old” or traditional sounding English words such as FATHOM, HEATHEN, HITHER, LATHE, LITHE, LOATHE, NETHER, SCATHE, SCYTHE, SEETHE, SHEATHE and TETHER. In fact, it only occurs in a relatively few “normal” English words such as MOTHER, LEATHER and HEATHER. Clearly /dh/ is a phoneme, and for purposes of defining our phoneme inventory, it gets a place along with the other fricatives. There are consequences that arise from its use though, in that because of the dominance of its used in function words, it is difficult to obtain “typical” characteristics of the phoneme (say in terms of its average duration) that are not swamped by the fact that it is used as a function word (which are often very short, regardless of what phonemes make them up). Now consider /zh/, the voiced equivalent of /sh/. This occurs in words such as BEIGE /b ey zh/. Again, while there are few direct minimal pairs, some exist (BEIGE vs BAIT or BANE) and so we classify /zh/ as a phoneme. Problems again arise from patterns of use. It has been argued (e.g. [97]) that /zh/ is a relative newcomer to the English language as in the past, words such as TREASURE and MEASURE were pronounced /t r eh z y uh r/ and /m e z y uh r/ and that it is relatively recently that these have been spoken as/ t r eh zh er/ and /m eh zh er/. /zh/ presents the same problems that arose with /dh/, but in this case they are even more difficult because of the very few words naturally occurring with /zh/. These problems with /zh/ and /dh/ can lead to problems in database collection and analysis (see Section 14.1). The sounds [ng] and [h] are noteworthy because they are severely restricted as to the posi- tions they can occur in. Primarily, /h/ can not occur at the end of a syllable or a word, and /ng/ can not occur at the beginning of a syllable or word. There are further issues with /ng/ in that as it only occurs in a post-vocalic position, it suffers from the loss of contrast that affects all nasals in this position. So while it is clear that SIN, / s ih n/, SIM, /s ih m/ and SING, /s ih ng/ all end in different sounds, it is less clear what the correct representation for words like THINK should be. In words such as this the last stop is in a velar position, and the nasal takes on a velar quality also (this is similar to the TEN BOYS effect we discussed earlier). The issue is then whether we should use a velar nasal /ng/ or a dental nasal /n/; there is no contrast possible in this position and so it is a question of convention which we should pick. As with other such issues, the main goal is to aim for consistency; if /ng/ is used here then it should be used in all other post-vocalic pre-velar stop positions. 8.2.3 Diphthongs and affricates So far we have concentrated on the paradigmatic dimension, by contrasting (say) HAT, HEAT and HOT and showing that in a frame of /h X t/ we can distinguish a number of phonemes. We were able to do this because a word such as HOT, clearly divides into three segments. We assign provisional phonemic status to each, swap in other sounds and thereby determine the system of phonemic contrasts. In many words however, it is not so easy to segment the word into distinct parts, and in such words we face a question of not only what the contrasts are, but how many positions are
- 226. 202 Chapter 8. Pronunciation there to have contrasts within. In words such as CHAT, the sound before the vowel can be represented as either a combina- tion of /t/ and /sh/ or as a distinct single phoneme /ch/. A similar case arises with words such as JUICE which can be represented as a combination of /d/ and /zh/ or as a distinct single phoneme /dz/. The argument for using two phonemes stems from the principal of keeping the numbers of phonemes as low as possible; the argument for a single phoneme arises from the idea that these act as a single coherent unit and are more than the sum of the two parts. There is no real way to settle this argument, so we will somewhat arbitrarily decide to use single phonemes /ch/ and /dz/. A similar situation arises with the diphthong vowels in words such as BOY, TOE, TIE and BAY. In each of these, the vowel does not hold a single steady state position, either acoustically or in terms of articulation. Rather, say TIE has an /ae/ sound followed by an /iy/ sound. We face the same situation that we did with affricates in that we could use a more minimal paradigmatic representation by saying two phonemes exist here, or invent new single phonemes to reflect the fact that these sounds seem to act as single coherent units. Again, there is no real objective way to decide, so we will follow our previous example and use a single symbol for each. One benefit from this is that, as well shall see below, this allows us to state a rule that every syllable is formed with exactly one vowel. 8.2.4 Approximant-vowel combinations The approximants /r/, /w/, /j/ and /l/ in many ways operate like normal consonants, so that we have contrasts in words such as ROT, WATT, YACHT and LOT. Of these /l/ is probably the most straightforward and the main point to note is that the acoustic patterns of dark /l/ in words such as PULL, and SMALL is very different from that with clear /l/, such as LOOK and LAMP. This does not pose much of a problem phonemically, but is an important different to bear in mind when trying to decide whether two /l/ units will join well together. The glides /w/ and /y/ are interesting in the ways they interact with vowels. Firstly, there is considerable phonetic similarity between /w/ and /u/ and /y/ and /i/, and in many phonetic analyses, words such as SHOW are transcribed as /sh o w/. If we follow our diphthong principle above, we rule out these cases and transcribe this as /sh ow/. The process of /w/ or /y/ following a vowel is sometimes called an off-glide. As mentioned in Section 7.4.1, one of the difficulties in determining a phonotactic grammar for English is that /y/ often acts as what we can call an on-glide. So in British English, speakers say TUNE as /t y uw n/ and DUKE as /d y uw k/. There is evidence that the /y/ should be attached to the vowel however, just like a diphthong, to give /t yuw n/ and /d yuw k/. Further evidence for this comes from other English accents; in the US, most speakers say /t uw n/ for TUNE and /d uw k/ for DUKE. The effect is systematic, and would seem easier to explain if only one factor (the realisation of a phoneme /yuu/) was changing between accents. Despite this argument, it is most common to regard VIEW as being composed of three segments, and we shall adopt this convention and say it is represented as /v y uw/. Finally, we consider /r/. This is perhaps the strangest and most interesting sound in human
- 227. Section 8.2. Formulating a phonological representation system 203 speech. Setting aside the phonemic question, consider for a moment how this differs phonetically across languages. In English, an alveolar approximant /r/ ([G] in IPA) is most common and it is quite common for this to have a retroflex (curling of the tongue) quality as well. In contemporary Scottish English, a tapped alveolar (D) is quite common, and is separate again from the “classical” Scottish trill [r] (now usually only seen on stage or in film). In French and many accents of German a uvular [J] sound is the norm, with other accents of German having a more alveolar quality. In Spanish, words such as PERRO are formed with a trilled [r], and this forms a minimal pair with the tapped [D] of words such as PERO. It seems incredible given all this variety, that somehow all these sounds are grouped together across languages and represented in writing as an r. Do they even have anything in common? While we will not go into cross language issues here, it is worth noting that words are quite commonly shared across languages, for example the words EMAIL, APRIL and METRE. In the case of METRE, we see that it contains an r in all languages and is pronounced in a wide variety of ways. If we study the acoustics however, we see that these different “types” of /r/ do share somewhat more in common in that they have characteristic energy bands at certain parts of the spectrum. When we consider the phonemics of /r/, we find that it has an interesting effect. First we distinguish between rhotic accents which have /r/ in post-vocalic position (in English, these are Scottish, Irish, General American) and non-rhotic accents which have no realised /r/ (most Eng- land English, Australian and some New York accents). For example, in Irish English, there is no “long a” sound [aa], which is found in RP BATH /b aa th/; this word is pronounced /b a th/. How- ever, in words which have an /a/ and a /r/ following it, the /a gets lengthened, and phonetically we have something like [h aa r d]. Being a phonetic realisation effect, this doesn’t concern us unduly, but it does show how significant (and unexpected to the unwary) the presence of an /r/ can be. More significant are the non-rhotic accents. In these, it is not the case that the /r/ is simply not present; rather (historically or by having some deep presence) it has the affect of altering the quality of the vowels before it. So in a word like HER, which might be represented as /h uh r/ in a rhotic accent, we do not get /h uh/ but rather /h @@/ which has an extended vowel. It is almost as if the /r/ has turned into a /@/ and is still present. Because of this, we are faced with a choice between transcribing non-rhotic FIRE as /f ai @/ or invent a new symbol /ai@/ which would be regarded as a triphthong. 8.2.5 Defining the full inventory From the above we should have sufficient knowledge to construct a phoneme inventory for any accent or language we are interested in. In practice we find that we don’t always have to design a phoneme inventory from scratch; for many languages or accents the existence of a significant computer speech research community means that phoneme inventories already exist. As these communities often develop and distribute resources using these phoneme inventories, it can be highly beneficial to adopt one of these rather that start from scratch. In the past, databases of speech which contained hand labelled phone or phoneme transcriptions were highly valued, but
- 228. 204 Chapter 8. Pronunciation as phonetic/phonemic transcription can now be done automatically, today by far the most useful resource is in fact a comprehensive lexicon. As the development of a large ( 100,000 word) lexicon can take years of work, any existing resources greatly help. Lexicons are in general tied to a specific accent of the language, and as such come with a phoneme inventory pre-defined. So in many cases a phoneme inventory may exist for the accent we are interested in. By and large, most phoneme inventories developed for computational work roughly follow the guidelines just described. As just pointed out, there are however several ways to resolve some of the issues, and so it is worth examining any predefined phoneme inventories in light of the above discussion to assess suitability. Most predefined phoneme inventories have not been developed with TTS in mind, rather they were developed for purposes of speech recognition or are just “general” phoneme inventories developed for a variety of uses. Experience has shown that minor modifications in predefined phoneme inventories can sometimes be beneficial for TTS purposes. The most common problem is that sometimes the phoneme inventory has some “phonetic” features (such as markers for aspirated or unaspirated, or a different label for the stop closure than the stop release), but these can normally be removed with some automatic processing. Table 8.1 gives a definition of a complete phoneme inventory based on a modification of the TIMIT phoneme inventory. This was originally designed for use with the acoustic-phonetic database of the same name and now one of the most well known phoneme inventories for General American [173]. Most of the modifications we have made have been to eliminate some of the purely phonetic distinctions in the original TIMIT inventory1. The main peculiarity of the TIMIT inventory relates to the fact that sequences of schwa followed by /r/ are represented by a single phoneme /axr/, and the vowel in words such as BIRD is treated similarly and represented as /b er d/. All other combinations of vowel and /r/ are treated as two separate phonemes. Table 8.2 gives a phoneme inventory for the British English RP accent, based on the University of Edinburgh’s MRPA (machine readable phonetic alphabet) system [276]. The consonant system is more or less the same; this is what we would expect from the facts of English accent variation. The vowel system is different, and in particular the treatment of “r” is different because RP is a non- rhotic accent. The pseudo-diphthong vowels /i@/, /u@/ and /e@/ are used in cases where there 1 For the record, the differences between this and the standard phone inventory are: 1. The /axr/ unstressed schwa+/r/ phoneme has been replaced by /er/. /er/ and the old /axr/ are now distinguished by stress alone 2. the close phones bcl, tcl etc have been removed 3. the nasal tap /nx/ has been merged with the /n/ nasal 4. /hv/ has been merged with /hh/ 5. the glottal stop phone /q/ has been removed 6. the syllabic consonants /en/, /em/, /eng/ and /el/ have been replaced by /ax/ + consonant sequences, e.g. /en/ → /ax n/ 7. /ix/ has been replaced by either /ax/ or /ih/, /ux/ by /uw/ and /ax-h/ by /ax/
- 229. Section 8.2. Formulating a phonological representation system 205 Symbol Example word Transcription Symbol Example word Transcription b BEE B iy l LAY L ey d DAY D ey r RAY R ey g GAY G ey w WAY W ey p PEA P iy y YACHT Y aa t t TEA T iy iy BEET b IY t k KEY K iy ih BIT b IH t dx MUDDY m ah DX iy eh BET b EH t jh JOKE JH ow k ey BAIT b EY t ch CHOKE CH ow k ae BAT b AE t s SEA S iy aa BOT b AA t sh SHE SH iy aw BOUT b AW t z ZONE Z ow n ay BITE b AY t zh AZURE ae ZH er ah BUTT b AH t f FIN F ih n ao BOUGHT b AO t th THIN TH ih n oy BOY b OY v VAN V ae n ow BOAT b OW t dh THEN DH e n uh BOOK b UH k hh HAY HH ey uw BOOT b UW t m MOM M aa M er BIRD b ER d n NOON N uw N ax ABOUT AX b aw t ng SING s ih NG axr BUTTER b ah d axr Figure 8.1 Definition of modified TIMIT phoneme inventory for General American, with example words “would be” an /r/ if the accent was non-rhotic. For the remainder of the book, we will use the modified TIMIT and MRPA phoneme inventories for General American and RP English accents respectively. 8.2.6 Phoneme names The above has focused on minimal pair analysis, which is the major task in defining a phoneme in- ventory. In defining the phoneme inventories shown in Tables 8.1 and 8.2 we have also performed an additional task in that we have explicitly given a name to each phoneme. The processes of min- imal pair analysis and naming are not connected; the names in theory could be anything, such that we could just name the discovered phonemes as 1, 2, 3 ... or a, b, c or anything else. In choosing a naming system, we have first made the decision to use plain ascii characters for all phonemes. The TIMIT phoneme inventory is exclusively alphabetic; the MRPA phoneme inventory is alphabetic apart from the use of “@” for sounds with schwa. As there are more phonemes that alphabetic
- 230. 206 Chapter 8. Pronunciation Symbol Word initial example Word final example Symbol Word initial example Word final example p PIP RIP j YIP - t TIP WRIT ii HEAT PEEL k KIP KICK ih HIT PILL b BIP RIB ei HATE PATE d DIP RID e HET PET g GAP RIG aa HEART PART ch CHIP RICH a HAT PAT dz GYP RIDGE uu HOOT POOT m MIP RIM u FULL PUT n NIP RUN ou MOAT BOAT ng * RING uh HUT PUTT f FOP RIFF oo HAUGHT CAUGHT th THOUGHT KITH o HOT POT s SIP REECE ai HEIGHT BITE sh SHIP WISH au LOUT BOUT v VOTE SIV oi NOISE BOIL dh THOUGH WRITHE i@ HEAR PIER z ZIP FIZZ e@ HAIR PEAR zh VISION u@ SURE PURE h HIP - @@ HER PURR l LIP PULL @ CHINA ABOUT r RIP CAR w WIP - Figure 8.2 Definition of British English MRPA phoneme inventory, with example words characters, we use two and sometimes three characters for a single phoneme. Often this is when we have a sound with two phonetic components such as /ch/ or /oy/, but this is also used for simple phonemes such as /th/ and /sh/. In naming these phoneme inventories we have also not worried whether the “same” phoneme in one accent has the same name in another accent. As we have no need to mix phonemic tran- scriptions across accents in normal synthesis, this is unproblematic (see however Section 17). Strictly speaking, the principles of phonemic analysis imply that this is a meaningless comparison anyway: phonemes only have status as units of contrast for a single accent, and so it is of no con- sequence whether we use the same symbol for the vowel in BUY in British or American English. In practice, it is clear from the TIMIT and MRPA definitions that the vowel sets are different but the consonant sets are nearly identical (this of course partly reflects the fact that vowel contrasts are the main source of phonemic differences between accents in English). We have also avoided the possibility of naming the phonemes after one of their allophones in the IPA alphabet. This is because it is generally easier to use a separate and self contained symbol set for phonemes and not
- 231. Section 8.2. Formulating a phonological representation system 207 use IPA symbols as this can lead to confusions about the phoneme/phone divide. In both the TIMIT and MRPA phoneme inventories, we have only used lower case letters. To some extent this is a purely aesthetic issue as we wish to avoid upper case (which has a somewhat “aggressive” quality) and mixed case (which can look ugly). There is an additional benefit in that upper case equivalents of the symbols can be used in cases where some emphasis is required, such as stress or other highlighting. In the above, we have chosen though to give simple mnemonic names to each phoneme, so that the first phoneme of TIP is named “t”, the first of SHIP is named “sh” and so on. Mnemonic names are usually helpful as we often have to enter words in the lexicon ourselves or perform transcriptions, and this is easier if we can easily recall the symbols to use. These names can sometimes trip us up, especially with vowels, so a little care should be taken. 8.2.7 Syllabic issues We have shown previously that in addition to simple or “linear” phonemic context, a major factor in determining the realisation of phonemes is their place in the syllabic structure. Phonemes at the starts of syllables or words can be realised differently from ones at the ends; the position of a phoneme in a consonant cluster is important, and stress effects, which are expressed at sylla- ble level, also have an important bearing on phonetic realisation. We therefore have to give the waveform generation component access to this syllabic information. If we actually know where the syllable boundaries lie within a word, it is a simply matter to assign the type of tree structures described in Section 7.4.4. Quite simply, the vowel serves as the nucleus, the consonants after the nucleus up until the end of the syllable are assigned to the coda, and the consonants between the vowel and the syllable start are assigned to the onset. In this way, the syllable structure is found deterministically from the syllable boundaries and phonemes. As we saw in Section 7.4.4, syllable boundaries can in the most part be determined by using the maximal onset principle. In the few cases where this is not so (recall our BOOKEND example) the boundary needs to be marked explicitly. As system designers we therefore face some choices. On the one hand, we can choose to explicitly mark every syllable boundary in the lexicon. This is costly however; many existing lexicons do not have syllable boundaries marked and the additional effort in marking these may be prohibitive. Alternatively we could determine the syllable boundaries automatically by application of the syllable onset rule. In doing so we will of course get the syllable boundaries wrong for the difficult cases just mentioned. A third possibility is a compromise between the two, where the difficult transcribed are marked by hand, and where the rest are transcribed by algorithm. For languages such as English, three levels of stress are necessary. While some phonological systems use more (e.g. SPE), our experience has shown that, at the lexical level at least, three is clearly sufficient. Following from a convention adopted for TIMIT, we use 1 for primary stressed syllables, 2 for secondary stressed syllables and 0 for reduced syllables. The stress numbers are placed adjacent and after the vowel, so, for example, the entry for INFORMATION would be /ih2 n . f ax0 . m ey1 . sh ax0 n/.
- 232. 208 Chapter 8. Pronunciation All syllables receive a stress mark, and hence most vowels will be marked 2. All words except function words must have a single primary stressed vowel. All normal schwa vowels (/ax/) receive a 0. There are a number of possibilities regarding whether to transcribe normal vowels with a 0 or a 2. One option is to regard 0 and 2 as weak and strong versions of one another, such that any vowel could receive either 0 or 2. Another option is to regard 0 as being only for schwa vowels; this means that in effect every other vowel is a 2 regardless. This last option has the benefit of simplicity, but some feel that this disregards phonetic facts. In traditional phonetics, the last vowel of HAPPY is quite often transcribed as /ih 0/ or /iy 0/. In addition, many think that it is most appropriate to transcribe the last vowel of HABIT as /ih 0 / giving /h ae1 b ih0 t/. Here we again face the issue of phonological organisation versus phonetic explicitness. Either option can be taken, but it is worth noting that direct experience showed that it was possible to take the 0 for schwa option everywhere, and label all other vowels as 2. This would then given /h a1 p iy2/ and /h a b ax0 t/. Finally, we come to the issue of phonotactics. This is an important part of defining a phonol- ogy; there is little point in spending time defining a phoneme inventory which define the paradig- matic relations if we then ignore the phonotactics which define the syntagmatic relations. Defining a phonotactic grammar is necessary for a number of practical reasons. Firstly, we need to define what we think the synthesiser should be able to say in terms of phoneme sequences. Even if we create a very simple grammar, that states for instance that /ng/ can’t occur at the start of a word, we can find any case where this is passed to the synthesiser by mistake and take corrective action. Secondly, we need to keep a check of the entries in the lexicon; as these entries are made by hand there is a chance that errors will be made. Hand in hand with the phonotactic grammar comes a set of conventions that we must establish regarding how words should be transcribed. Finally, having an explicit phonotactic grammar helps in database collection as we can easily tell how complete a particular database is with regard to phonemic sequence coverage. 8.3 THE LEXICON In our introduction to the nature of language, we explained that one of its fundamental properties is that there is an arbitrary relation between the meaning and form of a word. A word’s pronunciation cannot be inferred from its meaning, nor vice versa. Obviously for a language to work, these arbitrary relations need to be somehow stored, and the device used to do this is call a lexicon. We can talk about a speaker’s lexicon, meaning the lexicon that a particular individual uses to store his or her words; we can also talk about the language lexicon which is a store of all the meaning/form pairs (words) in that language (a traditional dictionary is this type of lexicon). For language engineering purposes however, we talk about the computer lexicon, which is a physical computer data object containing known descriptions of words. In machine understanding and generation systems, or any systems which deal with seman- tics and meaning, the lexicon is a vital component in a system as it is here that the connection
- 233. Section 8.3. The Lexicon 209 between meaning and spoken or written form is stored. In speech synthesis and recognition tasks however, the meaning element is largely ignored, so in principle the lexicon is not strictly speaking a necessary component. However, lexicons are in fact widely used in synthesis and recognition to store the correspondences between the two main forms of the word; that is the spelling and pronunciation of a word. As such these are called orthography-pronunciation lexicons. 8.3.1 Lexicon and Rules While the relation between the meaning and form of all languages is completely arbitrary, in general the relation between the two main forms, i.e. the spoken and written form, is not. The two main forms do not have an equal status; speech is the primary form of the language and writing is a secondary, invented form designed to record the spoken form. The degree to which these forms are related is dependent on many factors, and it is clear that in some languages the written form closely follows the spoken form (e.g Spanish), in others the relationship is discernible but complex (e.g. English and French) while in others it is nearly as arbitrary as the meaning/form relationship (e.g. Chinese). Hence for native words in some languages it is possible to infer one form from the other, and this mapping can be performed by a grapheme-to-phoneme algorithm. We will leave a discussion of these until Section 8.4, but for now let us assume that we can build such a system that always generates a pronunciation form from an orthographic form, even if the accuracy of doing this can be quite low for some languages. Due to this low accuracy of these rules for some languages we have to store the correspondences in a lexicon. It is important to realised that even for the more direct languages, the widespread use of English loan-words, technical words and acronyms means that at least some use of a lexicon is required. Most text-to-speech systems use a combination of a lexicon and rules to perform pronun- ciation. The justification for this two pronged strategy has changed somewhat over the years. In traditional systems, the explanation usually went something like this: “grapheme-to-phoneme rules are used as the main pronunciation mechanisms, but in addition we use an exceptions dic- tionary which deals with cases which don’t follow the normal patterns of the language”. Thus it was rules which were primary, and the lexicon was used for cases where the rules failed. The reasons for the primacy of the rule system were that, firstly, this was still in the days when using “knowledge” was a respectable way of building all types of speech, language and AI systems, and so using rules was a natural way that anyone would approach this problem. Secondly, memory space was limited, such that storing tens of thousands of words was prohibitive. Note that using rules to save memory typical incurs a processing cost, but at the time speed was not generally seen as a critical issue (ie, it was simply impossible to have enough memory to store the words, but it was possible to use slow rules, even if it resulted in a non-real time system). Finally, it takes time to develop a quality lexicon, and in the past there simply weren’t the resources and experience to have built a comprehensive lexicon. More common today is the justification “We use a lexicon to store pronunciations, and resort to rules for missing entries in the lexicon.” Hence the situation is reversed; the lexicon is seen as primary and the rules are there as backup when the lexicon fails.
- 234. 210 Chapter 8. Pronunciation The reader should be aware that there is still a lingering hangover in the literature from the earlier period regarding the rules vs lexicon issue. Rules are often inherently favoured over the lexicon and one often comes across statements stating that there is not enough memory to store all the words in a language. Such reasoning is rarely solid, as no decision can be taken about this until the actual size of the lexicon and the processing cost of the rules are quantified. For smaller footprints systems (say 10 megabytes), lexicon size will definitely be an issue, but for larger systems the size of the unit selection database will far out weigh the size of the lexicon. One also finds that alongside this we also see justifications that rules are a more sophisticated approach and that simply storing words in the lexicon is a dumb brute force approach. As we shall see below, this is a false argument; rules and lexicons are simply different techniques which trade-off memory and processing cost; both have their place and one is not a priori better, more sophisticated or even more accurate than the other. One occasionally encounters the opposite view, in which lexicons are seen as the only true way to handle pronunciation, and hence the use of very large lexicons is advocated. This can be dangerous too; while most people are aware that rules often make mistakes there is sometimes a failure to realise that significant errors can exist within the lexicon also; too often this possibility is ignored and the lexicon is taken as the absolute truth. Most accounts of the lexicon and rule setup in a TTS system would give a reader the idea that these were two completely separate mechanisms, which just happen to have been brought to bear on the same problem. In fact, the distinction between lexicon and rule approaches is not as clear cut as it seems, and in moderns TTS systems, it is best to think of what we normally think of as “lexicons” and “rules” as being at certain points on a scale. From the outside, the distinction seems clear; a lexicon is a declarative data structure (e.g. a list) which contains entries, a rule set is a series of processes which examine the grapheme sequence and generate a phoneme sequence. The distinction is however not always clear cut. Most rule systems (see below) use a context window on the orthography to help disambiguate which phoneme should be generated. If we take a word such as IRAQ for instance, it is quite possible that no other words with anything like this grapheme pattern will be found; so if the rules correctly predict this word, can we not say that the rules have learned the pronunciation for this and only this word, and that the rules have in fact implicitly stored the pronunciation? In machine learning, this phenomenon is called memorising the data: this is often consid- ered harmful if the system learns exactly those examples it was given but not any generalities; if this is the case we say that the model has over fitted the data. For a trainable model with enough parameters, it is entirely possible that it will end up memorising all the training data rather than learn the generalisations behind it. This is normally seen as an undesirable property in a rule sys- tem, but in our case, it can be argued that this is not strictly so. In most machine learning problems, the population of data is effectively infinite, and the training data is a small sample of the popula- tion. As the sample is not identical to the population care must be taken not to learn it too exactly, otherwise the trained algorithm will not perform well when presented with different samples from the population. As we shall see below, learning grapheme-to-phoneme rules is somewhat differ- ent, in that the population (all existent words) is large but finite, and there is a good chance of
- 235. Section 8.3. The Lexicon 211 encountering exactly the same words in different samples. Hence memorising common examples is not in itself harmful [76]. 8.3.2 Lexicon formats In speech technology, it is quite often to find existing lexicons in what we can term simple dic- tionary format. Such lexicons can be stored as simple ascii files, where each entry starts with the orthography, followed by the pronunciation and possibly other information. The entries are ordered by the orthography such that a few typical entries would look like this: deep d iy1 p deer d ih1 r defend d ax0 f eh1 n d define d ax0 f ay1 n degree d ax0 g r iy1 delicate d eh1 l ax0 k ax0 t delight d ax0 l ay1 t demand d ax0 m ae1 n d density d eh1 n s ax0 t iy2 It is clear that this format explicitly gives priority to the written form, as this is used to order and in- dex the entries. Dictionary formats can be problematic though, for the reason that the orthographic form of a word is not unique; homographs are fairly frequent and these must somehow be handled. Two common amendments to this style are to have multiple lines with the same orthographic form project p r aa1 jh eh2 k t project p r ax0 jh eh1 k t or to use more than one entry for the pronunciation on each line project p r aa1 jh eh2 k t, p r ax0 jh eh1 k t Even if this is done, we still somehow need to know that /p r aa1 jh eh2 k t/ is the noun form of project, it is not sufficient to say that project simply has two different pronunciations. While POS markers could be added, we soon find that when we also consider the issue of pronunciation variants (which are arbitrarily different ways of pronouncing the same word) or orthographic vari- ants (different ways of spelling the same word) we see that the simple dictionary style comes under considerable pressure, and looses its main positive feature, that it is easy to read and process. A solution to this is to build the lexicon as a relational database. In this, we have exactly one entry for each word. Each entry contains a number of uniquely identified fields, each of which has a single value. For for a simple word, the entry may just contain two fields, ORTHOGRAPHY and PRONUNCIATION. It is a simple matter to add more fields such as POS or SYNCAT. Each entry in a relational database can also be seen as a feature structure of the type we are now familiar with, and because of this similarity, we will use the feature structure terminology for ease of
- 236. 212 Chapter 8. Pronunciation explanation. Hence some simple words would look like ORTHOGRAPHY defend PRONUNCIATION d ax0 f eh1 n d ORTHOGRAPHY degree PRONUNCIATION d ax0 g r iy1 ORTHOGRAPHY define PRONUNCIATION d ax0 f ay1 n In this the orthography has no preferred status, and the entries can be indexed by the orthography or pronunciation field. The main advantage of using this style is that it is straightforward to add more fields to the entries while still maintaining a well defined structure. Hence we can add POS information to distinguish between homographs ORTHOGRAPHY project PRONUNCIATION p r ax0 jh eh1 k t POS verb ORTHOGRAPHY project PRONUNCIATION p r aa1 jh eh2 k t POS noun Throughout the book we have been using a convention whereby we give a name to each word, for instance PROJECT-NOUN for the last entry above. We should state that this name has no objective status; it is simple a “handle” that allows us to uniquely identify the word. We normally use the spelling as the handle and add some additional identifier if this isn’t sufficiently unique - again, this is just a convention, the handles could be 1,2,3,... or a,b,c,.... It is up to the system designer whether to include this in the entry or not: including it means that there is one unique value for each word in the lexicon, and this can have useful housekeeping functions. On the other hand, the developer needs to create these handles and ensure their uniqueness. For demonstration purposes
- 237. Section 8.3. The Lexicon 213 we will include them in the following examples, and so the above would look like NAME PROJECT-VERB ORTHOGRAPHY project PRONUNCIATION p r ax0 jh eh1 k t POS verb NAME PROJECT-NOUN ORTHOGRAPHY project PRONUNCIATION p r aa1 jh eh2 k t POS noun NAME DEFEND ORTHOGRAPHY defend PRONUNCIATION d ax0 f eh1 n d POS verb It is often useful to make use of lists as values as we want to describe the full range of possibilities without necessarily ascribing any causes to them. For example, many noun/verb pairs don’t differ in orthography or pronunciation, and so for our purposes we treat these as a single word (see Section 4.2). We still wish to record the fact that the word can have several parts of speech; this may for instance be used by the POS tagger NAME WALK ORTHOGRAPHY walk PRONUNCIATION w ao1 k POS {verb,noun} Orthographic and pronunciation variants can be handled this way also: NAME EITHER ORTHOGRAPHY either PRONUNCIATION {iy1 dh er0, ay1 dh er0} NAME COLOUR ORTHOGRAPHY {colour, color} PRONUNCIATION k ah1 l axr
- 238. 214 Chapter 8. Pronunciation 8.3.3 The offline lexicon Now let us consider the issue of certainty or knowledge with regard to pronunciation. To help in this, let us make a distinction between two types of lexicon. The first is the system lexicon: this is the database of word pronunciations actually used by the system at run time. The other is the offline lexicon; this is a database of everything we know about the words in our language: it is the sum of our recorded knowledge. The offline lexicon has all sorts of uses; we use it to label speech databases, to train the POS tagger, for help with text analysis and so on. The system lexicon is created from the offline lexicon, and there are a range of choices about what entries and fields to include. We should stress that regardless of the needs and requirements of the system lexicon, it is highly beneficial to have as large and comprehensive an offline lexicon as possible. When we consider the offline lexicon, it is not the case that a given word is strictly “in” or “not in” the lexicon. For instance, we may have entries that are incomplete, in that we know the values for some fields but not others. Most likely there will be missing words also, that is words that we know of but which we haven’t entered. Not every value in every field will be correct; errors may occur or there may be entries which we are not entirely certain of. More formally, we can evaluate our off-line lexicon in terms of coverage, completeness, and quality: Coverage This is a measure of how many words in the language are in the lexicon. For every word in the language, we can assign it to one of three categories: known Words which are present in the lexicon missing Words which exist, are known to the developers, but which are are not present in the lexicon. unknown Words which exist, are not known to the developers and are not in the lexicon. Completeness Every lexicon entry has a number of fields, but some of these may be unfilled. Quality A measure of how correct the values in an entry are. What is perhaps less clear is that these measures can be traded off against one another. Recall that we said (Section 4.2) that for our purposes the derived forms of a word are also considered words in their own right, so WALK, WALKING and WALKED are all distinct words. It is quite common for dictionary writers to only define the main, base form, so that if we come across an existing lexicon it is quite likely that we would find WALK but not WALKING or WALKED. We can easily create the empty entries for WALKED and WALKING, and in this way we have improved our coverage, but have left these entries incomplete as we have only specified the orthography, not the pronunciation. To complete these entries, we can use rules to add the pronunciations, for example by simply concatenating /ih0 ng/ to the end of the stem to create WALKING. In doing so, we don’t however know if the rule generated the correct form; so at this stage, the quality of the new entry is unknown. Finally by performing a manual check, we can tell if the rule has operated correctly. In this example we have seen how our knowledge of missing words (that is words known to us, but not in the lexicon) can be used to improve coverage. We have also seen that we can use rules to improve the completeness of the lexicon, but in doing so we don’t necessarily know if the rule
- 239. Section 8.3. The Lexicon 215 has generated the correct answer. This is one example of the way the rules and lexicon can interact. We can generalise this approach to providing more extensive coverage and completeness. With regard to the unknown versus missing distinction, with a well developed lexicon the most likely missing entries are of the type just mentioned, that is absent derived forms of known words: we can improve our coverage of missing words by the process just outlined. By contrast we can be sure that the number of unknown words is vast; this will include a huge number of technical or medical terms and of course many names. For unknown words, the task is really to find these words from sources outside our personal knowledge. One way of doing this is to conduct a search for unique words in any text corpus, or any collection of text material (e.g. the internet). Given such a list, we have effectively improved our coverage but done so in such a way that the completeness for all these new entries is limited to an orthographic definition. We can improve completeness as before, but this time we have to generate the base form of the pronunciation as well as any derived forms. We can find the base form by a grapheme-to-phoneme algorithm, and use the morphological rules to find the derived forms, and in that way general a complete lexicon. Our only issue now is with quality; as we have used a grapheme-to-phoneme algorithm there is a good chance that many entries will be completely wrong. All the automatically generated entries can be labelled as such, so that we know not to have the same degree of confidence in these as the other hand checked entries. In due course all these can of course be checked and edited so as improve the certainty and quality of the lexicon. These processes show the value of the off-line lexicon; it is in effect the full depository of our knowledge about words in the language. Creating this is highly beneficial in that gives us a full database of words from which to generate the system lexicon, to train or develop a grapheme- to-phoneme algorithm, and for other uses mentioned above. 8.3.4 The system lexicon If we generate the offline lexicon by the above process we will have a lexicon that is very large, but will have a considerable number of entries whose quality is unknown. In traditional TTS, the normal approach was to only have entries which were carefully checked in the lexicon, with the idea that the rules would be used for all other cases. However, it should now be clear that whether we use the rules at run-time to handle words not in the lexicon, or use the rules offline to expand the lexicon will not have any effect on the overall quality: all we are doing is changing the place where the rules are used. Given this large offline lexicon, we see then that the real debate about rules vs lexicons is not one of quality, but rather one of balance between run-time and off-line resources. If we take the case where we include the entire off-line lexicon in the system lexicon, we will have a system which uses a considerable amount of memory, but where the processing speed is minimal (simply the small amount of time taken to look up a word). If on the other hand we create a system lexicon that is only a small subset of the offline lexicon, this will result in a smaller footprint, but as the pronunciation of absent words will have to generate at run-time, the processing costs
- 240. 216 Chapter 8. Pronunciation could be considerable. One very useful technique for lexicon reduction is to take a set of rules, run them over all the words in the database and record which entries agree with the pronunciations generated by the rules. If we then delete all these entries, we have saved considerable space but will generate exactly the same quality output. In fact this technique is just the opposite of what we did before where we used the rules to create pronunciations for incomplete entries. There we used the rules to create the largest possible lexicon, while now we have used the rules to create the smallest possible lexicon. In both cases the quality of the output is the same. In fact, there are a number of possibilities beyond this. Considering the last point again, what we did was to use the rules to reduce the size of the lexicon, in effect the rules summarise the redundancy in the lexicon. This technique can be thought of as lossless compression; this is, the output is exactly the same as when the full lexicon is used, its just that the generalities of the lexicon have been exactly captured in a set of rules. If we desire a smaller system still, we can opt for lossy compression where the output is not the same as before and can be expected to contain some errors. After we delete the entries which the rules correctly predict, we further delete some less frequently occurring entries, resulting in a smaller system, but one in which a few more words will be wrong. We can tune this exactly to our space requirements right down to the point where we have no run-time lexicon at all and simply rely on the rules every time. We can move in the other direction too; if we have some additional memory and we find that the “decompression time” needed to use the rules is too slow, from a lossless compression system we can add in some correctly predicted common words back into the lexicon. This will have the effect of increasing memory but speeding the system up. 8.3.5 Lexicon quality As the lexicon is the repository of all knowledge regarding the orthography and pronunciation of words, it is vital that the quality of the entries that it contains is as high as possible. All too often this issue is overlooked, and unfortunately many lexicons that are used today have significant numbers of errors. It should be obvious enough that if the pronunciation of a word is entered incorrectly in the lexicon, when the TTS system comes to say that word it will sound wrong. Often listeners are completely thrown by this; they don’t perceive that the correct word has been spoken wrongly, rather they suffer a complete loss of comprehension. Other consequences arise from errors in the lexicon; if this is used to label the speech database, we might find a case where an instance of an /uh/ vowel is wrongly labelled as /uw/. When this unit is used in synthesis to saw a word such as /b uh k/ the result will be /b uw k/; again throwing the listener. This is even more serious than the first case because this will affect words other than the one where the mistake occurs. Personal experience has shown, that after the recording of a high quality speech database, the quality of the lexicon is the single biggest factor in determining overall TTS quality. A certain amount of skill is required to add entries by hand; the primary requirements are that one should have a familiarity with the pronunciation scheme being used and a knowledge of the accent in question. In many cases the person adding or checking entries has the same accent as the
- 241. Section 8.3. The Lexicon 217 lexicon itself, but this isn’t always the case, and so particular care needs to be taken when adding entries of a different accent; it is all too easy to mistakenly think we know how someone else says a word. That said, it is certainly not the case that one needs to be an expert phonetician; personal experience has shown that most developers can perform this task if given adequate training. Many errors result from a small number of basic causes. A primary source of error is that people make the pronunciation too close to the orthography. Inexperienced developers could for instance add /e/ phonemes at the ends of words where the orthography has an e, such that we would get /h ey t e/ instead of /h ey t/ for HATE. Other simple errors include using /j/ for /jh/, using /c/ for /k/ and so on. One simple trick is to find an existing checked word that we know rhymes with the word we are about to enter; so that when entering HATE we look up LATE and simply change the first phoneme. This is particularly useful when working with an accent one isn’t sure of. Another is to use an existing version of a synthesiser to playback the word synthesised with the new transcription; this technique very quickly exposes errors such as adding an /e/ to the ends of words. Finally, we can build an automatic pronunciation verification system, that simply checks whether the pronunciations for a entry is legal within the phoneme inventory, stress system and phonotactic grammar definitions. Those more experienced in phonetics often make the opposite mistake of creating lexicon entries that are too close to the phonetics. Here we see cases where many multiple variants are added because of a slight perceived difference or possibility of difference in a pronunciation. For the word MODERATE we might get /m o1 d ax0 r ax0 t/, /m o1 d r ax0 t/, /m o1 d eh0 r ax0 t/ and so on all being added because of perceived differences. In general, the tendency to “split hairs” and add many pronunciation variants for each word should be avoided. Even if two speakers do seem have have a difference in how they pronounce a word, it should be remembered that the phonemic representation is only approximate in any case, and can not be expected to cover all differences. Using too many variants can be harmful as it leads to a fragmentation of the data, and makes it difficult to learn and generalise. 8.3.6 Determining the pronunciation of unknown words Even with experienced lexicon developers, severe difficulties can arise because of unknown words. Recall that these are legitimate words which exist in the language but which the developer is not personally familiar with. To see the extent of this problem, take any large (i.e. unabridged) dictionary, open it at random and have a look - it usually doesn’t take long to find unfamiliar words. The profusion of words unknown to us is quite considerable as we see the dictionary is full of medical, chemical and nautical terms, old words that have fallen out of usage, and so on. If we then add to these the issue of names we see that the problems mount up. For many unknown words, even though the word’s meaning and form is unfamiliar to us, the patterns of letters seems familiar, and because of this we have a reasonable chance of guessing the correct pronunciation. Many are in groups of words that may be personally unfamiliar to the developer, but which are common enough for other people. Sailing terms are a good example of this, where for example,
- 242. 218 Chapter 8. Pronunciation most people can correctly read aloud bobstay, cringles and jibe even they don’t know what these words mean. Other terms are ore difficult however, such that most struggle with how to read aloud fo’c’sle boatswain and gunwale if they are not familiar with the terms. For such groups of words, the only answer is to consult a knowledgeable source; either a book on sailing or a sailor in person. In considering this issue, we come across an interesting philosophical issue regarding the pronunciation of words. Quite simply, how do we know what the correct pronunciation for any word is? Saying that word pronunciations are properly defined in traditional dictionaries avoids the point; this would imply that words not in a dictionary have no standard pronunciation, and misses the fact that dictionary writers try and listen to real speakers to elicit the actual pronunciation [484]. The best we can do is say that the “correct” pronunciation for a word is X because that is what the population of speakers believe. For most common words, this is unproblematic; and everyone of a particular accent will say the words WALK, DOG and HOUSE the same way. Problems do arise though because people do not in fact always pronounce words the same way, even within a single accent. The issue is whether we can establish a correct pronunciation, relegating the others to some “incorrect” status, or whether we have to accept multiple correct forms. Establishing the boundaries here is difficult, it seems equally absurd to insist on a “gold standard” of a single correct pronunciation, or an “anything goes” policy where all pronunciations are equally legitimate. There is a personal element in this as well; it doesn’t bother me whether someone says CONTROVERSY as /k o1 n t r ax0 v er2 s iy/ or /k ax0 n t r o1 v ax0 s iy2/, but somehow when someone says /n uw1 k uw2 l ax0 r/ for NUCLEAR this just seems wrong. To take another example, it is now quite common to hear American English speakers to pronounce the “silent” /l/ in words like CALM and SALMON - while this may seem clearly “wrong”, how can it be if millions of people now follow this habit? Difficulties most arise with regard to rare names. In most European countries, the native pronunciation of names is well established, and even if a name is borrowed from one language to make a new name in another language, we can often find the correct pronunciation in the original language. The situation in the United States is significantly different, in that the sheer diversity of names outweighs any other country. In the US, we see firstly that many names have moved considerably from their original pronunciation, but secondly many names are genuinely made up or so altered from the original that they effectively became new names2. This situation creates many cases where there is only a handful of families with a particular name, making it difficult for the lexicon developer to find the pronunciation. Even more problematic is that we find cases where a handful of families share the orthography for their name but have different pronunciations. It is often unknown even if these names share a common origin. How then do we determine the correct 2 Note that there is a consistent myth in that peoples; names were changed by officials during transit through the Ellis Island immigration service. What evidence we have suggests that this is not in fact the major source of new names, as some degree of consistency had to be kept between embarkation in the origin country and immigration procedure. Many minor changes in spelling and so on certainly happened, but it was not the case that immigration officials simply invented new names because they thought the real name was too awkward to handle.
- 243. Section 8.4. Grapheme-to-Phoneme Conversion 219 pronunciation for such a name? As Spiegel [409] says: Complicating the issue is that everyone is a pronunciation expert for their own name, and (only slightly less so) for the people they know. Sometimes people are unaware of other (valid) pronunciations for their names, even when the pronunciation is more common. Hence even if we decide that the correct pronunciation for smith is /s m ih1 th/ we will still find enough people who complain that the system got it wrong because their name is pronounced /s m ay1 th/. There is no objective real solution to this problem as this is one of system usage and user requirements. The real point is that developers who deploy TTS systems should be aware that the may not be any objective truth as to whether a system is saying a name correctly or not. 8.4 GRAPHEME-TO-PHONEME CONVERSION A grapheme-to-phoneme (G2P) algorithm generates a phonemic transcription of a word given its spelling. Thus it generates a sequence of phonemes from a sequence of characters3. 8.4.1 Rule based techniques Grapheme-to-phoneme conversion by rule is perhaps the classic application of traditional knowl- edge based rules in TTS. Indeed, in the past it was thought that this was how humans performed the task; when presented with an unknown word they would apply a set of rules to find the pro- nunciation. The most common approach is to process the character sequence left-to-right, and for each character apply one or more rules in order to generate a phoneme. It should be clear that these rules can’t operate in isolation, otherwise a t character would always result in a /t/ phoneme. Rather the character context of the character is used and in this was we can generate /th/ when t is followed by h. Most approaches use context sensitive rewrite rules of the type introduced in Section 5.2.4. Recall that these are of the form A → B/d/C which reads as “A is converted (rewrites to) d when preceded by B and followed by C”. In any well developed system, multiple rules will match a given context, and so a means of choosing amongst rules is required. A common technique for this is to order the rules and apply each in turn, so that 3 In this book, the use of the term grapheme-to-phoneme conversion relates specifically to algorithms which perform just this task; elsewhere this term is sometimes used for the process which uses algorithms and the lexicon, and indeed sometimes used for the entire process of text analysis. The term letter-to-sound rules is also encountered, but this is increasingly misleading as most modern techniques don’t used rules in the traditional way. A final complication regarding the terminology of this field is that strictly speaking a grapheme is a minimal unit of spelling rather than a single character, so in that the th is described as a single grapheme. As our input is always a sequence of characters, we are really performing character-to-phoneme conversion.
- 244. 220 Chapter 8. Pronunciation the first matching rule is the one chosen. As this is a purely hand written technique the developer has to order these rules him or herself. A common technique is to order the rules with the most specific first. Rule systems of this kind include [8], [307], [155], [226], [136]. These systems can often produce quite reasonable results for highly regular systems with a shallow orthography (e.g. Spanish). The results are generally very poor though for languages like English and French, which has increasingly led to the adoption of data driven and statistical techniques [300]. 8.4.2 Grapheme to phoneme alignment All the data driven techniques described below use algorithms to learn the mapping between char- acters and phonemes. They use an existing lexicon as training data, and often words in this are held out as test data and so these data driven techniques are only usable in cases where there a comprehensive lexicon is available. In general words do not have a one-to-one correspondence between characters and phonemes: sometimes two or more characters produce a single phoneme (e.g. gh → /f/ in rough) and sometimes they don’t produce a phoneme at all (e.g. the e in hate → /h ey t/). A pre-requisite for training then is to align the characters with the phonemes for each entry in the lexicon. Most techniques for performing this are based on a dynamic time warping, (DTW) tech- nique [343]. This works by assigning a cost to every phoneme that can align with a grapheme. “Good” correspondences are given low costs while “bad” correspondences are given high costs. The DTW technique is a dynamic program algorithm which effectively searches all possible align- ments and picks the one which has the lowest total cost. This is a significantly easier task than either writing the rules by hand or aligning every word in the lexicon by hand. For example, the character s maps to the phoneme /s/ with a low cost, but maps to /n/ with a much high cost. In many cases there is an obvious phoneme which maps from the character, but in other cases, e.g. h this is used to represent many different phonemes (/sh/, /th/, /dh/ etc) in writing and so has a number of low costs. A probabilistic version of this was used by Jelinek [239], who used the hidden Markov model Baum-Welch training technique (described in Section 15.1.8 to learn the probability that a character would generate a phoneme. This was extended by Taylor [433] who showed that it was possible to do this from a flat start meaning that no initial seeding is necessary. In general, the alignment method is independent of the actual algorithm used, so that any alignment technique can be used with any of the algorithms described below. 8.4.3 Neural networks The first data driven G2P algorithm was the NetTalk algorithm developed by Sejnowski and Rosen- berg [397], [396], [377]. This work became genuinely famous, not for its raw performance or use in practical text-to-speech, but because it was one of the first real demonstrations of a neural net-
- 245. Section 8.4. Grapheme-to-Phoneme Conversion 221 work in operation. A popular demonstration of the system speaking the pronunciations it had learned added to the fame of the system as it showed how the system performed better with each training iteration. NETtalk uses a feed-forward network consisting of an input layer, a hidden layer and an output layer. The input considers a character at a time, and uses a context window of three characters before and three after. This results in a total of 7 groups of units in the input layer of the neural network. Each group has a series of nodes which can be used in a binary coding fashion to represent each character (For example, if we have 5 nodes, then a would be 00001, b would be 00001, c would be 00001, and z would be 11010. A single group of units is used in the output and a single phoneme is therefore generated. A number of further studies have also used neural networks for G2P [191],[240] but this technique’s importance is mostly due to it being an powerful demonstration in the early days of machine learning. 8.4.4 Pronunciation by analogy Pronunciation by analogy is a popular data driven technique first introduced by Dedina and Nusbaum [127] and popularised by Damper and colleagues [422],[122],[123], [300], [27]. Its motivation comes from studies of human reading and draws from studies by Glushko [175] and Kay and Marcel [247]. The basic idea is that given an unknown word, a human considers known words which are orthographically similar, and adapts a pronunciation of the nearest word to use for the unknown word. For example, if we are confronted with a new word zate, we consider that while we don’t know the pronunciation of this, we do know of many similar words date, late, mate, all of which have the same ending pronunciation. In addition, we also know that most words that start with z generate a /z/ phoneme, and so we apply both these analogies and decide that zate should be pronounced /z ey t/. Marchand and Damper [300] give a review of this technique and describe it as belonging to the family of machine learning techniques known as memory based learning [119]. The algorithm works by comparing the characters in the input word with each entry in the lexicon. For each matching substring, a lattice is built with the input characters and the out- put phonemes (which are found by the alignment process). The lattice comprises a path of the characters c1,...,cN and a set of arcs which span character sequences in the path. Each arc represents one match between a sequence of characters and a sequence of output phonemes. Hence in through we would have an arc from c0 (a dummy start character) to c2, showing the match between the letters t h and the phoneme /th/. In addition a frequency count is kept of how many times this match is found in the lexicon. This lattice then represents a summary of all the matches in the lexicon for this word. Finally, a decision function is used to find the single best sequence of phonemes. The decision function can take a number of forms, and finding better methods of designing this function is the main goal of research in pronunciation by analogy. Marchand and Damper [300] give a review of some techniques. The original approach is to find the shortest paths through
- 246. 222 Chapter 8. Pronunciation the lattice, meaning choosing the matches which have the longest individual spans. If more than one shortest path is found, then the one with the highest frequency count is used. Other approaches include calculating the product of arc frequencies, the standard deviation of the path structure, the frequency of the same pronunciation, the number of different pronunciations generated, and the minimum of arc frequencies. 8.4.5 Other data driven techniques A number of other data driven techniques have also been used to G2P conversion. Pagel et al [343] introduced the use of decision trees for this purpose, a choice which has been adopted as a standard technique by many researchers [249], [198]. The decision tree works in the obvious way, by considering each character in turn, asking questions based on the context of the character and then outputing a single phoneme. This can be seen as a way of automatically training the context-sensitive rules described above. Other data driven techniques include support vector machines [114], [111], [18], transfor- mation based learning [67], [519] and latent semantic analysis [38]. Boula de Mareüil et al [62] describe a formal evaluation of grapheme-to-phoneme algorithms and explains the complexities involved in ensuring that accurate and meaningful comparisons can be performed. 8.4.6 Statistical Techniques The statistical approach has been somewhat of a latecomer to grapheme-to-phoneme conversion, perhaps because of the success of other data driven techniques such as pronunciation by analogy or the impression that context-sensitive rewrite rules are adequate so long as they can be automat- ically trained, e.g. by a decision tree. In recent years however a number of approaches have been developed which give a properly statistical approach. As mentioned previously, Jelinek [239] proposed a system to align graphemes and phonemes using the Baum-Welch algorithm original developed for speech recognition. His intention was to expand the lexicon to include unknown words that would then be used in a speech recogniser. His use of the statistical approach was limited only to the alignment step however, the actual grapheme-to-phoneme conversion was performed by other techniques. Taylor [433] took this work to its logical conclusion by using HMMs to perform the whole task. In this approach, states repre- sent the phonemes, each of which has a discrete output probability distribution over the characters which form the observations. The system therefore mimics a “writing” process where the under- lying form (phonemes) generates the surface representation (characters). G2P is then the opposite process of recovering the phonemes from the characters. If performed with a standard HMM, the results are quite poor but with various refinements the technique can be made to perform as well as any other. These include using context in the models, using an n-gram over the phoneme sequence and by performing pre-processing steps to make the alignment task easier. The main advantage of this technique is that no explicit alignment is required; the system implicitly generates a “soft”
- 247. Section 8.5. Further Issues 223 alignment during training. The use of the n-gram phoneme constraint also means that even if the answer is wrong, the generated phoneme sequence should be a likely one, and not simply a random sequence that is un pronounceable. A more common approach is to use technique proposed by Galescu and Allen where a search is performed using a joint n-gram model over both the grapheme and phoneme sequences [44], [172], [223], [86]. Typically, the alignment is performed in the normal way. This training data then gives a collection of short sequences of phonemes and graphemes which are aligned. For example, the word PHONEME gives { ph, f }, { o, ow }, { n, n }, { e, iy }, { m e, m }. Once the entire training lexicon has been converted to this format an n-gram is calculated over all sequences. At run time, a simple search gives the most likely sequence given the characters. The technique is completely general and allows for a variety of refinements in terms of n-gram smoothing and estimation techniques. It avoids the pre-processing steps of the HMM technique described above but does does however require a separate alignment step. 8.5 FURTHER ISSUES 8.5.1 Morphology It is common in linguistics to say that words are not atomic entities with regard to syntax or semantics but rather are composed of morphemes. If we consider the words BIG, BIGGER and BIGGEST we see there is a clear connection between the three words. The issue that concerns us is whether we should consider all of these as separate words, and store them in the lexicon, or just store the root form BIG and then use rules to derive the other forms. There is no right or wrong answer to this; the fundamental issue of one of size vs speed. If we have plenty of storage space and require quick processing, it is better to store all words. If storage space is limited it may be better to store only the root and generate other examples no the fly. In modern TTS systems, the first approach is generally taken as memory is cheap, and even for large lexicons, the size of these is swamped by the size of the unit selection database. It is traditional to defined two types of morphology; inflectional, where (in English) we typically add well defined suffixes to words to make other words in the same class (as above). This is in contrast to derivational where we make new words that often have different parts of speech or different meanings, such that we form CREATION from CREATE. While in general it is possible to store all common derived and inflected forms of words, the set of possible words is very large indeed; for example we can attach the suffix -ESQUE to virtually and noun or proper noun (e.g. TAYLORESQUE). We can deal with this in a number of ways, but the most common approach is to store all roots and common inflections and derivations in the lexicon. When other inflections or derivations are encountered, these are treated as unknown words and passed to the grapheme-to-phoneme convertor. This then can attempt a basic morphological decomposition which attempts to find
- 248. 224 Chapter 8. Pronunciation known morphemes within the word. Either this succeeds for all parts of the word, in which case the pronunciation can be formulated by using these parts, or only some of the morphemes are identified, in which case a mixture of a lookup table for the known morphemes and a grapheme- to-phoneme conversion for the unknown morphemes can be used. The difference may be slight for English, but this approach is certainly more attractive for languages with rich morphology such as German [370], [460], [483]. 8.5.2 Language origin and names A further clue to how to pronounce a word is given by the origin of the word. While we know that late is pronounced /l ey t/, and that the final e is not pronounced; we also know that latte is pronounced /l ah t ey/ with a pronounced final e. The reason is of course that LATTE is an Italian word and so receives an “Italian” pronunciation. We should note however that the word is not pronounced the way an Italian person would say it, but rather an English pronunciation using an approximation of the Italian phoneme sequence. Quite how “natively” words such as this should be spoken is by and large a cultural issue, many native English speakers pronounce the final /t/ in RESTAURANT many do not. Given, there have been approaches which have used analysis of character sequences to find the language of origin, typically using n-grams trained on separate lexicons from each language [94], [289]. A final point concerns the types of words that are most likely to be used by grapheme- to-phoneme algorithms. The approach taken by most machine learning research is to arbitrarily divide the lexicon into training and test data, for example by assigning every tenth word to the test set. When using G2P algorithms for lexicon compression purposes this is a reasonable thing to do, but when using the G2P algorithm for rare words that are unknown to the system, this becomes less valid. In a system with a large, well developed lexicon, the words most likely to be processed by a G2P convertor are vary rare, new formations (e.g. email) loan words or proper names. The chances are that G2P algorithms will fare worse on these than normal words, so care should be taken in assuming results taken from a normal word test set apply to the words we are likely to find in real usage. Taking this into account, a better approach would be to design or train G2P systems on just those words which we are most likely to see. For instance Spiegel [409] describes a system which uses an extensive set of rules to model proper names. The system includes all the types of names found in the United States and is designed primarily for directory enquiry and other telephony applications. 8.5.3 Post-lexical processing Some systems use the output from the lexical lookup process and/or grapheme-to-phoneme con- version directly as the input to the synthesizer. In some systems however, an additional step often known as post-lexical processing is performed which modifies the output in some way. The most significant factor in determining how much post lexical processing is the specific form of
- 249. Section 8.6. Summary 225 the input to the synthesizer. This was discussed in Section 8.1.3, where we saw that we could choose high level phonemic representation or a low level detailed phonetic representation. A low level representation will require significant processing, for example to turn an /n/ into an /m/ in a labial context, but as many modern systems use high level representations often no post-lexical processing of this type is required. Some processes operate at a purely phonological level, and so regardless of the type of input to the synthesizer, some processing may still be required. English doesn’t really have any clear demonstrations of this, but when we consider French we see that the process of liaison must be modelled. In general, the word LES is pronounced /l ey/, but when followed by a word starting in a vowel, it is pronounced /l ey z/ as in LES AMIS → /l ey z a m iy z/. 8.6 SUMMARY Basis for pronunciation • Using a phonemic or phonetic representation helps reduce the size of the specification for the synthesizer. • A choice has to be made between a high level abstract representation or low level detailed representation • In general, modern systems use high level representations. Pronunciation systems • Phoneme inventories can be designed using the principles of phonetic similarity and minimal pairs. • There is no single correct phoneme system for a language but if care is taken an adequate system is normally not too hard to define. • In many cases, suitable phoneme sets already exist, especially for well studied accents and languages Lexicons and rules • Pronunciations can be generated by lexicon look up or by algorithm. These two techniques should be viewed as points on a scale. • lexical lookup is fast by expensive in memory • using rules is expensive in processor time but cheap in memory Grapheme-to-phoneme conversion • Grapheme-to-phoneme conversion is the process of guessing a word’s pronunciation from its spelling.
- 250. 226 Chapter 8. Pronunciation • Traditional systems used context-sensitive rules. These are generally inaccurate for lan- guages like English, but can still produce good results for languages with more regular spelling. • Many data driven techniques exist, including neural networks, pronunciation by analogy and decision trees. • Some statistical techniques have recently been proposed.
- 251. 9 SYNTHESIS OF PROSODY This chapter is concerned with the issue of synthesising acoustic representations of prosody. The input to the algorithms described here varies but in general takes the form of the phrasing, stress, prominence and discourse patterns which we introduced in Chapter 6. Hence the complete process of synthesis of prosody can be seen as one where we first extract a prosodic form representation from the text, described in Chapter 6, and then synthesize an acoustic representation of this form, described here. The majority of this chapter focuses on the synthesis of intonation. The main acoustic representation of intonation is fundamental frequency (F0), such that intonation is often defined as the manipulation of F0 for communicative or linguistic purposes. As we shall see, techniques for synthesizing F0 contours are inherently linked to the model of intonation used, and so the whole topic of intonation, including theories, models and F0 synthesis is dealt with here. In addition, we cover the topic of predicting intonation form from text, which was deferred from Chapter 6 as we first require an understanding of intonational phenomena theories and models before explaining this. Timing is considered the second important acoustic representation of prosody. Timing is used to indicate stress (phones are longer than normal), phrasing (phones get noticeably longer immediately prior to a phrase break) and rhythm. 9.1 INTONATION OVERVIEW As a working definition, we will take intonation synthesis to be the generation of an F0 con- tour from higher level linguistic information. Intonation is probably the most important aspect of prosody (the others being phrasing and prominence) in that it has a richer degree of expression. As we explained in Chapter 6, the primary semantic content of prosody is in affective representations such as emotion and attitude, and intonation is the primary way in which this is communicated. A sentence can be made to sound believing or incredulous simply by changing the intonation. Into- nation is also used augmentatively in that many prominences are indicated with changes in local F0. Phrasing is also often indicated by a lowering of F0 at phrase ends and a subsequent raising 227
- 252. 228 Chapter 9. Synthesis of Prosody of F0 at phrase beginnings. Finally even if intonation is not used for any affective or augmentative purpose, all utterances have characteristic intonation patterns which must be modelled correctly if the speech is to sound natural. 9.1.1 F0 and pitch In this book we have used the terms F0 (fundamental frequency) and pitch interchangeably, as there are few cases where the difference matters. Strictly speaking, pitch is what is perceived, such that some errors or non-linearities of perception may lead this to be slightly different to fundamental frequency. Fundamental frequency is a little harder to define; in a true periodic signal this is simply defined as the reciprocal of the period, but as speech is never purely periodic this definition itself does not suffice. An alternative definition is that it is the input or driving, frequency of the vocal folds. In prosody, F0 is seen as the direct expression of intonation; and often intonation is defined as the linguistic use of F0. The relationship between the two is a little more subtle than this though as it is clear that listeners do not perceive the F0 contour directly, but rather a processed version of this. The exact mechanism is not known, but it is as if the listener interpolates the contour through the unvoiced regions so as to produce a continuous, unbroken contour. 9.1.2 Intonational form There is probably no area of linguistics were there is less agreement than when it comes to the nature of intonational form. In the following sections we will cover many of the well known models and theories, and it will be clear that there is enormous variation in how researchers have chosen to represent intonation. Disagreements occur in all other areas of linguistics; it is just the degree to which the theories differ in intonation that is so noticeable. For instance, when we consider the theories of verbal phonology described in Section 7.4, we do in fact find that there are quite a number of approaches to the problem, but all researchers at least agree that there is such a thing as phonology. We can give an overview of the situation as follows. The function of intonation is to express emotion and attitude, to augment the verbal component by firstly emphasising or marking words, and secondly by giving phrases and utterances characteristic patterns which show their structure and type. Although rarely expressed as such, this much is uncontentious1 At the other end of the scale, it is widely accepted that F0 is the main acoustic manifestation of intonation. The disagreement stems around how any intermediate intonational form should be represented. In some early studies, researchers did actually try to map directly from intonational semantics to acoustics [107], [287], but these studies were criticised by many, summarised clearly by Ladd [269]. Basically, in measuring say the F0 of anger, we must of course do this with reference 1 Any disagreement would be based around the notion of whether there was a pure linguistic intonational core, and a separate para-linguistic or extra-linguistic component.
- 253. Section 9.1. Intonation Overview 229 to particular speakers and if we did this, the difference in basic physiological behaviour between speakers might well swamp any subtleties in the effects of anger itself. Put more plainly, regardless of how high or low a speakers voice, a woman, child or man plainly exhibits clear regularities in how these effects are expressed, and we are missing something if we state that the mean value of anger is say 150Hz. This leads us to the conclusion that there must be some sort of intonational form that speakers of a language/accent share, and which is readily identifiable to those speakers. This is obviously somewhat similar to the idea of the form of words and phonemes in verbal language, and often just this parallel is drawn such that we are trying to find the intonational equivalents of notions such as word, phoneme, phone and allophone. The basic problem is that while nearly everyone agrees that some representation of prosodic form must exist, we have no equivalent test to the minimal pair test for phonemes. Recall from Section 7.3.2, that this said that if two sounds could be found that distinguished two words, then we could call these phonemes. If however we had two sounds which clearly sounded different (say dark and light [l] in English), but which did not distinguish words, then these were not phonemes but allophones. Crucially, this test lies on the extremely convenient ability of listeners to tell whether two words are the same or not - this can be done without even knowing what the words mean. In intonation, and other aspects of prosody, we simply do not have such a test, and the “semantic space” while perhaps simpler overall, doesn’t arrange itself into nice discrete categories the way the words in a lexicon do. We should realise from the start however that not every model was designed with the same goals in mind; for instance, the British school (explained in Section 9.3.1) evolved from old style prescriptive linguistics, and even quite recent publications were concerned with the teaching of “correct” intonation patterns to non-native learners of English. The aims of autosegmental- metrical (AM) (Section 9.3.3) school were to provide a theory of intonation that worked cross linguistically such that phenomena from African and Asian tone languages, pitch accent lan- guages such as Swedish, and intonation languages such as English could all be described with a unified model. An aim of the Fujisaki model (Section 9.3.5) was to create a model that followed known biological production mechanisms, whereas the aim of the Tilt model (Section 9.3.6) was to create a model solely for engineering purposes. The models differ significantly in what they take as the primary form of intonation. In the AM model this is quite abstract while in the Tilt model this is quite literal or “acoustic”. These differences in primary form should not be taken to mean that the proponents of these models do not believe that there should be more abstract or more concrete representations, just that the “best” representation happens to lie where they describe it. In the many synthesis schemes based on the AM model there are other, more phonetic or acoustic levels, and in the Tilt model there is always the intention that it should serve as the phonetic description of some more abstract higher level representation.
- 254. 230 Chapter 9. Synthesis of Prosody 9.1.3 Models of F0 contours Imagine that we have an adequate representation of intonational form and that we have extracted such a representation from the input text. How then should we go about generating an F0 contour? We have seen a consistent pattern throughout the book where mappings between one domain and another are performed by means of either explicit rule, or, in more recent systems, machine learning algorithms such as decision trees or neural networks. The general formulation of this is as a mapping between a set of features and a label: X(f1, f2,...fN) → Li This approach worked for semiotic classification homograph disambiguation, POS tagging phrase break prediction, prominence prediction and many other problems described in Chapters 4 and 5. It would therefore seem reasonable to try such an approach for intonation synthesis. Strictly speaking, the algorithms used in text analysis are called classifiers; as they attempt to classify an entity using features which descrive it. This is a labelling process of assigning one label from a pre- defined set to the entity. In F0 and timing (discussed below) the problem is similar but now we have to assign a continuous value to our entity, and techniques which perform this are called regression algorithms. In many cases the same algorithms can be use for both; for example decision trees are really classifiers but can be adapted to become regression algorithms by replacing the labels at the nodes with distributions. Neural networks are really regression algorithms but can be converted to classifiers by interpreting each output node as representing a binary part of a label. In practice standard regression algorithms are difficult to use directly for F0 prediction, be- cause of a serious mismatch in the nature of the input and output representations. In general the input representation is a representation of intonational form as just discusses (for example, pitch accent types and positions) whereas the output is a continuous list of real valued numbers. In particular, the feature combination needs to generate not one, but a sequence of F0 values, which is further complicated because the number of values in this sequence can vary from case to case: X(f1, f2,...fN) → F01,F02,...F0M, Furthermore, when we come to performing a bottom up analysis, where we wish to describe the continuous contour with a discrete set of linguistic units, we find the mapping even more complicated. Because of these problems, it is nearly universal to use some sort of acoustic model which acts as an interim representation between the abstract intonational form and the F0 contour. This model has a fixed number of parameters per unit, and so we can use one of the standard mapping algorithms (such as CART) to generate these parameters from the intonational form: X(f1, f2,...fN) → P1,P2,...PL, As a second step we can use the specialised model to generate the F0 values themselves:
- 255. Section 9.2. Intonational Behaviour 231 Y(P1,P2,...PN) → F01,F02,...F0M, These models usually work on a principle whereby they encode or stylize the real F0 con- tour. In other words, they give a tractable simplification of the F0 contour in which the salient linguistically important aspects of the contour are preserved, while less necessary parts are dis- carded. At its simplest, the parameters would be a discrete list of target points in the contour. From these, a continuous F0 contour can be synthesized. If we keep the number of points constant for each linguistic unit, then we have a situation where the parameters of the model align nicely with the linguistic description, while hopefully still allowing accurate synthesis of the contour. The values of these parameters can then be learned by a machine learning algorithm or written by hand, as we now have a much simpler situation where we can align input and output pairs. The models described below generally take a more sophisticated approach, but in essence the principle is the same: to make the mapping between the discrete prosodic form units and the F0 contour more tractable. 9.1.4 Micro-prosody Micro-prosody is a term used to denote some of the lowest level, more detailed aspects of prosody. Often this includes the interaction between verbal effects and prosodic effects on the main acoustic components of prosody such as F0. It has been widely observed [405], [270] that the phones of an utterance have an affect on the F0. While these effects are less noticeable than the effect of prosody proper, they need to be handled if a complete model of F0 synthesis or analysis is to be performed. Typical effects include a sudden raising or lowering of F0 at a boundary between a voiced and unvoiced section of speech. The degree to which micro-prosody needs to be synthesized is debat- able. Most intonation algorithms synthesise smooth contours and simply ignore micro-prosody altogether and unit selection systems have the fine micro-prosodic detail already included in the units they use. If however any automatic analysis of F0 is to be performed, the micro-prosody must be handled. One common approach is simply to use an algorithm to smooth the F0 contour, in an attempt to make it appear more like the stylised intonation contour that we believe represents the underlying intonation pattern. Modern statistical techniques can handle miro-prosody more directly, by considering it as noise in the data. 9.2 INTONATIONAL BEHAVIOUR In this section we describe some basic intonational phenomena and attempt to do so in a theory neutral way.
- 256. 232 Chapter 9. Synthesis of Prosody 9.2.1 Intonational tune The intonation tune can be broadly described as the core pitch pattern of an utterance. Tunes differ from one another in type and in association. By using different types of tunes, the speaker can express for example an emotion such as surprise, disbelief or excitement. The association of the tune connects the prosodic part of intonation (the tune itself) to the verbal part, by associating parts of the tune with particular words. By shifting association from “john” to “match” in examples 118 and 119, one can convey different effects. By varying the type of tune, one can also express different effects as in example 120. (118) John went to the match (as opposed to Harry) (119) John went to the match not the theatre (120) John went to the match (disbelief: but he hates football ! Describing tune type is perhaps the most difficult issue in prosody. Tune schemes can be broadly divided into those which classify tunes using dynamic features (rises and falls) and those which use static features (tones). Theories also vary in the size of the units they use. Global descriptions make use of a few basic patterns that cover the entire phrase, atomistic theories make use of smaller units that combine together to form larger patterns. Jones [242] is at the global end of the scale, the British school [333], [193] uses sub-phrase units (see Section 9.3.1 , while the AM school [352], [351] and the Dutch school [446] use units which are smaller still (Sections 9.3.2, 9.3.3. Much of the discussion on the subject of tune centres around how to describe pitch accents. A pitch accent is commonly manifested in the F0 contour as a (relatively) sudden excursion from the previous contour values. This is where association comes in, as pitch accents only occur in conjunction with prominent syllables, and in doing so attract attention to that syllable. Pitch accents can only occur in association with prominent syllables (see Section 6.3 on prominence), but need not occur on all prominent syllables. Most work agrees that the prosodic phrase as described in Section 6.2 is the basic domain of intonational tune patterns. Hence the phrase not only groups words together, it also serves as the start and end of each section of tune. In many models, we have a nuclear accent, which occurs once per phrase. The other main area of interest in tune description concerns what happens at the ends of intonation phrases. Often F0 is low at a phrase boundary, but in many circumstances F0 is high. For instance, if another phrase directly follows the current one, a continuation rise may be present. If the tune is that of a yes/no question, the final pitch may also be high. The British school deals with these effects by using different nuclear accent and tail configurations. The AM model makes use of high and low boundary tones which distinguish the different types of contour. It would be impossible to show all the possible types of intonational tune for English, but six common tunes which vary because of their nuclear accent types are shown in Figures 9.1 to 9.6. These examples are not comprehensive and other theories may classify these contours differently.
- 257. Section 9.2. Intonational Behaviour 233 Fo time anna came with manny * Figure 9.1 Typical Fall Accent. “Anna came with Manny.” The nuclear fall is on the stressed syllable of “manny” denoted by an *. The fall is a commonly found intonation accent and is often used in neutral declarative situations. Fo time anna came with manny * Figure 9.2 High Fall, “Anna came with Manny !”. This shape corresponds to a British “high fall”, +raised or pitch level 4. In this particular utterance there is still a single intonation phrase, and the word “anna” also has an accent, but this accent is pre-nuclear. Some may argue that there is no phonological distinction between fall and high fall, and that the high fall is really just an extra prominent fall. The examples merely demonstrate some of the intonational effects that can be produced. 9.2.2 Downdrift It has been observed by many researchers that there is often a gradual downdrift in the value of F0 across a phrase [446], [447], [351], [107], [285], [167]. How downdrift (often referred to as declination) is dealt with by different theories varies widely. Ladd [267] gives a review of some of the different theories. Many treat downdrift as an automatic physiological effect arising from changes in sub-glottal pressure during the course of an utterance [287], [107]. This account gives the speaker little conscious control over declination. The approach of the Dutch School [446], has been to use three
- 258. 234 Chapter 9. Synthesis of Prosody Fo time it was amazing * Figure 9.3 Rise fall accent, “It was Amazing !”. Spoken with a sense of wonderment, this accent is similar to a fall, but with a much larger preceding rise. The peak value of the F0 contour is also later than with a simple fall accent. Fo time * with anna manny come did Figure 9.4 Fall rise accent, “Did Manny come with Anna ?” A peak in the F0 contour occurs in the stressed syllable of “manny” (*). After coming down from the peak, the contour rises slowly and finishes with a sharp rise at the end of the phrase. This type of accent is often used for simple yes/no questions. parallel declination lines, which refer to a baseline, a mid-line and a line corresponding to the top of the speaker’s normal range. The contour must follow one of these lines or be rising or falling between them. Fujisaki’s model is more flexible in that the rate of declination and initial starting value can be varied, but the overall effect is still automatic [167]. Liberman and Pierrehumbert [285] show that the final F0 value for utterances is invariant under a wide range of utterance lengths and pitch ranges which is inconsistent with the view that declination slope is constant. They propose an exponential decay downdrift effect, with the additional feature of final lowering at the end of the phrase. Figure 9.7 shows three different views of declination. A major claim of Pierrehumbert’s thesis [351] was that more than one factor was responsible for the downdrift of F0 contours. As with many other theories, she proposed that the phonetic dec- lination effect exists, but also argued that the major contribution to the downdrift of utterances was downstep which was a phonological effect and therefore controllable by the speaker. Figure 9.8
- 259. Section 9.2. Intonational Behaviour 235 Fo time did come with anna * manny Figure 9.5 Low Rise, “Did Manny come with Anna ?!”. This accent shape may at first glance look similar to the fall-rise, but differs in that the stressed syllable (*) of the word which carries the nuclear accent is not a peak but a valley. Thus the F0 contour rises from the nuclear accent. Quite often this accent is preceded by a falling F0. This accent can be used to convey incredulity or disbelief. Fo time did come with anna manny * Figure 9.6 High Rise, “Did Manny come with Anna ?”. Here the accent falls on the first syllable of Anna. There is no valley as with the low rise, and the F0 on the nuclear syllable is much higher. High rise accents are often used for yes/no questions where the speaker is looking for confirmation in a statement, as in “ok?” or “ right?”. It is similar in many ways to the low rise, with the F0 contour rising from the nuclear accent, the main difference being that the nuclear accent occurs considerably higher in the speaker’s pitch range, and is often not preceded by a falling section of contour. shows a downstepping and non-downstepping version of the same sentence. These two sentences are not only different in F0 shape, they also have subtly different meanings. (The first sounds more excited, the second sounds more relaxed and confident.) 9.2.3 Pitch Range In music, if a sequence of notes is repeated an octave higher than the original, the tune remains the same, even though the frequency values of the notes are different with respect to the original. Rather it is the constancy of the pattern that gives the perceptions of the tunes being the same.
- 260. 236 Chapter 9. Synthesis of Prosody 0 0.5 1.0 1.5 (a) (b) (c) Fo Fo Fo Figure 9.7 The Dutch model, figure(a), has three declination lines, which refer to a baseline, a mid-line and a line corresponding to the top of the speaker’s normal range. The contour must follow one of these lines or be rising or falling between them. Pierrehumbert’s system (figure (b)) scales the pitch range 0.0 to 1.0 for normal speech but allows higher levels. The contour is not required to follow any of these declination lines - they are merely the “graph paper” on which the F0 contour is produced. Note how the lines converge with respect to time. The Fujisaki model (figure (c)) only specifies a baseline, which decays exponentially. The same effect is observable in intonation: by increasing the overall pitch of an utterance while keeping the basic tune pattern constant, the perceived tune remains the same. The relation- ship between intonational tunes in different pitch ranges is not as simple as the musical equivalent. It has been shown that the increase or decrease in pitch range need not be constant throughout the phrase, as utterances always tend towards a fairly constant final F0 value. Pitch range varies for a number of reasons. In single isolated utterances it can be used for increasing the overall emphasis of the utterance. When one “raises one’s voice” in anger one is using an increased pitch range. Pitch range factors also have a role to play in longer utterances. If
- 261. Section 9.2. Intonational Behaviour 237 really really good Fo Fo (a) (b) really really good time time Figure 9.8 Two utterances of the phrase “really, really good”. Figure (a) has the two “really”s at the same pitch level, with a fall on “good”. In Figure (b) each word is downstepped relative to the previous word. a speaker started at the same F0 level with every intonation phrase, the speech would sound very bland. Speakers use a variety of pitch ranges to be more expressive. The boundaries between pitch range and prominence effects are not clearly defined. For example, in many systems it is difficult to say if an unusually high accent is due to extra (local) prominence or a temporary shift in pitch range. 9.2.4 Pitch Accents and Boundary tones Many theories of intonation define pitch accents as the fundamental unit of intonation. While we defer specific systems of pitch accent descriptions for the section below on intonational models, here we attempt to describe some pitch accent phenomena in theory neutral terms. In most models that use pitch accents, we see a basic distinction between the pitch accent’s inherent properties and its setting. In general, by inherent properties, we mean the shape of the F0 contour of the pitch accent. To a degree, we can separate the F0 pattern of the pitch accent from the surrounding contour, but
- 262. 238 Chapter 9. Synthesis of Prosody this is only for convenience, in reality the pitch accents rarely have distinct beginnings and ends. The most common type of pitch accent takes the form of a peak in the F0 contour. The actual shape of this map be a rise followed by a fall, or a rise followed by a relatively level section and then a fall. The degree of the excursion is important, and generally speaking the higher the excursion the more prominent the pitch accent will sound. There are many other types of pitch accent also. One common type is the simple fall, where the contour is level, falls over a short region and then continues at a reasonably level rate. Another important type is the low pitch accent, or rising pitch accent. The local shape of this is a little harder to define, but basically this type of accent is marked by a negative excursion in the F0 contour. This can take the shape of a fall followed by a rise, or just a specific low point in the contour. This is often associated with question intonation, or cases where the speaker is doubting the listener and so on. Beyond these, theories differ as the range of additional types. The important point about pitch accent shape is that too often these are thought of as simple peaks in the contour; falls and low pitch accents are just as important. The pitch accent’s setting in the overall contour is also vital. This is governed by two factors; its height, meaning how high or low it occurs in the speaker’s range and its alignment with the phonetic content. In terms of height, we find a few characteristic patterns. In general, pitch accents height decreases as a function of time throughout the utterance, as we just mentioned on the section about downdrift. We also find that as with the size of the excursion, the accent height is correlated with the degree of prominence (in fact in many models, accent height and excursion size are governed by the same parameters). The issue of alignment is more subtle. While intonation does has a form of its own, its interaction with the verbal phonetic content of the utterance is important to how it is produced and perceived. We have already seen that association is a key link between the prosodic and verbal component as it determines which syllable or which word is accented. In addition to this, the precise position of the accent in time with respect to the accented syllable is very important. Some models [260], [259], [258], talk of early accents and late accents, and some argue that exactly the same intrinsic pitch accent shape can give rise to quite different overall perceptions if only the alignment differs. Quite detailed studies of alignment have been conducted [465], [266] and [132], and found that this is a complex issue, and it is not even clear just what should be aligned with what. Some suggest that the start of the vowel in the accented syllable should act as one anchor point, with the peak or mid point of the pitch accent acting as the other. Many other alignment models are however possible and in general the interactions between the intonation and phonetics are quite complicated. The other major intonational effect is the behaviour at phrase boundaries. The F0 at these points can be systematically raised or lowered for specific effect. The term boundary tone is often used as the equivalent intonation unit to pitch accent. Strictly speaking, this term is only applicable to AM models, but it has gained wide enough usage that we use it here to describe general behaviour at boundaries regardless of whether the particular model uses tones or not.
- 263. Section 9.3. Intonation Theories and Models 239 9.3 INTONATION THEORIES AND MODELS 9.3.1 Traditional models and the British school The British School of intonation includes contributions made as far back as Palmer [344]. Other major contributions in this school have come from O’Connor and Arnold [333], Crystal [117], and Halliday [193]. All these variants on Palmer’s original theme use dynamic features such as rise and fall to describe intonation. In the account given by Crystal, the most important part of the contour is the nucleus which is the only mandatory part of an intonation phrase. The nuclear accent can take one of several configurations, e.g. fall, fall-rise, low rise. Other parts of the contour are termed the tail (T) which follows the nucleus, the head (H) which starts at the first accented syllable of the intonation phrase and continues to the nucleus, and the pre-head (P) which precedes the head. The intonation phrase has a grammar of (P) (H) N (T), where the brackets denote optional elements. The relationship between the form and acoustics of this school is the most loosely defined of all the models described here; this is hardly surprising however as none of the originators of this system had the technology to analyse F0 contours in detail. The description of form is related to actual contour shapes that are found, but the descriptions should not be interpreted too literally. Both “fall” and “rise-fall” accents have rises followed by falls, the difference being that the rise in the fall accent is much smaller and earlier in the syllable than the rise in the rise-fall accent. Halliday describes his tones using rise and fall terminology, but does away with the standard naming system preferring simply to name his tones 1, 2, 3, 4 5. Some more formal descriptions have been proposed for use with the British School phonol- ogy. In particular, two models which have been designed for speech synthesis purposes are those of Isard and Pearson [231] who use Crystal’s phonology and Vonwiller et al [479], who use Hall- iday’s. Both these synthesis models use the full range of the British school tune descriptions and Isard and Pearsons’s scheme is capable of variations in prominence and pitch range. 9.3.2 The Dutch school The Dutch School [446], [447] is based on the principle of stylization of F0 contours. Stylization in the Dutch system involves taking a F0 contour and attempting to fit a series of straight lines as closely as possible to the original contour. This stage is useful as it reduces the amount of data needed for further analysis: a small number of straight lines are easier to deal with than a continually varying F0 contour. From these stylizations, a series of basic patterns can be found - this process is called standardization. The version of the theory presented in t’Hart and Cohen [446] describes contours in terms of three declination lines - high, middle and low. Pitch accents are realised by rising and falling between these declination lines. An example of a stylized and standardized contour is shown in Figure 9.9 (from Willems [494]).
- 264. 240 Chapter 9. Synthesis of Prosody 0.0 0.3 0.6 0.9 1.2 1.5 100 200 300 500 50 and the weather turned out fine Fo (Hz) time (seconds) Figure 9.9 Example of a standardized contour in the Dutch system. The dotted lines denote the three declination lines and the thicker solid line shows the path of the F0 contour. The first excursion to the top declination line is a head accent (British school). The second accent which rises to the top line and then falls to the baseline is a fall accent. The rise at the end is a continuation rise. Because of the stylization process, the continuously varying nature of the F0 contour is eliminated, and because of the standardization process, the contour description is further reduced into a small number of units (rises, falls etc). The idea of F0 stylisation has proved popular with many models and techniques outside the Dutch school [481], [121]. 9.3.3 Autosegmental-Metrical and ToBI models Ladd terms this group the autosegmental-metrical (AM) model of intonation. Its origins are in Liberman’s [284] and Bruce’s [286] work, but its first full exposition was in Pierrehumbert’s seminal thesis [351]. From that basic model, a number of alternatives and amendments have been proposed, but the basics remained unchanged. Most significantly, the intonation part of the ToBI scheme comes directly from Pierrehumbert’s work in this school. Before these early publications, it was rare to associate the word “phonology” with prosody, but these early works promoted the idea that intonation and prosody could be described and modelled with much the same technical apparatus as normal phonology. Hence this school is often referred to as intonational phonology [266], [269]. The AM model describes intonation as a series of high and low tones. By using a system of diacritics which distinguish tones located on accented syllables from those occurring at boundaries and between accents, Pierrehumbert argued that intonation could be described by patterns of two basic tones, which she called H (high) and L (low). Pitch accents can be represented as either a single or double tone. Every pitch accent has a starred tone (*) which signals that it is that tone which is directly associated with the accented syllable. The possible pitch accents are H*, L*, H* + L, H + L*, L + H* and L* + H. At phrase boundaries, boundary tones can be found, which
- 265. Section 9.3. Intonation Theories and Models 241 are marked with a (%). Phrase tones are used to show path of the contour from the last (nuclear) accent to the phrase boundary, marked with a (-). Unlike the British school analysis, there is no strict division of the contour into regions such as head and nucleus. Both nuclear and pre-nuclear accents can be any one of the six types described above. The nucleus accent is distinguished because the phrase and boundary tones that follow it allow a much larger inventory of intonational effects. Each tone forms a target from which F0 contours can be realised by using interpolation rules. As with many other theories, the AM model retains the idea of a declination baseline, but says that the downdrift commonly observed in F0 contours is mainly due to the phonological effect of downstep which again is controllable by the speaker. In her original work, Pierrehum- bert proposed that the downstep effect is triggered by the speaker’s use of a sequence of H L H tones, using evidence from African tone languages as justification (see Figure 9.8 for examples of downstepping and non-downstepping contours). The history of the AM model is particularly interesting in that it has a dual character of being intended as a “pure” linguistic theory in the MIT, Chomsky/Halle phonology tradition, and as a working model for the Bell Labs TTS system. It is important to realise that from a theoretical linguistic point of view, the model as just described is not intended to be a “complete” model of intonation as used in human communication. Rather, the AM school divides the general intonation system into a part which is properly linguistic, a part which is para-linguistic and a part that is extra-linguistic. The distinction is seen as important within this school as the properly linguistic part, i.e. that described above, is the main area of interest as it forms part of the general grammar of a speaker’s language faculty. For a complete description however, we require the para-linguistic effects too, which include for instance the degree of prominence of a pitch accent, and extra- linguistic effects such as global F0 settings. For our purposes, the distinctions between these subsystems are less important, but it is important to realise that the system of H* + L accents and so on is only a partial representation of an utterance’s intonation. 9.3.4 The INTSINT Model The INTSINT model of Hirst et al [212], [213], [211], was developed in an attempt to provide a comprehensive and multi-lingual transcription system for intonation. The model can be seen as “theory-neutral” in that it was designed to transcribe the intonation of utterances as a way of annotating databases and thereby providing the raw data upon which intonational theories could be developed. Hirst has described the development of INTSINT as an attempt to design an equivalent of IPA for intonation. As stated in Section 6.10, there is no such thing as completely theory neutral model as all models make some assumptions. Nevertheless, INTSINT certainly fulfills its main goals of allowing a phonetic transcription of an utterance to be made without necessarily deciding which theory or model of intonation will be subsequently used. INTSINT describes an utterance’s intonation by a sequence of labels each of which repre- sents a target point. These target points are defined either by reference to the speaker’s pitch range,
- 266. 242 Chapter 9. Synthesis of Prosody in which case they are marked Top (T), Mid (M) or Bottom (B), or by reference to the previous tar- get point, in which case they are marked Higher (H), Same (S) or Lower (L). In addition, accents can be marked as Upstepped (S) or Downstepped (D). Hirst [212] describes this system in detail and shows how it can be applied to all the major languages. Several algorithms have also been developed for extracting the labels automatically from the acoustics and for synthesizing F0 contours from the labels2. Applications of this model to synthesis include Veronis et al [475]. 9.3.5 The Fujisaki model and Superimpositional Models Fujisaki’s intonation model [166] takes a quite different approach to the models previously dis- cussed in that it aims for an accurate description of the F0 contour which simulates the human production mechanism. Fujisaki’s model was developed from the filter method first proposed by Öhman [335]. In the model, intonation is composed of two types of components, the phrase and the accent. The input to the model is in the form of impulses, used to produce phrase shapes, and step functions which produce accent shapes. This mechanism consists of two second order critically damped FIR filters (these are intro- duced fully in Section 10.4). One filter is used for the phrase component, the other for the accent component. The F0 contour can be represented by equations 9.1, 9.2 and 9.3. lnF0(t) = lnFmin + I ∑ i=1 Api Gpi (t −T0i)+ J ∑ j=1 Aaj (Gaj (t −T1j )−Gaj (t −T2j )) (9.1) where Gpi (t) = α2 i te−αit for t ≥ 0 0 for t 0 (9.2) and Gaj (t) = min[1−(1+βjt)e−βjt,θ] for t ≥ 0 0 for t 0 (9.3) 2 In addition to being an expert on multi-lingual intonation, Hirst is also know for being able to do a particularly fine rendition of “Don’t Cry For Me Argentina”.
- 267. Section 9.3. Intonation Theories and Models 243 time time Fo Amplitude (b) (a) Figure 9.10 Three phrase components of differing amplitude. Graph (a) shows the path of the F0 contour, while graph (b) shows the input impulses. Fmin baseline I number of phrase components J number of accent components Api magnitude of the ith phrase command Aaj magnitude of the jth accent command T0i timing of the ith phrase command T1j onset of the jth accent command T2j end of the jth accent command αi natural angular frequency of the phrase control mechanism of the ith phrase command βj natural angular frequency of the accent control mechanism of the jth accent command θ a parameter to indicate the ceiling level of the accent component. Although the mathematics may look a little complicated, the model is in fact very simple. Each phrase is initiated with an impulse, which when passed through the filter, makes the F0 contour rise to a local maximum value and then slowly decay. Successive phrases are added to the tails of the previous ones, thus creating the type of pattern seen in figure 9.10. The time constant, α, governs how quickly the phrase reaches its maximum value, and how quickly it falls off after this. Accents are initiated by using step functions. When these step functions are passed through the filter they produce the responses shown in figure 9.11. The accent time constant, β, is usually
- 268. 244 Chapter 9. Synthesis of Prosody time Fo time Amplitude (b) (a) Figure 9.11 Three accents of differing duration. Graph (a) shows the path of the F0 contour, graph (b) shows the input step function for the second accent. As the accent duration increases, the accent becomes less like a peak and develops a flat top. much higher than α, which gives the filter a quicker response time. This means the shape produced from the accent component reaches its maximum value and falls off much more quickly than the phrase component. Phrases are initiated by impulses, which have the effect of resetting the baseline when fed through the filter. Figure 9.10 shows the F0 contour produced by the phrase component. A number of amendments have been proposed to the Fujisaki model aimed at giving it more flexibility for modelling languages other than Japanese for which it was originally designed. For instance, Van Santen et al [465], [467], [411] have proposed a model which allows more than one type of phrase contour, and allows the modelling of micro-prosodic effects as well as detailed alignment of the contour to the phones. The principle of superimposition is attractive in itself as it seems intuitive to model different parts or functions of the F0 contour separately and then combine these to produce the final contour for the utterance. This has lead to other types of superimpositional models, for instance the superposition of functional contours model of Bailly and Holm [31], discussed in detail in Section 9.6.4. 9.3.6 The Tilt model The Tilt intonation model developed by Taylor [434] [439] [436] was developed with the explicit intent of creating a practical, engineering model of intonation. To this extent, issues of purely linguistic concern (such as phonological rules in the AM model) or biological plausibility, as in the
- 269. Section 9.3. Intonation Theories and Models 245 rise amplitude amplitude fall rise duration fall duration Figure 9.12 A pitch accent split into rise and fall components. +1 +0.5 0.0 −0.5 −1.0 Figure 9.13 Five pitch accents with differing values of tilt, from +1 (pure rise) to -1 (pure fall). Fujisaki model were ignored. While the Tilt model is no more plausible as a means of intonation production that say concatenative synthesis, it was designed so that its parameters would have a clear linguistic interpretation. The Tilt model describes intonation abstractly as a series of events, and in this sense it fol- lows the AM model. The main difference is that instead of having a fixed set of categories for pitch accents and boundary tones; it uses a set of continuous parameters instead. The reason for this is that Taylor thought the evidence for the particular categories defined in the AM model as being weak. The basic argument was that the AM model acted like a parallel to the verbal phonological model. With verbal language, phonetically we have a continuous space, either ar- ticulatory or acoustic, but cognitively, this is divided up into a discrete set of phonological units, ie the phonemes. The AM model follows the same policy with intonation, but Taylor’s concern was that there was no equivalent for the minimal pair test to decide how the phonetic space should be divided up. In fact, as far as he was concerned, no objective test could be developed to find the pitch accent categories of a particular language. The solution was to abandon the idea of a discrete number of categories and instead describe the intonational events with a small number of continuous parameters. While the possibility of some objective test being developed is still possible, being an engineering model purely, it made sense to err on the side of caution and use a model which at least described F0 contours accurately and effectively.
- 270. 246 Chapter 9. Synthesis of Prosody In the Tilt model, there are two types of event accent and boundary (these are effectively the same as the pitch accents and boundary tones in the AM model). Each event comprises two parts, a rise and a fall, as shown in Figure 9.12. Between events straight lines called connections are used. All variation in accent and boundary shape is governed by firstly changing the relative sizes of the rise and fall components, and secondly by how the event aligns with the verbal component (basically, whether it is higher or lower in the pitch range, or occurs earlier or later with respect to the accented syllable). This then gives a total of six parameters, 4 describing the shape of the event and 2 describing how it aligns. This basic model (called the rise/fall/connection model [439]) models contours accurately but is not particularly amenable to linguistic interpretation. The Tilt model itself is a transformation of the 4 parameters for each event (rise and fall amplitude, and rise and fall duration) into the 3 Tilt parameters. Amplitude and duration are given by Amplitude = A = |Arise|+|Afall| Duration = D = |Drise|+|Dfall| The Tilt parameter itself is used to define the general shape of the event, independent of its am- plitude, duration or alignment. To find the value of Tilt, we first calculate an interim value, Tilt amplitude, found by taking the ratio of the differences of the rise and fall amplitudes divide by the sum of the rise and fall amplitudes: tiltamp = |Arise|−|Afall| |Arise|+|Afall| and likewise for duration tiltdur = |Drise|−|Dfall| |Drise|+|Dfall| these two quantities are highly correlated, which allows us to combine them into the single final Tilt parameter: tilt = tiltamp +tiltdur 2 Figure 9.13 shows a number of event shapes for different values of the Tilt parameter. 9.3.7 Comparison Many other models have been proposed in addition to the above, but the models described are perhaps the most popular and at least cover the range of models available. An entire book could be devoted to discussing the philosophical, historical, theoretical and practical issues surrounding why one model has the features it does and why it may or may not be better than another model. For now, however we will simply describe some of the differences in the model to help better understand why these differences occur.
- 271. Section 9.3. Intonation Theories and Models 247 Purpose Not all models were designed for the same purposes. The AM models are usually not described as models at all, but as theories as to how intonation actually works in human communication. Issues such as cross-language studies and links with other components of human grammar and so on are particularly important in this group. The Fujisaki model is often said to be biologically justifiable, meaning that it mimics the actual articulation of the human production mechanisms. The Tilt model is only described in practical terms, as a tool to get the job done and has no claims to model any aspect (cognitive or biological) of human behaviour. These different approaches are of course quite common in all areas of language study; what is of note is that intonation seems to be one of the few areas where engineering and scientific approaches still have enough commonality to constitute a single field. Phonological vs Phonetic vs Acoustic The AM model is phonological, the INTSINT model phonetic and the Fujisaki and Tilt models acoustic. While these differences are obviously important, one should be wary as to placing too much significance on these as being statements about what the model developers think is the “real” nature of intonation. For example, Hirst proposed INTSINT as an intonational equivalent to IPA that would allow researchers to label what intonation information a contour contained independent of linguistic theory, in just the same way that a phonetician might make detailed transcriptions of verbal phenomena. Once done, a phonological theory can be built using these transcriptions as data, and so it is not correct to say that Hirst thinks intonation is “phonetic” in character. Likewise, in the AM model, there is the assumption that production includes phonetic and acoustic components, and in acoustic models the assumption again is that there are higher level more abstract representations. In fact, there is nothing to stop creating a single formulation where the AM model would serve as the phonological representation, the INTSINT model as a phonetic representation and the Tilt model as the acoustic representation. Tones vs Shapes There is a basic distinction in most models as to whether intonation is an inherently tone or pitch- level based system (as in AM models), or whether it is based on pitch shapes or pitch dynamics (as in the Dutch model, Fujisaki model or Tilt). In theoretical linguistics this debate is seen as particularly important as researchers are attempting to find the “true nature” of intonation, with “conclusive” evidence occasionally being produced for one view over the other. The issue is hard to resolve, and it may be case where intonation is actually a mixture of both. Superimpositional vs Linear Some models have a superimpositional nature (Dutch, Fujisaki) where pitch accents are seen as being relatively short term which “riding” on top of a phrases which are seen as being relatively long. By contrast, in linear models (AM/Tilt) contours are composed of linear sequences of in-
- 272. 248 Chapter 9. Synthesis of Prosody tonational units. Often a grammar is used which states which units can follow each other, but so long as the sequence lies within the grammar, any unit can follow any other. Being brief, the main reason for supporting the superimpositional model is that F0 contours do seem to exhibit patterns of global behaviour, where phrases define particular F0 patterns. The main reason for supporting linear models comes from speech production concerns. Here, we see that if we support a superimpositional model then the speaker has to pre-plan the utterance some way ahead of time. This doesn’t seem to correspond to known facts in speech production, where we know for example that a speaker can change the course of the F0 contour at any time. Finally we should note that some papers have shown how labels in one model can be trans- lated to labels in another [436], [374]. 9.4 INTONATION SYNTHESIS WITH AM MODELS We now turn our attention to this issue of intonation synthesis itself. In Chapter 6 we described a variety of techniques for generating the prosodic form representations of phrasing and prominence from text. We only made passing reference to the issue of generating intonational form from text as this really required a thorough discussion on the issue of intonational form first. With this job now done, we can describe techniques for both generating intonational form descriptions from text and for generating F0 controls from intonational form. While it is possible to build synthesis algorithms with any of the models outlined in Section 9.3.3, we will focus our attention on the most popular techniques, which divide into those based on AM models (described here) and those based on acoustic models (described next). 9.4.1 Prediction of AM labels from text We approach the task of predicting intonation from text in much the same way as we predicted phrasing and prominence from text. With the AM model, our job is to assign an intonational label to each syllable, and no matter what algorithm is used, this is formulated as a mapping from a set of features describing the syllable and its context to a label. A typical approach of this type is given by Lee [280] who trained a decision tree to generate ToBI labels using linguistic features such as position in phrase, part of speech tags and syntactic information. In general though, any of the standard text prediction machine learning techniques introduced in Chapter 5 can be used on this task, and the approach is by and large the same as for prominence prediction and phrasing prediction. Many other approaches for the prediction of AM and other intonation representations from text have been proposed, including HMMs [325], decision trees/CART [513], [378] and rules [462], [365].
- 273. Section 9.4. Intonation Synthesis with AM models 249 9.4.2 Deterministic synthesis methods Given a representation of intonational form predicted from text, we next attempt to synthesize an F0 contour from this. In this section we will examine two rule based deterministic proposals for doing this, the first developed for use the with original Pierrehumbert model and the second, developed more specifically for use with ToBI. Anderson, Pierrehumbert and Liberman [16] proposed a synthesis scheme (hereafter called APL) which generated an F0 contour from an arbitrary AM tone description. Recall that a feature of the AM model is that it is linear in the sense that there is no global pattern to intonation, each intonational unit in the sequence operates largely independently. This is reflected in the synthesis algorithm where an acoustic model made of target points is used. In the algorithm, each intonational event generates a set of target points which define the basic shape of the event. Pitch range in the model is expressed by two lines, the reference line which expresses the default or “at rest” value of F0, and the baseline which is the minimum F0 value, and the value that F0 falls to at the end of a phrase which is terminated by an L- L% boundary tone. Declination as a global trend is absent from the model; rather, downdrift is synthesized by a trend where each successive pitch accent is in general placed lower in the pitch range. Between pitch accents, phrase accents and boundary tones, linear interpolation is used to “fill in” the F0 values. An important nuance in the algorithm is its treatment of phrase final behaviour. Following extensive studies reported in previous work [285], Anderson et al added a component to the algorithm which dropped the F0 contour to the baseline after the nuclear accent, regardless of whether the model itself would have predicted such a value. This greatly increases the naturalness of the contours produced as they have a proper impression of finality compared to other synthesis algorithms they investigated. A more recent algorithm designed to work with the ToBI scheme was proposed by Jilka, Mohler and Dogil [241]. This in many ways extends the original work by Anderson et al as it too synthesises the ToBI events as isolated units, places them in context in the utterance, and then uses interpolation to cover the regions between. The algorithm improves on the original work in that it pays particular attention to where target points should be placed with respect to their syllables, and for instance models in some detail the fact that accents that occur early in phrases are often later with respect to their syllable position than accents which occur later in phrases (its as if the phrase boundaries are repelling the accents in both directions). In this algorithm, a topline and baseline are used as the reference points. When downstepping occurs, it triggers a lowering in the topline, and this is the general mechanism by which downdrift occurs. Example rules for the position of the target points are given below:
- 274. 250 Chapter 9. Synthesis of Prosody Intonational event context resultant time position H* First syllable in phrase 85% Last syllable in phrase 25% One syllable phrase 50% All other cases 60% L* First syllable in phrase 85% Last syllable in phrase 25% One syllable phrase 50% All other cases 60% L+H* (H part) First syllable in phrase 90% Last syllable in phrase 25% One syllable phrase 70% All other cases 750% L+H* (L part) Normal case 0.2s before H* If voiceless region 0.2s before H* (reference point) 90 %of voiced region to left of referenc 9.4.3 Data Driven synthesis methods Black and Hunt [51] described a data-driven technique for synthesising F0 contours from ToBI labels. In fact, their work can be seen as a data-driven version of the APL algorithm, in which the target points are learned from the data rather than specified by hand. The approach was based on learning the mapping between ToBI labels, which had been marked on a database by hand, and F0 contours extracted automatically. The APL model defines target points at very specific locations, but these are difficult to determine bottom-up so the Black and Hunt approach instead uses three target points for each ToBI label. These are place the beginning, middle and end of the syllable associated with the label. This gives us a representation where the input is a ToBI label and a set of features describing the linguistic context of the label, and the output is three target points. In this form, the mapping is now amenable for use by a standard machine learning algorithm. Black and Hunt in fact used a linear regression technique, with features such as lexical stress, numbers of syllables between the current syllable and the end of the phrase, identity of the previous labels and so on. Once learned, the system is capable of generating a basic set of target points for any input, which we then interpolated and smoothed to produce the final F0 contour. Other data driven techniques such as CART have proven suitable for synthesizing from AM representations [292], [340]. 9.4.4 Analysis with Autosegmental models The biggest difficulty with the AM models is the difficulty in labelling corpora. Several studies have been conducted in labelling with ToBI with the general conclusion that while labellers can often identify which syllables bear pitch accents, they are very poor at agreeing on which particular
- 275. Section 9.5. Intonation Synthesis with Deterministic Acoustic Models 251 accent is correct. As we explained in Section 6.10, the inter-labeller agreement with models such as ToBI is worryingly low, and in many practical database labelling projects many of the pitch accent distinctions have been dropped [488], [487]. Two major problems stem from this. Firstly, any database which has been labelled with ToBI will have a significant amount of noise associated with the pitch accent label classes. Secondly, for any large scale machine learning or data driven approach, we need a considerable amount of labelled data to the extent that it is impractical to label data by hand. As we shall see in Chapters 15 and 16, virtually all other aspects of a modern data driven TTS system’s data are labelled automatically, and so it is a significant drawback if the intonation component can not be labelled automatically as well. Because however the level of human labeller agreement is so low, it is very hard to train a system successfully on these labels; we can hardly expect an automatic algorithm to perform better than a human at such a task. One solution that is increasingly adopted is to forgo the distinction between label types altogether, see for instance [488]. While the break index and boundary tone components are often kept, only a single type of pitch accent is used; in effect the labellers are marking whether a word is intonationally prominent or not. However, it should be clear that such an approach effectively reduces ToBI to a data driven system of the type described below. 9.5 INTONATION SYNTHESIS WITH DETERMINISTIC ACOUSTIC MOD- ELS Here we describe how to synthesise F0 contours using the Fujisaki and Tilt models and as we will see, the approach is similar in both cases. Synthesis from the Fujisaki or Tilt parameters is of course given by the models themselves, and so presents no problems. For TTS purposes then, the issue is to generate the Fujisaki or Tilt parameters from the text. Two basic approaches can be taken in this regard; either we try and generate the parameters directly from the text, or we take a two stage approach where we first generate an abstract, discrete prosodic form representation from the text, and then proceed to generate the model parameters from these. 9.5.1 Synthesis with superimpositional models The Fujisaki model is most commonly used with Japanese, but has been used or adapted to many other languages. In Japanese, we find that the range of pitch accent phenomena is narrower than in languages such as English, which means that the model’s single type of accent is particularly suited. In addition, the nature of intonation in Japanese means that accents are marked in the lexicon, which greatly simplifies the problem of prominence prediction. Hence a simple approach to this, which uses accent information from the lexicon alone, is often sufficient. A common approach therefore is to determine phrase breaks and prominent syllables from the text, and then phrase by phrase and syllable by syllable generate the input command parameters for the Fujisaki
- 276. 252 Chapter 9. Synthesis of Prosody model using one of the standard machine learning techniques. Hirose et al [206], [205] describes such an approach for Japanese. In doing so, he makes use of the basic lexicon tone distinctions in Japanese to given where accents occur, thus simplifying the prominence and accent prediction process considerably. To generate the accent commands, he uses parts of speech, syllable structure, relative positions of these, numbers of proceeding phrases and punctuation. Syntactic units are also optionally used. A number of neural network algorithms are used, including recurrent networks. In addition, multiple regression analysis in a vein similar to the ToBI Black and Hunt [51] technique was investigated, in most cases the results were very similar regardless of which algorithm is used (this of course follows the pattern for other prosody prediction problems). One significant advantage that Hirose makes note of is that even when the prediction algo- rithm generates incorrect parameters for the Fujisaki model, the result is still a “natural” contour because the model itself is tightly constrained. In other words, the contour produced is some con- tour, just not the right one. By comparison, when prediction is poor with target models, the result may be a contour that bears no similarity to the type that a human could produce. 9.5.2 Synthesis with the Tilt model Dusterhoff et al [146] described a three stage technique for synthesising intonation with the Tilt model. First, they predict where accents are placed; this is effectively a prominence prediction algorithm of the type described in Section 6.8. Next, for each accent, the three intrinsic and 2 extrinsic Tilt parameters are generated. This is done with a decision tree, using features such as syllable position in phrase, length of phrase and so on. Finally, the Tilt model itself generates the F0 contour. In general though there are many similarities between synthesis with Tilt and the Fujisaki model, so a technique which works with one can probably be adapted to the other. 9.5.3 Analysis with Fujisaki and Tilt models An advantage that the Tilt model has over many others is that automatic analysis is significantly easier using this model. Taylor [436] describes a complete technique for first finding pitch accents and boundary tones and then determining the Tilt parameters from the local F0 contour in the surrounding region. Many algorithms have also been developed for automatically finding the Fujisaki model parameters from data [165], [71], [314], [7], [379]. 9.6 DATA DRIVEN INTONATION MODELS A common criticism of the Fujisaki model [285], [436] is that there are many contours which the model can not generate. These include gradual low rising contours which occur when a low accent occurs early in a phrase, and the effect of phrase final lowering where the contour falls past the normal baseline to a final baseline in declarative utterances. Suggestions that fix the problems
- 277. Section 9.6. Data Driven Intonation Models 253 have been made, for instance by including phrase-final negative phrase commands and so on, but such suggestions detract from the elegance of the original model. The Tilt model can be criticised as over generating in that there are contours it can generate which are never observed. When we consider simpler stylisation models we see that these models often massively over generate. To solve these problems, we can of course refine the models, but this can result in endless “tinkering” in which the models loose their originally elegance and simplicity. The basic problem is that when we design a model, we wish it to be elegant and easy to interpret, but we always run the risk that some intonational phenomena that we haven’t taken into account may appear and be outside the model’s capabilities. We then face the choice of adding to the model, which often appears as a “hack” or designing a completely new model. We can however consider a quite a different approach where we completely abandon the idea of creating an explicit model, but instead infer the model from data. Bailly [29] has considered this issue and asked whether explicit intonation models are required at all; perhaps we should concentrate on databases and labelling, and leave the specifics of the model to be determined by modern machine learning techniques. As we shall see in Chapter 13, it is convenient to describe synthesis techniques in terms of three generations, where the first generation used explicit hand written rules, the second generation uses basic and quite constrained data driven approaches, and a third generation which used general data driven and statistical approaches. The same pattern can be seen in intonation synthesis, where for example the APL algorithm is first generation and the Black and Hunt decision tree ToBI algorithm is second generation. The techniques described here can be considered third generation as they match closely the approaches used in the HMM and unit selection synthesis techniques described in Chapters 15 and 16. In the following sections, we give an overview of the techniques used, but as many of these follow the synthesis techniques described in Chapters 15 and 16, we wait until then to give a fully formal account. 9.6.1 Unit selection style approaches The unit selection synthesis technique described in Chapter 16 uses an entirely data driven ap- proach, whereby recorded speech waveforms are cut up, rearranged and concatenated to say new sentences. Given the success of this approach in normal synthesis, a number of researchers have applied these algorithms to F0 synthesis [296], [310], [311]. F0 contours extracted from real utterances are by their very nature perfect. The basic idea therefore is to collect a database of naturally occurring F0 contours, and use these at synthesis time for new utterances. It is extremely unlikely that one of the complete contours in the database will be exactly what we want; for this to be the case we would have have to have exactly the same text for the synthesis and database utterance. At a smaller scale, say at the phrase, word or syllable level, it is however possible to find exact or close matches and so the database processed to create a set of F0 units. Each unit has a set of features describing it, and at run time, these features are compared with features generated by the prosodic prediction. Sometimes exact matches are found,
- 278. 254 Chapter 9. Synthesis of Prosody but often only close matches are found in which case a distance, called the target cost is calculated from the desired features to the ones for each unit in the database. The lower the distance, the closer a unit is to the desired features. Obviously we wish to avoid large discontinuities in F0 values when we move from one unit to the next, and so a second distance, called the join cost is used to give a measure of how similar two units are at their edges. For F0 contours, a simple metric such as absolute F0 difference is often sufficient. The idea then is to find the best sequence of F0 units for our desired features, and this is calculated by summing the target and join costs for a complete sequence of units. Formally, this is given by: C(U,S) = T ∑ t=1 T(ut,st)+ T−1 ∑ t=1 J(ut,ut+1) (9.4) where U is a sequence of units from the database, S is the sequence of desired features, ut is one unit, st is one set of desired features and T(.) and J(.) are the target and join costs respectively. A search is then conducted to find the single sequence of units Û which minimises this cost: Û = argmin u T ∑ t=1 T(ut,st)+ T−1 ∑ t=1 J(ut,ut+1) (9.5) The advantage of this approached is that we are guaranteed to synthesise a natural contour, the only question is whether it will be the “correct” natural contour for the linguistic context. This approach is appealing in that is simple and should work for all languages and genres. Furthermore, as database sizes grow, the quality of this technique will continually improve as we will become more and more likely to find good matches in the database. 9.6.2 Dynamic System Models The main drawback of the above unit selection approach is that the algorithm suffers from a “hit or miss” problem with respect to finding matches in the database. In other words, we have a fixed number of units, and it is sometimes a matter of luck whether we find a good match or not. While the technique can produce excellent F0 contours, occasionally no good matches are found and the results can be poor. Unit selection approaches can be seen as ones where certain F0 contours are memorised by the system. We can contrast this with statistical approaches, where the aim is to learn the general nature of all F0 contours, by learning model parameters (not memorising) from a limited amount of training data. The advantage is that the technique is more robust when we need to synthesize contours we have not previously come across. The downside is of course that in building our model we have made some assumptions, and so the reproduction of F0 contours may not be as accurate for cases that we have in fact seen. Ross and Ostendorf developed a model based on the dynamic system model, a model well known in control engineering (also know as a Kalman filter). This model is described by a state equation and an observation equation, given respectively as follows:
- 279. Section 9.6. Data Driven Intonation Models 255 xk+1 = Fjxk +uj +wk (9.6) F0k = Hjxk +bj +vk (9.7) The idea here is that the state equation describes a trajectory through time. The above is a first order IIR filter of the type we will describe in Section 10.4, for now all we need know is that by appropriate setting of the parameters Fj, uj and wk, we can generate a wide variety of trajectory shapes. By inferring these parameters from data, we can train this equation to generate contours. The term uj can be seen as an input, and can be thought of as the equivalent of the phrase or accent command in the Fujisaki model. wk is a term for modelling noise, and is what gives the model its statistical nature; for now we can think of this as an error term in the model, which allows us to robustly and statistically learn the other parameters. Rather than use this equation as is, we also use the observation equation which performs a mapping from each point in the state trajectory to an actual F0 value, again with trainable param- eters determining the exact nature of this mapping. While we should be somewhat cautious about giving explicit interpretations to statistical models, we can think of the overall process as follows. The state equation generates a “true” intonation contour that properly expressed the actual lin- guistic manifestation of the prosodic form. These trajectories are smooth, clean and simple. We know however that real F0 values, as measured by pitch detection algorithms (see Section 12.7) are noisy, in that the contours are rarely smooth or predictable at the lowest level. The observation equation models this part of the system. If we used the output of the state equation to model F0 values directly, our problem would be significantly harder as we would then have to model every perturbation in the contour, and would miss the general and underlying trends in the contour. During synthesis, the model is “driven” by means of the input parameters uj and bj and also optionally by varying Fj and Hj. This can be done in a number of ways, but Ross and Osterndorf use simple equations which link the usual linguistic features (position within the phrase, stress value of the syllable etc) to change the values of the parameters every syllable. During synthesis, the noise terms are ignored. Training the model is somewhat complicated, but in essence we attempt to generate F0 contours on a training corpus and adjust the parameters until the best match between is found between the F0 contours produced by the model and the natural ones. This approach is very similar to the expectation maximisation (EM) technique used to train hidden Markov models. This is explained in full in Section 15.1.8. 9.6.3 Hidden Markov models The dynamical system model is a natural choice for statistical generation of F0 contours as it is well suited to the job of generating continuous trajectories. If it has any weakness we can point to the fact that the state trajectories are limited to being those of a first order filter, the noise terms have to be Gaussian and the training process can be quite intricate. An alternative is to use hidden
- 280. 256 Chapter 9. Synthesis of Prosody Markov models (HMMs) as these are in general easier to train and allow more complexity with regard to noise/covariance terms. HMMs are however similar to dynamic system models in certain ways in that they have a hidden state space, within which the model dynamics operate, and a projection from this to an observation space. We have used HMMs before for various tasks including part-of-speech tagging and phrase-break prediction. In those cases, the output distributions were discrete and gave the probability of seeing (say) a word given a POS tag. It is also possible to have HMMs with continuous output distributions which give the probability of seeing a continuous value given a state. This sort of HMM can therefore be used for F0 generation. In the dynamic system model, the hidden state space is a continuous trajectory, and this seems very natural for F0 contours. HMMs by contrast have a discrete state space, and when we move from one state to another during generation, the observation probabilities suddenly change. In this mode an HMM will generate extremely unnatural contours where we have a series of constant F0 values, a jump to a different series of constant F0 values and so on. No real F0 contours have behaviour anything like this. If we however adopt the formulation of Tokuda et all [453], we can generate from the HMM states while also taking into account the learned dynamics of F0 behaviour. Inanoglu [230] describes a system where F0 for various emotions can be accurately generated by HMMs (diagrams of this are shown in Figures 15.14 and 15.15). This method of HMM synthesis is the subject of Chapter 15 and so we leave further explanation of HMM F0 synthesis until then. We finish however with the observation that by using this formulation, extremely natural F0 contours can successfully be generated with HMMs. 9.6.4 Functional models Given the considerable difficulties in constructing a model of prosodic form, never mind any difficulties in using such a model, we can ask whether an explicit model of prosodic form is really necessary. While experience has shown that some notion of prosodic form is necessary (see Section 9.1.2), it may not in fact be necessary to model this explicitly; rather we can learn a model of prosodic form and keep it as a hidden part of a more general model [501], [502]. The superposition of functional contours (SFC) [216], [321], [217], [31] is a model in which the mapping from the meaning and linguistic structure to F0 is attempted in a unified fashion. The SFC model is based on the idea that there are a number of identifiable prosodic functions3 and that each has a characteristic prosodic pattern. As the scope of each function is variable, and as we can have several independent functions operating on a particular part of the utterance (for instance syntax and emotion), then we need to combine these effects together to form a single prosodic pattern. In the SFC model, the syllable is used as the unit of prosodic control and each is described by three F0 values and a duration value representing the prosodic lengthening to be applied. Each 3 Here the word “function” means “purpose” and is a combination of affective and augmentative prosody.
- 281. Section 9.7. Timing 257 prosodic function is assigned a generator, whose purposes is to generate the prosodic output for the particular sequence of syllables given when that prosodic function is active. A training database therefore comprises a set of utterances, each marked with the multiple prosodic functions and the 4-parameter descriptions for each syllable. Each generator is then trained by an iterative procedure which learns globally what prosodic patterns each generator should produce. In the original paper Bailly and Holm used neural networks as the generators, but in principle, any trainable model can be used for this, including the dynamic system models and HMMs described above. The advantage of this approach is that realistic and natural prosody can be generated without the need for determining which syllables receive pitch accents and what type they should be. All this is modelled and optimised implicitly in the model. 9.7 TIMING Many aspects of intonation, prominence and phrasing are manifested in differences in the timing of speech. It is therefore the task of the timing module in a TTS system to assign a temporal structure to the utterance. The first question then to ask is how exactly should we specify these timing values? Most commonly this is formulated in terms of unit durations, although other models assign timing values to significant points in the utterance [35]. The question then is, what type of units should we assign these durations to? It is well known from experimental phonetics that the temporal properties of speech are complex. If one talks more quickly, it is not the case that every part of the utterance contracts by a constant factor. Emphasis generally lengthens sections of speech, but in very particular ways. At a micro-level, we know that the articulators all have specific dynamic properties and when a phone is spoken at a faster rate, some articulators may undershoot their target more than others. A completely comprehensive approach to timing would therefore attempt to model all these factors, but in general it is very hard to determine any reliable anchor points at a level lower than the phone. In practical TTS, the complex sub-phone timing patterns are often ignored as firstly it is difficult to predict the interactions and secondly it is subsequently difficult to know what to do with the sub-phone information once obtained. In practice then the most common approaches are to predict syllable durations or phone durations. Syllables are attractive as they seem to be the natural units of prosody; we have seen elsewhere that they are the units of prominence and therefore pitch accent placement. In addition, we usually think of the syllable as being the natural unit of rhythm. Phones are attractive as they are the smallest units we can predict and therefore allow the most detail in prediction. 9.7.1 Formulation of the timing problem Unlike intonation, timing synthesis can timing synthesis can in fact be formulated as a regression algorithm directly as our goal is to predict a single continuous value from a set of features. Re-
- 282. 258 Chapter 9. Synthesis of Prosody gardless of whether we are predicting a phone duration or syllable duration, the problem can be formulated as a mapping from a set of features to a single value, expressed in seconds: X(f1, f2,...fN) → di Apart from the fact that we are predicting a continuous value instead of a class label, this formulation is the same as we have been using throughout the book for other classification prob- lems and so it is natural to apply these approaches directly to the duration problem. In addition to these, a number of specialised approaches have also been developed for use with this problem. 9.7.2 The nature of timing The most common view of timing is that it is governed by both phonetics and prosody. Regardless of other factors, the fundamentals of articulatory speech production give a basic temporal pattern to speech. The articulation required to produce some vowels, such as /ii/ mean that they are inherently longer than others such as /o/. In addition to these, largely “automatic” aspects of timing, we also have phonetic timing factors that are part of the allophonic system of the language. The classic example in English is with words such as MELT and MELD where we see that the vowel in MELD is considerably longer than in MELT. This is a consistent phenomenon (BENT vs BEND, TIGHT vs TIDE), but is not a physiological phenomenon alone; it is quite possible to speak these word pairs with equal durations, and some accents of English, most noticeably Scottish do in fact do this. In general however, it seems that the identity and interactions of phones has a strong bearing on the duration of each. In addition to this we have prosodic factors. It is well observed that the durations at the ends of phrases are longer than elsewhere. The extent of this phrase final lengthening is not completely understood, but roughy speaking we expect the last syllable in a phrase to be longer than otherwise, and that the amount of lengthening in each phone increases towards the end of that syllable. It is important to realise that more than timing is affected by the presence of a phrase end; voice quality changes here too and the in general the speech is spoken with considerably less vocal effort. The other main prosodic factor affecting timing is prominence, with prominent syllables be- ing considerably longer than non-prominent ones. Although not a proper prosodic factor, lexical stress behaves in the same way so that lexically stressed syllables are longer than unstressed syl- lables. It is hard to determine whether intonation has much of an effect on duration as all pitch accents lie on prominent syllables and so it is difficult to separate the effects of each. One ob- served phenomenon occurs when a complex intonation form is used with a very short, say single syllable, utterance. If we say the name SUE? with a doubting intonation pattern which has a low accent followed by a rise, we find that the length of the word seems to stretch to accommodate the complex pitch pattern.
- 283. Section 9.7. Timing 259 9.7.3 Klatt rules One of the most widely set of deterministic rules was developed by Dennis Klatt [253], [10]. Although designed for use with the MITalk formant synthesiser, these rules can be used with any synthesis technique. In this, the basic properties of each type of phone are given by two values, a minimum duration M and an inherent duration . The inherent duration is a sort of average, but is best thought of as the duration of the phone spoken in a neutral canonical context. The final is then given by Duration = [(inherent duration −minimum duration)×A]+minimum duration (9.8) The term A is calculated by the successive application of a set of rules, where each firing rule multiples A by a factor. The factors are calculated as follows: clause-final lengthening If a segment is the vowel or in the coda of a clause-final syllable, then A = 1.4A. non-phrase-final shortening If a segment is not in a phrase final syllable A = 0.6A. If the seg- ment is a phrase-final postvocalic liquid or nasal A = 1.4A. non-word-final shortening If a segment is not in a word final syllable A = 0.85A. Poly-syllabic shortening If a vowels is in polysyllabic words, A = 0.80A. Non-initial-consonant shortening If a consonants is in non-word-initial position, A = 0.85A. Unstressed shortening Unstressed segments are half again more compressible; so that minimum duration = minimum duration/2. If the segment is unstressed or reduced, set A as follows: • vowel in word medial syllable: A = 0.5A • other vowels: A = 0.7A • prevocalic liquid or glide: A = 0.1A • all others: A = 0.7A Prominence if the segment is in a prominent syllable, A = 1.4A. Post vocalic context of vowels A vowel is modified by the consonant that follows it in the fol- lowing ways: • no following consonants, word final: A = 1.2A • before a voiced fricative: A = 1.6A • before a voiced plosive: A = 1.2A • before a nasal: A = 0.85A • before a voiceless plosive: A = 0.7A • before all others voiced fricative: A = A These effects are less in non-phrase final positions, in which case set A = 0.7+(0.3A)
- 284. 260 Chapter 9. Synthesis of Prosody Shortening in clusters • vowel followed by a vowel: A = 1.2A • vowel preceded by a vowel: A = 0.7A • consonant surrounded by consonants: A = 0.5A • consonant followed by a consonant: A = 0.7A • consonant preceded by a consonant: A = 0.7A In addition, Klatt also states that pauses should be 200ms long, and that non-reduced vowels preceded by voiceless plosives should be lengthened by 25ms. 9.7.4 Sums of products model Van Santen [411], [464] proposed the sums-of-products model which can be seen as a semi- trainable generalisation of the Klatt model described above. Van Santen makes a few key obser- vations regarding the nature of duration prediction. Firstly he notes that many possible feature combinations will never be observed in the training data (this of course is the same sparse data problem we have come across before). Crucially however, he notes that duration modelling can be seen as a well behaved regression modelling problem, in which certain properties of the prob- lem allow us to make assumptions that can help ease data sparsity. Specifically, he states that to a large degree the factors influencing duration are monotonic (or “directionally invariant” in his terms). In practice this means, while we know that emphasis and phrasing both affect the length of a phone, and that while strictly speaking we can not consider these features as being indepen- dent, their general influence on the duration is well behaved, such that emphasis always lengthens a phone as does phrasing; it is not the case that the effects of these features ever reverses. This means that it is possible to build a model where the interactions of these features do not have to be comprehensive. In general, models of this type will not work because they don’t model interactions be- tween the features. In van Santen’s approach however, the process of building the model is to use knowledge to design the features. Hence the features actually used in the model will often be transformations of the ones generally available in the utterance structure. This process of hand design helps produce features which can operate more independently than the original features. The parameters can be set in a number of ways; how this is done is independent of the model itself. A sampling approach is one possibility. 9.7.5 The Campbell model Campbell [83] proposed a model that uses the syllable as the fundamental unit of duration. In this a syllable duration is predicted from a set of linguistic features after which the individual phone durations within the syllable are then calculated. This approach has the attraction in that it is more modular, where we have one component modelling the prosodic part of the duration and another modelling the phonetic part.
- 285. Section 9.8. Discussion 261 The prosodic part of the Campbell model uses a neural network to calculate a syllable du- ration. This is an attractive approach as the neural network, unlike the Klatt or sums-of-products model, can model the interactions between features. A somewhat awkward aspect of the model however is the fact that the syllable duration itself is of course heavily influenced by its phonetic content. If left as is, the phonetic variance in model duration may swamp the prosodic variance that the neural network is attempting to predict. To allow for this, Campbell also includes some phonetic features in the model. Campbell maps from syllable to phone durations using a model based on his elasticity hypothesis which states that each phone in a syllable expands or contracts according to a constant factor, normalised by the variance of the phone class. This operates as follows. A “mean” syllable is created containing the correct phones with their mean durations. This duration is then compared to the predicted duration, and the phones are either expanded or contracted until the two durations match. This expansion/contraction is performed with a constant variance, meaning that if the vowel is expanded by 1.5 standard deviations of its variance, the constant before it will be expanded by 1.5 standard deviations of its variance. The problem of the absolute syllable duration can be easily solved by having the neural network predict this z-score instead of an absolute duration, an approach followed by [456]. This then frees the neural network from phonetic factors completely and allows it to use only prosodic features as input. In general this model is quite effective as it solves the feature explosion problem by positing a modular approach. One significant weakness however is that the elasticity hypothesis is demon- strably false. If it was true then we would expect the z-scores for all the phones in a syllable to be the same, but this is hardly ever the case; depending on context, position and other features the z- scores for phones across a syllable vary widely. This is a problem with only the second component in the model and a more sophisticated model of syllable/phone duration interaction could solve this. In fact, there is no reason why a second neural network could not be used for this problem. 9.7.6 Other regression techniques As we might expect by this stage in the book, most of the usual classification/regression algo- rithms have been applied to the duration prediction problem. These include decision trees [372],, neural networks for phone prediction [109], [157], genetic algorithms [319] and Bayesian belief networks [182]. Comparative studies of decision trees and neural networks found little difference in accuracy between either approach, [473], [187], [72]. 9.8 DISCUSSION In this chapter, we have considered techniques for synthesizing the acoustics of prosody. Any discussion on this topic will draw more or less the same conclusions as given in the discussion in Chapter 6, where we explained that there was an enormous range of models, theories and explana-
- 286. 262 Chapter 9. Synthesis of Prosody tions for prosodic phenomena. The consequence of this is that developing engineering solutions to practical prosodic problems was considerably more difficult than with the verbal component of language. Hence the problem is complicated because there is no commonly agreed way of representing the linguistic form of the phenomena we are attempting to synthesize. While many of the AM models and deterministic acoustic models provide useful and ad- equate representations for intonation, the trend is clearly towards the data driven techniques de- scribed in Section 9.6. These have several advantages; besides bypassing the thorny theoretical issues regarding the “true nature” of intonation, they have the ability to automatically analyse databases, and in doing are also inherently robust against any noise that can occur in the data, whether it be from errors in finding F0 values or from other sources. The purpose of all the prosody algorithms described in this chapter is to provide part of the specification which will act as input to the synthesizer proper. In the past, the provision of F0 and timing information was uncontested as a vital part of the synthesis specification and most of today’s systems still use them. As we shall see in Chapters 15 and 16, however, some third generation systems do not require any acoustic prosody specification at all, making used of higher level prosodic representations instead. Rather than use F0 directly, stress, phrasing, and discourse information are used. While such an approach completely bypasses all the problems described in this chapter, it does have the consequence of increasing the dimensionality of the feature space used in unit selection or HMM synthesis. It is therefore a practical question of tradeoff whether such an approach is better than the traditional approach. Most work on the acoustics of prosody has focused on the modelling of F0 and timing. This has mainly been because these are the parts of prosody which are most readily identifiable and separable from the influence of verbal factors. In addition, these are the aspects of speech which signal processing algorithms find easiest to modify. It is quite clear however that prosodic factors have a strong influence on a range of other acoustic aspects of the signal. For instance at the ends of phrases the vocal effort is considerably reduced, and the overall voice quality is quite different than elsewhere. To date, only a few studies have considered these effects [424], [417], [183]. If the prosodic influence on voice quality is not modelled within a TTS system the naturalness with definitely suffer, but it is unclear whether these affects should be modelled explicitly, or implicitly within a unit selection or HMM synthesizer. 9.8.1 Further Reading Articles on all the models and techniques described here are readily available, but there are few which consider all models together and offer comparisons. The best account of the AM model is given by Ladd [269]. Ladd not only describes the model and its development, but also serves as a solid introduction to the whole field of intonational phonology and the practical phonetics associated with it. Accounts of the INTSINT model are given in Hirst [212], the Tilt model in Taylor [436], the Fujisaki model in Fujisaki [166], the SFC model in Bailly and Holm [31], the Dutch model in [447], the British school in O’Connor and Arnold [333].
- 287. Section 9.8. Discussion 263 9.8.2 Summary Prosodic form and acoustics • The history of intonation research is full of attempts at finding a system of prosodic form which could act as the equivalent of phonemes in the verbal component of language • No consensus has been reached on this issue, which has lead to the large number of quite different models that have been developed. • F0/pitch and timing are the two main acoustic representations of prosody, but this is largely because these are quite easy to identify and measure, other aspects of the signal, most notably voice quality are heavily influenced also. Intonation models • The British school describes intonation contours as comprising a pre-head, head, nucleus and tail, and each has an inventory of shapes. • The AM school (which includes ToBI) describes intonation in terms of abstract High and Low tones. Diacritics (*, %, -) are used to specify which tones align with syllables and boundaries. The tones can be combined in various ways to form an inventory of pitch accents (e.g. H* +L). • The INTSINT model provides a way of labelling intonation across theories and languages, and provides a way of labelling data for further analysis in much the same was as the IPA is used in phonetics. • The Fujisaki model describes intonation as a sequence of commands with continuous param- eters which when fed into the appropriate filters generates an F0 contour. The model has an accent component and phrase component which are superimposed to give the final contour • The Tilt model describes intonation with a sequence of events, each of which is described by a set of continuous parameters. • Many more intonation models exist, some of which are modifications of the above. Intonation Synthesis • The acoustic models, such as Tilt and Fujisaki, have synthesis algorithms “built-in” and hence synthesis is fairly trivial with these. • The phonological models usually have independently developed synthesis algorithms • In recent years, the notion of explicit models has been challenged by a number of data-driven techniques which learn intonation effects from data. • Some models, such as SFC, do away even with notions of explicit prosodic form and learn the all the processes and representations of intonation from data.
- 288. 264 Chapter 9. Synthesis of Prosody Timing • The purpose of a timing algorithm is to assign a time, usually in the form of a duration to each linguistic unit. • Nearly all algorithms take the form of a regression algorithm, where a number of linguistic features are used as input to a algorithm which outputs a time.
- 289. 10 SIGNALS AND FILTERS This chapter introduces the fundamentals of the field of signal processing, which studies how signals can be synthesised, analysed and modified. Here, and for the remainder of the book, we use the term signal in a more specific sense that before, where we take it to mean a waveform that represents a pattern of variation against time. This material describes signals in general, but serves as a precursor to the following chapters which describe the nature of speech signals and how they can be generated, manipulated and modified. This chapter uses the framework of Digital Signal Processing, a widely adopted set of techniques used by engineers to analyse many types of signals. 10.1 ANALOGUE SIGNALS A signal is a pattern of variation thats encodes information. Signals which encode the variation of information over time can be represented by a time waveform, which is often just called a waveform. Figure 10.1 shows an example speech waveform. The horizontal axis represents time and the vertical axis represents amplitude, hence the figure shows how the amplitude of the signal varies with time. The amplitude in a speech signal can represent different physical quantities: for example, the variation in air pressure in front of the mouth, the displacement of the diaphragm of a microphone used to record the speech or the voltage in the wire used to transmit the speech. Because the signal is a continuous function of amplitude over time, it is called an analogue signal. By convention, analogue signals are represented as a continuous function x over time t. This is, if we give a value for time, the function x(t) will give us the amplitude at that point. Signals can be classified as periodic, that is signals which repeat themselves over time, or -20000 -15000 -10000 -5000 0 5000 10000 15000 20000 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 ampl i t u de time Figure 10.1 An example speech wave of /... ii th .../ 265
- 290. 266 Chapter 10. Signals and Filters -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 0 2 4 6 8 10 12 14 16 amplitude time (a) The sine wave -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 0 2 4 6 8 10 12 14 16 amplitude time (b) The cosine wave Figure 10.2 Sine and cosine waves aperiodic, that is signals which do not. Strictly speaking, no speech sound is ever exactly periodic because there is always some variation over time. Such signals are termed quasi-periodic. In Figure 10.1 it should be clear that the vowel sound is quasi-periodic, whereas the fricative is aperiodic. In the subsequent sections we will build a framework for describing signals, by first considering periodic signals and then moving on to considering any general signal. 10.1.1 Simple periodic signals: sinusoids The sinusoid signal forms the basis of many aspects of signal processing. A sinusoid can be represented by either sine function or a cosine function x(t) = sin(t) x ′ (t) = cos(t) It may not be immediately clear why the sine and cosine function, which we probably first en- countered in trigonometry, have anything to do with waveforms or speech. In fact it turns out that the sinusoid function has important interpretations beyond trigonometry and is found in many places in the physical world where oscillation and periodicity are involved. For example, both the movement of a simple pendulum and a bouncing spring are described by sinusoid functions. Plots of these functions against time are shown in Figure 10.2, and from these it should be clear that both are periodic, that is that they exactly repeat over time. The period T is the length of time between any two equivalent points on successive repetitions. In the simple sinusoid, the period is T = 2π; that is, for any give point, the signal has the same value 2π seconds later. From Figure 10.2 we see that the sine and cosine function have identical shape, and differ only in their alignment. That is, if we shifted the cosine wave π/2 to the right, it would exactly match the sine wave: sin(t) = cos(t +π/2)
- 291. Section 10.1. Analogue signals 267 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 0 50 100 150 200 250 300 amplitude time (a) The cosine function with phase shifts −1,−2,−3,−4 -1.5 -1 -0.5 0 0.5 1 1.5 0 0.05 0.1 0.15 0.2 amplitude time (b) The cosine function with frequency 100Hz Figure 10.3 The extra term in the cosine function is called the phase shift, denoted φ, and can be used to move sinusoids earlier and later in time. When φ = π/2 the cosine exactly aligns with the sine, but different values will produce different alignments, as shown in figure 10.3a. It should be clear, that because the signal has period 2π, φ need only have values in the range 0 to 2π to account for all possible phase shifts. Because a sine can be represented by a cosine with an appropriate phase shift and vice versa, we can use either sines or cosines exclusively. By convention, cosines are generally used, giving the general formula: x(t) = cos(t +φ) We define frequency, F, as the number of times the signal repeats in unit time, and this is clearly the reciprocal of the period: F = 1 T (10.1) The frequency is measured in cycles per second, Hertz, and so the frequency of our sinusoid is 1/2π Hz. A sinusoid of a different frequency can be generated by multiplying the variable t inside the sine function. For a frequency of 1Hz, we multiply t by 2π, and so for any frequency F expressed in Hz, we multiply t by 2πF. Figure 10.3b shows a sinusoid.with frequency 100Hz. x(t) = cos(2πFt +φ) To save us writing 2π everywhere, a quantity called angular frequency is normally used, which is denoted by ω and has units radians per second: ω = 2πF = 2π T (10.2)
- 292. 268 Chapter 10. Signals and Filters -1.5 -1 -0.5 0 0.5 1 1.5 0 0.005 0.01 0.015 0.02 0.025 0.03 amplitude time Figure 10.4 A square wave of fundamental frequency 100Hz The parameters ω and φ can thus be used to create a sinusoid with any frequency or phase. A final parameter, A is used the scale the sinusoid and is called the amplitude. This gives the general sinusoid function of: x(t) = Acos(ωt +φ) (10.3) 10.1.2 General periodic signals Many non-sinusoidal waveforms are periodic: the test for a periodic waveform is that there exists some period T, for which the following is true: x(t) = x(t +T) = x(t +2T) = x(t +3T)... (10.4) That is, at multiples of T the value is always the same. The lowest value for which this is true is called the fundamental period, denoted T0. The frequency F0 = 1/T0 is called the fundamental frequency , and its angular equivalent is ω0 = 1/2πT0. A harmonic frequency is any integer multiple of the fundamental frequency, 2F0,3F0,.... It can be shown that any periodic signal can be created by summing sinusoids which have frequencies that are harmonics of the fundamental. If we have a series of harmonically related sinusoids numbered 0,1,2,...,N, then the complete signal is given by: x(t) = a0cos(0×ω0t +φ0)+a1cos(1×ω0t +φ1)+a2cos(2×ω0t +φ2)+... (10.5) were appropriate setting of the amplitudes ak and phase shifts φk is used to generate the given waveform. As 0 × ω0t = 0 and a1cos(φ0) is a constant, we can combine cos(φ0) and a0 into a new constant A0, leaving the first term as simply this new A0. This can be thought of as an “amplitude offset” (or “DC offset” in electrical terminology), so that the waveform does not have to be centred on the y-axis. In the most general case, the number of terms can be infinite, so the general form of
- 293. Section 10.1. Analogue signals 269 -1.5 -1 -0.5 0 0.5 1 1.5 0 0.005 0.01 0.015 0.02 0.025 0.03 amplitude time (a) superimposed sine waves of the first and third harmonics -1.5 -1 -0.5 0 0.5 1 1.5 0 0.005 0.01 0.015 0.02 0.025 0.03 amplitude time (b) superimposed sine waves of the first to seventh harmonics -1.5 -1 -0.5 0 0.5 1 1.5 0 0.005 0.01 0.015 0.02 0.025 0.03 amplitude time (c) synthesised wave of the first and third harmon- ics -1.5 -1 -0.5 0 0.5 1 1.5 0 0.005 0.01 0.015 0.02 0.025 0.03 amplitude time (d) synthesised waves of the harmonics 1 to 7 -1.5 -1 -0.5 0 0.5 1 1.5 0 0.005 0.01 0.015 0.02 0.025 0.03 amplitude time (e) synthesised waveform of harmonics 1 to 29 Figure 10.5 Fourier synthesis of a square wave for different numbers of harmonics
- 294. 270 Chapter 10. Signals and Filters equation 10.5 is: x(t) = A0 + ∞ ∑ k=1 akcos(kω0t +φk) (10.6) Equation 10.6 is a specific form of the Fourier series, one the most important concepts in signal processing. The process as just described is known as Fourier synthesis, that is the process of synthesising signals by using appropriate values and applying the Fourier series. It should be noted that while the sinusoids must have frequencies that are harmonically related, the amplitudes and phases of each sinusoid can take on any value. In other words, the form of the generated waveform is determined solely by the amplitudes and phases. To demonstrate some of the properties of Fourier synthesis consider the task of synthesising a square wave, as shown in Figure 10.4, and defined by: x(t) = ( 1 for 0 ≤ t ≤ 1 2 T0 −1 for 1 2T0 ≤ t ≤ T0 (10.7) This signal can be generated by adding only the odd harmonics of the fundamental, using the same phase for every harmonic, and choosing appropriate amplitudes: ak = ( 4/kπ k = 1,3,5,... 0 k = 0,2,4,6,... (10.8) Figure 10.5 shows the process of creating a square waveform by adding more and more harmonics to the fundamental sinusoid. We have only been able to generate the square waveform because we knew the appropriate values for ak and φk for every harmonic. The converse problem, of finding unknown ak and φk for a known waveform function x(t) is termed Fourier analysis and will be described in section 10.1.4. 10.1.3 Sinusoids as complex exponentials While the general form of the sinusoid, x(t) = Acos(ω0t + φ) faithfully represents a sinusoid, it turns out that for types of operations we consider below, a different form of the sinusoid, known as the complex exponential greatly simplifies the mathematics and calculations. The basis of this representation is Euler’s formula which states: ejθ = cosθ+ jsinθ (10.9) where θ is any number and j = √ −1. The inverse Euler formulas are: cosθ = ejθ +e− jθ 2 (10.10) sinθ = ejθ −e− jθ 2j (10.11)
- 295. Section 10.1. Analogue signals 271 From here on, we will make frequent use of complex numbers, which comprise of a real part x, and an imaginary part jy, where j = √ −1, such that z = x+ jy. If we add an amplitude and set θ = ωt +φ, we get; Aejω+φ = Acos(ωt +φ)+ jAsin(ωt +φ) (10.12) where A, ω and φ are all real and have the same meanings as before. At first glance, this seems insane - why on earth should we replace our previous, easily interpretable representation of a sinusoid, A cos (ωt + φ), with a different equation, that includes an extra, imaginary term, jAsin(ωt +φ), and then rewrite it as an exponential? The reason is purely that it makes manipula- tion much easier. As a demonstration, consider the following, which allows us to represent phase differently: x(t) = Aejωt+φ = Aejφ ejωt (10.13) ejφ is a constant, not dependent on t, and we can combine this with the amplitude A to form a new constant, X, giving: x(t) = Xejωt = Aejωt+φ = Aejωt ejφ (10.14) Here, the new quantity X, is a complex amplitude and describes both the original amplitude and the phase. The purely sinusoid part, ejωt is now free of phase information. This is highly significant, because it turns out that in general, we want to describe, modify and calculate amplitude and phase together. (From here on, amplitude terms in complex exponential expressions will generally be complex - if the pure amplitude is required, it will be denoted |A|.) While the complex exponential form has an imaginary part. jsin(ωt) to the waveform, in physical systems such as speech, this does not exist and the signal is fully described by the real part. When requiring a physical interpretation, we simply ignore the imaginary part of ejωt. Using the complex exponential form, we can write a more general form of the Fourier syn- thesis equation: x(t) = ∞ ∑ −∞ akejkω0t (10.15) where ak is a complex amplitude representing the amplitude and phase of the kth harmonic: ak = Akejφk (10.16) Equation 10.15 is the general form of Fourier synthesis which sums harmonics from −∞ to +∞. It can be shown that for all real valued signals (that is ones like speech with no imaginary part), the complex amplitudes are conjugate symmetric such that a−k = ak. In this case, the negative
- 296. 272 Chapter 10. Signals and Filters harmonics don’t add any new information and the signal can be faithfully constructed by summing from 0 to ∞: x(t) = ∞ ∑ 0 akekω0t (10.17) In fact, the summation from −∞ to +∞ keeps the mathematics easier, so we often use the complete form. 10.1.4 Fourier Analysis What if we know the equation for the periodic signal x(t) and wish to find the complex amplitudes ak for each harmonic? The Fourier analysis equation gives this: ak = 1/T0 Z T0 0 x(t)e− jkω0t dt (10.18) It should be clear, that integrating a sinusoid over a single period will give 0 for all sinusoids1 Z T0 0 sin(ω0t)dt = 0 This holds for any harmonic also - while we have more positive and negative areas, as we have an equal number of them then the sum will always be zero. In exponential form, we can therefore state: Z T0 0 ekω0t dt = 0 Now consider the calculation of the inner product of two harmonically related sinusoids, where one is the complex conjugate of the other Z T0 0 ejkω0t e− jlω0t dt this can be written: Z T0 0 eω0(k−l)t dt When k 6= l, this quantity k−l is an other integer, which will represent one of the harmonics. But we know from equation 10.1.4 that this will always evaluate to 0. However, when k = l, then k−l = 0 so the integral is Z T0 0 ejω00t dt = Z T0 0 1dt = T0 1 This can be seen by using the interpretation that integration over an interval is finding the area between the curve and the x-axis. From the sine wave in Figure 10.2 we see the area between 0 and π will cancel out the area between π and 2π.
- 297. Section 10.1. Analogue signals 273 -1.5 -1 -0.5 0 0.5 1 1.5 0 0.005 0.01 0.015 0.02 amplitude time (a) The square wave and sine waves of the first harmonic -1.5 -1 -0.5 0 0.5 1 1.5 0 0.005 0.01 0.015 0.02 amplitude time (b) multiplication of square wave and sine wave of the first harmonic -1.5 -1 -0.5 0 0.5 1 1.5 0 0.005 0.01 0.015 0.02 amplitude time (c) The square wave and sine wave of the second harmonic -1.5 -1 -0.5 0 0.5 1 1.5 0 0.005 0.01 0.015 0.02 amplitude time (d) multiplication of square wave and sine wave of the second harmonic -1.5 -1 -0.5 0 0.5 1 1.5 0 0.005 0.01 0.015 0.02 amplitude time (e) The square wave and sine wave of the third harmonic -1.5 -1 -0.5 0 0.5 1 1.5 0 0.005 0.01 0.015 0.02 amplitude time (f) multiplication of square wave and sine wave of the third harmonic Figure 10.6 Fourier synthesis of a square wave for different numbers of harmonics. We can see that the curve in figure d has symmetry about the x-axis and so there is an equal amount of area above and below the x-axis. The total area is therefore 0, hence the contribution of the second harmonic to a square wave is 0. Figures b and f have more area beneath the x-axis than above, hence the total area will sum to a non-zero value.
- 298. 274 Chapter 10. Signals and Filters Using these results, we can state the orthogonality property of sinusoids: Z T0 0 ejkω0t e− jlω0t dt = ( 0 if k 6= l T0 if k = l (10.19) Imagine we have a signal x(t) = ejlω0t of known fundamental period but unknown harmonic (that is, we know ω0 but not l). We can calculate the inner product of this signal with each harmonic k = 1,2,3... of the fundamental. All the answers will be zero except the one for when the harmonics are the same (k = l). As a general periodic signal is just the sum of a set of harmonics, we can use this method to find the value of every harmonic in a signal. Mathematically, we show this by starting with the Fourier series: x(t) = ∞ ∑ −∞ akejkω0t If we multiply both sides by the complex conjugate exponential and integrate over period T0 we get: Z T0 0 x(t)e− jlω0t dt = Z T0 0 ∞ ∑ −∞ akejkω0t e− jlω0t dt = ∞ ∑ −∞ ak Z T0 0 ejkω0t e− jlω0t dt = ∞ ∑ −∞ ak Z T0 0 e− j(k−l)ω0t dt = alT0 (10.20) and if we replace l with k we get the Fourier analysis definition. Fourier analysis can easily be demonstrated graphically. Figure 10.6 shows the multiplica- tion of a square wave with sinusoids of different harmonics. It can be seen from the graphs that the first and third harmonics have a non-zero area, and so their Fourier coefficient will be non-zero. The multiplication of the square wave and second harmonic produces a curve which is symmetri- cal about the x-axis and so the area beneath the x-axis will exactly match the area above, resulting in a total area of 0. One interesting point about the Fourier synthesis and analysis equations is that while they have a similar form, one is a sum while the other is an integral. This is solely because periodic signals are composed of a integer number (possibly infinite) of discrete harmonics, necessarily enforcing a discrete frequency description. The waveform however, is a continuous function. As we shall see, there are several variations on these equations in which frequency can be continuous or time discrete.
- 299. Section 10.1. Analogue signals 275 -1.5 -1 -0.5 0 0.5 1 1.5 -1500 -1000 -500 0 500 1000 1500 amplitude frequency /tmp/tmpSdIp-B using 1:2 (a) real spectrum -1.5 -1 -0.5 0 0.5 1 1.5 -1500 -1000 -500 0 500 1000 1500 amplitude frequency /tmp/tmphGyZno using 1:2 (b) imaginary spectrum 0 0.2 0.4 0.6 0.8 1 1.2 1.4 -1500 -1000 -500 0 500 1000 1500 amplitude frequency /tmp/tmp-07j6o using 1:2 (c) magnitude spectrum -4 -3 -2 -1 0 1 2 3 4 -1500 -1000 -500 0 500 1000 1500 amplitude frequency /tmp/tmpaPw0bS using 1:2 (d) phase spectrum Figure 10.7 The two ways of showing the full spectrum for a square wave. For a square wave, the real part is zero for all harmonics, and the phase is π for all harmonics. 10.1.5 Frequency domain The previous sections have shown how arbitrary periodic signals can be composed of harmonics with appropriate amplitudes and phases. It is often very useful to study the amplitude and phases directly as a function of frequency. Such representations are referred to as spectra and are de- scribed as being in the frequency domain. By contrast, waveforms as a function of time are said to be in the time domain. Because the harmonic coefficients are complex, we can plot the spectrum in either Carte- sian or polar form. The Cartesian form has a real spectrum and imaginary spectrum as shown
- 300. 276 Chapter 10. Signals and Filters -1.5 -1 -0.5 0 0.5 1 1.5 -1500 -1000 -500 0 500 1000 1500 amplitude frequency /tmp/tmphBTt1l using 1:2 (a) real spectrum -1.5 -1 -0.5 0 0.5 1 1.5 -1500 -1000 -500 0 500 1000 1500 amplitude frequency /tmp/tmpDlzYtT using 1:2 (b) imaginary spectrum 0 0.2 0.4 0.6 0.8 1 1.2 1.4 -1500 -1000 -500 0 500 1000 1500 amplitude frequency /tmp/tmpY-nEpG using 1:2 (c) magnitude spectrum -4 -3 -2 -1 0 1 2 3 4 -1500 -1000 -500 0 500 1000 1500 amplitude frequency /tmp/tmpi4kvjA using 1:2 (d) phase spectrum Figure 10.8 Spectra for square wave with delay. The real and imaginary spectra are clearly differ- ent from figure 10.7. The magnitude spectrum is the same, and the phase spectrum is the same apart from the additional shift. in Figures 10.7a and 10.7b . Alternatively, we can plot the coefficients in polar form with a mag- nitude spectrum and phase spectrum, as shown in Figures 10.7c and 10.7d. For the Cartesian interpretation, we see that the all the real parts are 0, and only the imaginary spectrum has values. This is equivalent to saying that no cosine functions were used, only sines: this can be seen in Figure 10.5. This can also be seen from the phase spectrum, were all the phases have a value of π/2, again showing that only sine waves are used. Now consider the square wave that is identical to the one we have been analysing, except that it is shifted 1/4 period to the left. If we subject this to Fourier analysis, we get the spectra
- 301. Section 10.1. Analogue signals 277 -6 -4 -2 0 2 4 6 0 0.0020.0040.0060.008 0.01 0.0120.0140.0160.018 0.02 amplitude time Figure 10.9 A waveform created from harmonics 1,3,5 and 7 of a “square” wave. The magnitude values are the same as the standard definition, but the phase values for harmonics 3 and 9 are set to 0. This will sound exactly the same as the waveform shown in 10.5d, even though the shapes of the two waveforms are quite different. shown in figure 10.8. This shows that the imaginary part is now zero, and the real part spectrum now contains the non-zero coefficients. The phase spectrum is also different, while the magnitude spectrum is the same as for the original square wave. This result is highly significant as it shows that only the magnitude spectrum is unaffected by a delay in the signal, and so we can say the magnitude spectrum best captures the essential characteristics of the shape of the square wave. The difference in the two square waves is in their relative position in time - one is slightly later than the other. This agrees with what we expect from human hearing: if we play a note on a piano and then play the same note at some other time, the notes sound exactly the same, even though their position with respect to some reference time t = 0 will generally be different. In fact, from extensive studies, we know that the human ear is largely insensitive to phase information in discerning speech and other sounds, such that if two signals have the same magnitude spectrum but different phase spectra, in general they will be perceived as sounding the same2. This fact not only applies to the square wave as a whole, but to the individual harmonics as well - if we synthesise a waveform with the magnitude spectrum of a square wave, and any phase spectrum, it will sound exactly the same as the normal square wave. As the square wave shape is only created by careful alignment of the sinusoids, it will generally be the case that a wave created with an arbitrary phase spectrum will not form a square waveform at all; nevertheless, any two waveforms with the same magnitude spectra will sound the same regardless of phase or the shape of the waveform. Figure 10.9 shows a waveform created with the magnitude values of a square wave. This is an important fact for speech analysis and synthesis. For analysis, we generally only need to study the magnitude spectra and can ignore phase information. For synthesis, our signal 2 Phase information is however used to localise sounds, that is, we use the phase difference between the signal arriving at each ear to estimate where the source of the signal is
- 302. 278 Chapter 10. Signals and Filters generator need only get the magnitude information correct, the phase is much less important. We will now discuss the frequency range over which the spectrum is defined. First note that when considering frequency plots of continuous signals, it is important to realise that the upper limit in the graph is arbitrarily chosen - in fact all continuous signals have frequency values to ∞. Just because humans can only hear frequencies up to say 20,000Hz doesn’t mean to say they stop being physically present above this level. Secondly note that the spectra for the square wave in Figures 10.7 and 10.8 are plotted with negative as well as positive values for frequency. Our initial definition of frequency stated that it was the inverse of the period, and since we can’t have a negative period, why negative frequency? The use of negative frequency comes from the fact that a true continuous signal, is defined in time from −∞ to ∞. As we shall see later, time and frequency share a duality, with neither being more “real” than the other. As time extends from −∞ to ∞, so does frequency. If we take the cosine sinusoid definition and write it in terms of complex exponentials using the inverse Euler equation 10.11, x(t) = cos(ωt) = ejωt +e− jωt 2 x ′ (t) = sin(ωt) = ejωt +e− jωt 2j we see that this signal actually has two frequencies, a positive ω and a negative ω. This is a general property of all real signals: they will have conjugate symmetry around ω = 0, i.e. the vertical axis in the spectrum. Conjugate symmetry is a complex property, in which the real part of the spectrum is reflected in the vertical axis, and the imaginary part is reflected in the vertical axis and again in the horizontal axis. This is clearly seen in Figures 10.7 and 10.8. Signals which are not purely real, such as ejωt itself, only have one frequency value. Because of this, it can be argued that complex exponentials really are a simpler form that the sine or cosine form. 10.1.6 The Fourier transform In general, we need to analyse non-periodic signals as well as periodic ones. Fourier synthesis and analysis can not deal with non-periodic signals, but a closely related operation, the Fourier Transform is designed to do just this. By definition, the Fourier transform is: X(jω) = Z ∞ −∞ x(t)e− jω dt (10.21) Its similarity to Fourier analysis (Equation 10.18) should be clear: ak = 1/T0 Z T0 0 x(t)e− jkω0t dt
- 303. Section 10.1. Analogue signals 279 -1.5 -1 -0.5 0 0.5 1 1.5 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 amplitude time (a) -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 -1500 -1000 -500 0 500 1000 1500 amplitude frequency /tmp/tmp_V_JV0 using 1:2 (b) -1.5 -1 -0.5 0 0.5 1 1.5 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 amplitude time (c) -0.2 0 0.2 0.4 0.6 0.8 1 -1500 -1000 -500 0 500 1000 1500 amplitude frequency /tmp/tmped7xML using 1:2 (d) -1.5 -1 -0.5 0 0.5 1 1.5 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 amplitude time (e) -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -1500 -1000 -500 0 500 1000 1500 amplitude frequency /tmp/tmpvld6dE using 1:2 (f) Figure 10.10 Real spectra for a signal of constant shape but decreasing period. As the period decreases, the harmonics get closer, but the envelope of the spectrum stays the same. In the limit, period → ∞, the harmonics will be infinitely close and the spectrum will be be a continuous function of frequency.
- 304. 280 Chapter 10. Signals and Filters The differences are that firstly we are now computing an integral from −∞ to ∞ rather than over a single period: as our signal now may not have a period we can not of course integrate over it. Secondly, the result is a continuous function over frequency, not simply the complex amplitudes for a set of discrete harmonics. For non-periodic signals we have to abandon the idea of harmonics: the magnitude and phase are instead expressed as a continuous function of frequency. The Fourier transform can be explained as follows. Imagine we have a rectangular pulse waveform x(t), which has values for a finite duration T, defined as follows: x(t) = ( 1 for −1/2T ≤ t ≤ 1 2T −1 otherwise This is equivalent to just a single period of a square wave, centred at t = 0 with values −1 every- where else. This is clearly not periodic. But imagine we artificially make a periodic version by replicating its shape as shown in Figure 10.10a. This gives us our familiar square wave and by Fourier analysis we get the spectrum, the real part of which is shown in Figure 10.10b. Now imag- ine we perform the same trick, but move the repetitions further apart - in effect we are increasing the period of our waveform (Figure 10.10c). After performing Fourier analysis, we get the real spectrum shown in Figure 10.10d. It should be immediately apparent that the overall shape (the envelope) is the same, but that the harmonics are now much closer together, as we would expect of a waveform of longer period. If we extend the period further still (Figure 10.10e), we get the real spectrum shown in Figure 10.10f and here the harmonics are closer still. As T0 → ∞, the “harmonics” in the spectrum become infinitely dense, resulting in a continuous function. Let us now calculate the continuous spectrum of the rectangular pulse directly. From its definition, the Fourier transform of the waveform is: X(jω) = Z T/2 −T/2 e− jωt dt = e− jωT/2 − jω − −ejωT/2 − jω We can use the inverse Euler formula to simplify this as: X(jω) = sin(ωT/2) ω/2 (10.22) Any function of the form sin(x)/x is known as a sinc function. The real spectrum of this is shown in Figure 10.10b, and we can see that this has the exact shape suggested by the envelope of the harmonics in Figure 10.10. Figures 10.10c and 10.10d show the waveform and Fourier transform for a pulse of shorter duration. The Fourier transform is now more “spread-out”, and this demonstrates an important of the Fourier transform known as the scaling property.
- 305. Section 10.1. Analogue signals 281 -1.5 -1 -0.5 0 0.5 1 1.5 -0.01 -0.008 -0.006 -0.004 -0.002 0 0.0020.0040.0060.0080.01 amplitude time (a) -0.1 -0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 -1500 -1000 -500 0 500 1000 1500 amplitude frequency (b) -1.5 -1 -0.5 0 0.5 1 1.5 -0.01 -0.008 -0.006 -0.004 -0.002 0 0.0020.0040.0060.0080.01 amplitude time (c) -0.1 -0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 -1500 -1000 -500 0 500 1000 1500 amplitude frequency (d) Figure 10.11 Waveform and real part of spectrum for two rectangular pulses of different duration. These figures demonstrate the scaling principle whereby compression in one domain gives rise to spreading in the other. The inverse Fourier transform is: x(t) = 1/2π Z ∞ −∞ X(jω)ejωt dt (10.23) and when we compare this to the Fourier synthesis equation (Equation 10.18) x(t) = ∞ ∑ −∞ akejkω0t
- 306. 282 Chapter 10. Signals and Filters -1.5 -1 -0.5 0 0.5 1 1.5 -2000 -1500 -1000 -500 0 500 1000 1500 2000 amplitude frequency -0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 -0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08 amplitude time Figure 10.12 Figure a) is the real part of a spectrum, and figure b) is its inverse Fourier transform. Along with Figure 10.11, these figures demonstrate the duality principle. we can see that the discrete frequency representation of a fundamental frequency multiplied by a harmonic index has now been replaced with a continuous frequency variable ω. Let us now use the inverse Fourier transform to find the waveform for a defined spectrum. Consider a spectrum which is 1 within a certain frequency region and -1 elsewhere: X(jω) = ( 1 for −ωb ≤ t ≤ ωb −1 otherwise (10.24) The inverse Fourier transform of this is: x(t) = 1/2π Z ωb −ωb ejωt dt = 1 2π e− jωbt − j2πt − −e− jωbt − j2πt = sin(ωbt) πt (10.25) This is shown in figure 10.12. When we compare equation 10.25 to equation 10.22 and figure 10.11 to figure 10.12 we see that the Fourier transform of a rectangular pulse is a sinc spectrum, and the inverse Fourier transform of a rectangular spectrum is a sinc waveform. This demonstrates another special property of the Fourier transform known as the duality principle. These and other general properties of the Fourier transform are discussed further in section 10.3.
- 307. Section 10.2. Digital signals 283 10.2 DIGITAL SIGNALS We now turn our attention to digital signals. All the signals we have discussed up until now are analogue signals in which the amplitude and time dimensions are represented by real numbers. By contrast, Digital signals are represented by a sequence of integers. In the “real-world”, most signals are analogue, including speech signals, where the amplitude is a measure of sound pressure level or some other physical quantity. Digital signals are mostly found in computers. 10.2.1 Digital waveforms A digital signal is represented by a sequence of evenly spaced points, the value at each point being the amplitude, represented by an integer. This integer amplitude value can be interpreted as the digital manifestation of sound pressure level, or voltage or some other such analogue quantity. Formally, a digital sequence is represented as follows: x[n] = ...,x−2,x−1,x0,x1,x2,... (10.26) For digital sequences, it is usual to use n and write x[n], to distinguish from continuous time t in analogue signals. Each point in the sequence is called a sample, and the distance between one sample and the next is called the sample period, Ts. The sample rate or sample frequency is given by Fs = 1/Ts. The sample rate limits the frequencies which the digital signal can describe. The highest frequency that a digital signal can contain is called the Nyquist frequency and is exactly half the sample rate. The choice of sample rate is a compromise between wanting to have the highest possible quality (high sample rate) and the lowest storage or transmission cost (low sample rate). Practical data from human perception studies shows that the limit of human hearing is about 20,000Hz and this figure would lead us to use sampling rates of 40,000 Hz. Indeed, compact discs and digital audio tapes have sampling rates of 44,100Hz and 48,000Hz respectively. However, it can be shown that for speech signals, very little useful speech energy is found in the higher frequencies, and so sample rates of 16,000Hz are often used and found to be perfectly adequate for accurate represen- tation of speech signals. Where space is tight, lower sampling rates are used, most commonly in the telephone which has a standardised sample rate of 8000Hz. While speech at this sample rate is normally quite intelligible, it is clear that it sounds much poorer than when higher sample rates are used. The bit range describes the dynamic range of a digital signal, and is given by the range of the integer used to store a sample. Widely used values are 16 bits, which gives an absolute dynamic range of 216, meaning the amplitude of the sample can range from -32,768 to 32,767. Dynamic range is normally expressed in the logarithmic decibel scale. In digital signals, we take the reference as 1 and therefore the dynamic range of a 16 bit signal is: 10log(216 ) ≈ 96dB (10.27)
- 308. 284 Chapter 10. Signals and Filters To save on space and transmission bandwidth, telephones and other systems frequently use 8 bits (48db) and 12 bits (72db). At 8 bits, speech quality is noticeably worse than at 16 bits. 12 bits is a reasonable compromise, but if space is not an issue (and with most types of speech synthesis it is not), 16 bits is normally used. 10.2.2 Digital representations We use somewhat different notation when dealing with digital signals. First, an integer n is used instead of t for the time axis, where n is related to t by the sample period: n = tTs (10.28) From this, we can relate the continuous signal x(t) and digital signal x[n] by x[n] = x(n/Ts). We use square brackets “[ ]” rather than parentheses (round brackets) “( )” when describing digital functions. Strictly speaking, a signal x[n] is in the n-domain but we shall informally refer to it as being in the time domain. Secondly, it is common to use normalised frequency, rather than natural frequency. From 10.28 we see that the relationship between frequency, angular frequency and normalise angular frequency is ω̂ = ω Ts = 2πF Ts (10.29) Using n and normalised angular frequency ω̂ makes calculations easy as the equations are the same regardless of sample rate. To convert to real frequency, the normalised frequency is simply multiplied by the sample rate. In general, we use the same terms for amplitude and phase for both analogue and digital signals. 10.2.3 The discrete-time Fourier transform For a continuous signal, we can move from the time domain to the frequency domain via the Fourier transform, but as we are now dealing with digital signals, this formula cannot be used di- rectly. We will now derive an equivalent transform for digital signals. Starting with the previously defined Fourier transform (Equation 10.21) X(jω) = Z ∞ −∞ x(t)e− jωt dt we can use this on a digital sequence with the relation x[n] = x(n/Ts). The Fourier transform is calculated as a integral over a range, which can be interpreted as finding the area under the curve over this range. The numerical method of the Riemann sum says that we can approximate a continuous area function by adding up the areas of separate rectangles. The area of each rectangle is the value of the function at that point multiplied by the width of the rectangle. In our case,
- 309. Section 10.2. Digital signals 285 the width of the rectangle is the sample period Ts. This gives a numerical version of the Fourier transform: X(jω) = ∞ ∑ −∞ (x(nTs)e− jωt )Ts Ts is independent of the integral, we move it outside, and use x[n] in place of x(nTs) giving: X(jω) = Ts ∞ ∑ −∞ x[n]e− jωnTs We can get rid of the Ts by making use of normalised frequency, ω̂ = ω/Ts. The result is the discrete-time Fourier transform or DTFT: X(ejω̂ ) = ∞ ∑ −∞ x[n]e− jω̂n (10.30) It should be clear that the DTFT is a transform in which time is discrete but frequency is contin- uous. As we shall see in later sections, the DTFT is a very important transform in its own right but for computational purposes, it has two drawbacks. Firstly, the infinite sum is impractical, and secondly we wish to obtain a representation for frequency that is useful computationally, and such a representation must be discrete. 10.2.4 The discrete Fourier transform For the first problem, we choose a sufficiently large interval L such that the samples outside 0 n L are zero. For the second problem, we will use a similar operation to time domain sampling and “sample” the spectrum at discrete intervals. Just as n = t.Ts, we define the spacing between each discrete frequency as Fs/N = 1/(NTs) where N is the total number of frequency values we want. Each angular frequency therefore lies at ωk = 2π NTs giving the transform: X(ejω̂k ) = L ∑ n=0 x[n]e− j ˆ ωkn If we set L = N (discussed below), and use the above definition of ωk, we can write 10.2.4 in a more compact usable form: X[k] = N−1 ∑ n=0 x[n]e− j 2π N kn (10.31)
- 310. 286 Chapter 10. Signals and Filters which gives us the standard definition of the Discrete Fourier Transform or DFT. The interpretation is the same as for the Fourier transform defined in section 10.1.6; For each required frequency value, we multiply the signal by a complex exponential waveform of that frequency, and sum the result over the time period. The result is a complex number which describes the magnitude and phase at that frequency. The process is repeated until the magnitude and phase is found for every frequency. The DFT states that the length of the analysed waveform and the number of resultant fre- quency values are given by N. This makes intuitive sense: if we only analyse a few samples, we can only expect a somewhat crude version of the spectrum; by analysing more samples, we have more information “to go on” and can fill in more detail in the spectrum: that is, the points in the spectrum (ωk) get closer and closer. As N → ∞, we analyse all the waveform values, and the dis- tance between the frequency points in the spectrum → 0. This is exactly the case that the DTFT describes. Both the DTFT and DFT have inverse transforms: x[n] = 1 2π Z π −π X(ejω̂ ).ejω̂n dω̂ (10.32) x[n] = 1 N N−1 ∑ n=0 x[n]ej 2π N kn (10.33) In practice an algorithm called the Fast Fourier Transform or FFT is commonly used to calculate the DFT. This is not another type of transform, merely a fast way of calculating a DFT: the result for the FFT is exactly the same as for the DFT. The FFT only works for values of N which are powers of 2, hence it is common to find N set to 256, 512. If a DFT of a different N is required, a time signal of the next highest power of 2 is chosen, and the extra values are set to 0. So for example, if we have a signal of length 300, we set N = 512, “pad” the extra 212 values with 0, and the calculate a 512 point FFT. The use of the FFT is pervasive in practice, and one often hears references to “the FFT of the signal” etc. We repeat, in terms of the outcome, the FFT and DFT are exactly the same and hence taking “the FFT of the signal” is no different from “taking the DFT”. An overview of the FFT can be found in Oppenheim and Schafer [342]. 10.2.5 The z-Transform We now introduce another important transform, the z-transform 3. The z-transform is perhaps the most useful of all transforms in practical digital signal processing. It is defined as: X(z) = ∞ ∑ n=−∞ x[n]z−n (10.34) 3 what? you can’t be serious - another transform!
- 311. Section 10.2. Digital signals 287 This should look familiar - if we use z = ejω̂, we get: X(ejω̂ ) = ∞ ∑ −∞ x[n]e− jω̂n which is the discrete-time Fourier transform, Equation 10.30. Just like the DTFT, the z-transform is a transform which operates on a digital signal but which produces a continuous function. An important property of the z-transform is its affect on a delay. If we define a signal y[n] = x[n−nd] that is, a delayed version of x[n]. By use of a dummy variable m = n − nd, we can find its z- transform as follows Y(z) = ∞ ∑ n=−∞ y[n]z−n = ∞ ∑ n=−∞ x[n−nd]z−n = ∞ ∑ m=−∞ x[m]znd −m = ∞ ∑ m=−∞ x[m]z−m znd The summation in m is just the z-transform of x defined in m rather, than n and so we can write: Y(z) = X(z)znd (10.35) Hence delaying a signal by nd multiplies its z-transform by znd . This important property of the z-transform and is discussed further in section 10.3.5. Why use yet another transform? The reason is that by setting z to ejω̂ we greatly simplify the DTFT, and this simplification allows us to perform sophisticated operations in analysis and synthesis with relatively simple mathematical operations. The key to the ease of use of the z- transform, is that equation 10.34, can be re-written as: X(z) = ∞ ∑ n=−∞ x[n](z−1 )n (10.36) That is, the transform can be written as a polynomial in z−1, so that the techniques of polynomial solving can be brought to bear. A final transform, the Laplace Transform, is frequently used in continuous time signal processing, and is the continuous time version of the z-transform. The z-transform and the Laplace transform can be thought of as generalisations of the DTFT and Fourier transform respectively. The Laplace transform is not required in this book, as all our signal processing is performed on digital signals. We mention it here for completeness only.
- 312. 288 Chapter 10. Signals and Filters 10.2.6 The frequency domain for digital signals For continuous signals, we showed that the frequency domain extends to from −∞ to ∞ . However, for digital signals, we stated that a frequency of half the sampling rate, the Nyquist frequency, was the highest possible frequency that could be represented by a digital signal. What then is the range of the frequency domain? As for continuous signals the range extends from −∞ to ∞, but no new information is found outside the Nyquist range. Rather, the spectrum repeats itself at multiples of the Nyquist frequency. In other words, the spectrum of a digital signal is periodic. It is for this reason that the spectrum of digital signals, calculated from the DTFT is described as X(ejωt), rather that X( jω) as is the normal for the Fourier transform. The “period” of the spectrum is twice the Nyquist frequency, as the spectrum is defined from −FN to +FN. In terms of normalised frequency, the frequency range is just −π to +π and so the “period” is 2π. Just as with real continuous signals, real digital signals exhibit conjugate symmetry in the frequency domain. Hence in a 512 point DFT the first 256 values will be conjugate symmetric with the second 256. For this reason only half a digital spectrum is normally plotted, and this will have half the number of values used to calculated the DFT. 10.3 PROPERTIES OF TRANSFORMS At first sight the number of transforms can seem unnecessarily large. (recall that each has an inverse transform also.) Why do we need so many? The differences are simply attributable to us wanting to use both continuous and discrete representations for time and frequency, and to make analysis easier (in the case of the z-transform and Laplace transform). The differences between them can be clearly shown in the following table: Discrete Time Continuous time Discrete Frequency Discrete Fourier Transform (DFT) Fourier analysis, periodic waveform Continuous Frequency Discrete Fourier Transform (DTFT) Fourier Transform Continuous variable z-transform Laplace Transform We now describe various general properties of the transforms we have introduced. For demonstration purposes, we will describe these using the Fourier transform, but in general the following hold for all our transforms. 10.3.1 Linearity All the transforms are linear. If we have an input x(t) = αs1(t)+βs2(t), then X(jω) = αX1(jω)+ βX2(jω). The proof of this for the Fourier transform is:
- 313. Section 10.3. Properties of Transforms 289 X(jω) = Z ∞ −∞ (αx1(t)+βx2(t))e− jωt = α Z ∞ −∞ x1(t)e− jωt +β Z ∞ −∞ x2(t)e− jωt = αX1(jω)+βX2(jω) (10.37) 10.3.2 Time and Frequency Duality The Fourier transform and inverse Fourier transform have very similar forms. If we have a function g(.) and define a signal as g(t) then its Fourier transform will be G(jω) = F {g(t)} (here we use the notation F to denote the operation of the Fourier Transform). If we define the spectrum in terms of our function as g(ω) and take the inverse Fourier transform, we get G(t) = 2πF {g(ω)}. It should be clear that the form is the same apart from the 2π term, which is required because of our use of angular frequency. g(t) F − → G(w) (10.38) g(ω) I F ← − − G(t) (10.39) Figure 10.12 demonstrates this property. 10.3.3 Scaling We have already described the Fourier transform of a pulse of different durations in section 10.1.6. The shorter duration pulse has a spectrum that is more spread out, and the longer duration pulse has a more contracted spectrum. This is known as the scaling property, which states that: com- pressing a signal will stretch its Fourier transform and vice versa. This can be formally stated as: x(at) F − → 1 |a| X(j(w/a)) (10.40) 10.3.4 Impulse Properties From the scaling property, we know that as the duration of the pulse approaches zero, the spectrum spreads out more and more. In the limit, we reach the unit impulse, which has zero width. This is a particularly important signal. For digital signals, this is simply defined as a signal, δ[n], which has a value 1 at time 0 and a value 0 elsewhere. For continuous-time signals, the definition is a little trickier, but is described as a signal which has values of zero everywhere apart at time 0 (i.e. zero width), and at that time it is described by an integral which defines its area (not height) as 1: δ(t) = 0 for t 6= 0 (10.41) Z ∞ −∞ δ(t)dt = 1 (10.42)
- 314. 290 Chapter 10. Signals and Filters As expected from the scaling property, the Fourier transform of an impulse is a function that is “infinitely stretched”, that is, the Fourier Transform is 1 at all frequencies. Using the duality principle, a signal x(t) = 1 for all t will have a Fourier transform of δ(ω), that is, an impulse at time ω = 0. This is to be expected - a constant signal (a d.c. signal in electrical terms) has no variation and hence no information at frequencies other than 0. δ(t) F − → 1 (10.43) 1 F − → δ(ω) (10.44) 10.3.5 Time delay A signal x(t) that is delayed by time td is expressed as x(t −td). If we denote the Fourier transform of x(t) as X(jω) and the Fourier transform of x(t −td) as Xd(jω), we see that from the Fourier transform definition that Xd(jω) is: Xd(jω) = Z ∞ −∞ x(t −td)e− jωt dt (10.45) To simplify the integration, we use a dummy variable v and make the substitution v = t −td: Xd(jω) = Z ∞ −∞ x(v)e− jω(v+td ) dv = e− jωtd Z ∞ −∞ x(v)e− jωv dv Xd(jω) = e− jωtd X( jω) So delaying a signal by td seconds is equivalent to multiplying its Fourier transform by e− jωtd . If we apply this to a unit impulse, x(t −td) = δ(t −td), we get: Xd(jω) = Z ∞ −∞ δ(t −td)e− jωt dt = e− jωtd Z ∞ −∞ δ(v)e− jωv dv and from the impulse property, we know the Fourier transform of of δ(t) is 1, leaving: Xd(jω) = e− jωtd (10.46) At first sight, this may seem strange as the Fourier Transform for a shifted impulse seems very different for a normal impulse which simply had a Fourier Transform of 1. Recall however, that the magnitude of e− jωtd , will be 1, and so the magnitude spectrum will be the same as the delta function. The phase of e− jωtd is simply a linear function of td - this is as we should expect, the longer the delay td the more the phase spectrum will be shifted. It should be noted that the above result is the Fourier transform result of the z-transform delay derived in equation 10.35.
- 315. Section 10.3. Properties of Transforms 291 10.3.6 Frequency shift From the duality principle, it then follows that if we have a signal x(t) = ejω0t (i.e. a sinusoid), the Fourier transform will be 2πδ(ω − ω0), that is a single impulse at the frequency ω0. This is entirely to be expected - the frequency domain impulse can be interpreted as a harmonic and is consistent with the Fourier analysis result. More generally, if we multiply a signal by e jω0t we cause a frequency shift of ω0 in the Fourier transform: x(t)ejω0t ← X(j(ω−ω0)) (10.47) Shift in the frequency domain corresponds to modulation in the time domain. 10.3.7 Convolution Convolution is defined as the overlap between two functions when one is passed over the other, and is given by: f(t) = g(t)⊗h(t) = Z h(τ)g(t −τ)dτ (10.48) If we take the Fourier transform of this, we get: F(jω) = F nZ ∞ τ=−∞ h(τ)g(t −τ)dτ o = Z ∞ t=−∞ Z ∞ τ=−∞ h(τ)g(t −τ) e− jωt dt = Z ∞ τ=−∞ h(τ)dτ Z ∞ t=−∞ g(t −τ)e− jωt dt (10.49) from the time-delay property (equation 10.3.5) we know that: Z ∞ τ=−∞ g(t −τ)e− jωt = e− jωτ G(jω) and substituting this into 10.49 gives F(jω) = Z ∞ τ=−∞ h(τ)e− jωτ G(jω) The R h(τ)e− jωτ term is simply the Fourier transform of h(t) defined in terms of τ rather than t, so we can state: F(ω) = G(ω)H(ω) (10.50) Hence, convolution in the time domain corresponds to multiplication in the frequency do- main. It can also be proved that multiplication in the time domain corresponds to convolution in the frequency domain.
- 316. 292 Chapter 10. Signals and Filters 10.3.8 Analytical and Numerical Analysis The main practical difference between the transforms is whether an analytical or numerical method is used in the calculation. For transforms and inverse transforms that contain integrals, it should be clear that we must be able to find the analytical form of integral. This can be a blessing in that we can use analytical techniques to find and succinctly describe the transforms of some signals, but this can also be a difficulty, because first we have to be able to represent the signal as an analytic function, and second we have to be able to find its integral. When dealing with arbitrary signals, say from a recording, it is nearly certain that we will not know the form of the signal, and so a numerical method must be used. 10.3.9 Stochastic Signals All the transforms we have examined either integrate or sum over an infinite sequence. This will only give a meaningful result if the sum of the signal over that range is finite; and it can easily be shown that this is the case for all periodic and many non-periodic signals. There are however signals for which this is not so, and in such cases we say that the Fourier (or other transform) does not exist for these signals. One important class of such signals are stochastic signals, which are signals generated by some random process. The “noisy” signals (for example found in fricatives) that we have been mentioning up until now are such signals. In trying to characterise these signals we face two problems. Firstly, we cannot exactly describe them in the time domain as the very nature of the random process means that there is no strict pattern in the samples. Secondly, as the transforms we have introduced can not be calculated, we can not express these signals directly in the frequency domain either. We get round these issues by focusing on averages within the signal, and specifically we use a measure of the self similarity of a signal called its autocorrelation function. For a signal x[n], this is defined as R(j) = ∞ ∑ n=−∞ x[n]x[n− j] (10.51) which is the expected value of the product of a stochastic signal with a time shifted version of itself. This function does have a Fourier transform, and this gives the power spectral density of the signal. 10.4 DIGITAL FILTERS A filter is a mathematically defined system for modifying signals. In signal processing filters are put to many uses; as their name might suggest they can “filter out” some unwanted portion of a signal. In speech synthesis, we are interested in them mainly because of the acoustic theory of speech production, (section 11.1), which states that speech production is a process by which a
- 317. Section 10.4. Digital Filters 293 glottal source is modified by a vocal tract filter. This section explains the fundamentals of filters and gives insight into how they can be used as a basis for modelling the vocal tract. Digital filters are of course filters which operate on digital signals. Here, we describe a specific but powerful kind of filter, called the linear time invariant (LTI) filter. Using the notation x[n] 7→ y[n] to describe the operation of the filter (“7→” is read “maps to”), we state the time invariance property as: x[n−n0] 7→ y[n−n0] for any n0 (10.52) which states that the system behaviour is not dependent on time. A linear system has the property of scaling, such that if x[n] 7→ y[n] then αx[n] 7→ αy[n] and superposition, such that if x1[n] 7→ y1[n] and x2[n] 7→ y2[n] then: x[n] = x1[n]+x2[n] 7→ y[n] = y1[n]+y2[n] These combine to form the linearity principle: x[n] = αx1[n]+βx2[n] 7→ y[n] = αy1[n]+βy2[n] (10.53) The linearity property combines with the principle of Fourier synthesis - as the filter linearly combines the components in a waveform, we can construct and analyse a filter for components separately and not have to worry about their interactions. 10.4.1 Difference Equations Digital LTI Filters are often described in terms of their time domain difference equation, which relates input samples to output samples. The finite impulse response or FIR filter has the differ- ence equation: y[n] = M ∑ k=0 bkx[n−k] (10.54) The infinite impulse response or IIR filter has the difference equation: y[n] = M ∑ k=0 bkx[n−k]+ N ∑ l=1 aly[n−l] (10.55) As well as operating on the input, the IIR filter operates on previous values of the output and for this reason these filters are sometimes referred to as feedback systems or recursive filters. The
- 318. 294 Chapter 10. Signals and Filters -2 -1 0 1 2 0 2 4 6 8 10 amplitude time /tmp/tmpos4p_z using 0:1 (a) Input signal -6 -4 -2 0 2 4 6 0 2 4 6 8 10 amplitude time /tmp/tmpwO0Pqx using 0:1 (b) Output signal Figure 10.13 Operation of an FIR filter on an example, finite waveform FIR is a special case of an IIR filter, formed when the coefficients {al} are set to zero. (Note that the {al} coefficients do not have a value for 0 as the output for y[n] must be calculated before it can be used on the right hand side of the equation.) Consider a simple FIR filter with M = 3 and bk = {3,−1,2,1}. As M = 3, this is known as a third order filter. From the FIR definition equation, we can see that this will modify any input sequence x as follows: y[n] = 3x[n]−x[n−1]+2x[n−2]+x[n−3] (10.56) Figure 10.13 shows an example signal x = {0,1,2,1,0,−1,−2,−1,0,0} and the output of the filter for this example signal 10.4.2 The impulse response While the above clearly showed how our filter created an output signal y[n] from our given in- put signal x[n], it isn’t immediately clear how the filter will modify other, different input signals. Rather than having to examine the output with respect to a wide range of different input signals, it would be useful to have a characterisation of the effect of the filter that is independent of the specific form of the input. Such a characterisation is called the response of the filter, and a particu- larly useful response is provided by using a unit impulse as input. The output is then the impulse response, denoted h[n]. A unit impulse, δ, is a special function which has value 1 at n = 0, and 0 elsewhere: δ[n] = n 1 n = 0 0 n 6= 0 (10.57)
- 319. Section 10.4. Digital Filters 295 -6 -4 -2 0 2 4 6 0 2 4 6 8 10 amplitude time /tmp/tmpszD3oV using 0:1 (a) FIR filter 0 0.5 1 1.5 2 0 5 10 15 20 25 30 amplitude time /tmp/tmpHwWEdb using 0:1 (b) IIR filter Figure 10.14 Example impulse responses for FIR and IIR filters. Note while the number of coef- ficients in the IIR filter is actually less than for the FIR filter, the response will continue indefinitely (only values up to 30 samples are shown) The output of the example FIR filter for an impulse is given in figure 10.14a, which shows that the impulse response is simply the coefficients in the filter. Because of this, the duration of the impulse will always be the order of the filter. As the order is finite, the impulse response is finite which gives rise to the name of the filter. Consider now an example first order IIR filter, with a1 = −0.8, and b[0] = 1. Its difference equation is therefore: y[n] = x[n]−0.8y[n−1] (10.58) If we give this filter the unit impulse as an input, then we can calculate the output for any n after 0. The output is shown in Figure 10.14b. First of all we observe that the output is not of a fixed duration: in fact it is infinite as the name of the filter implies. Secondly we observe that the response takes the form of a decaying exponential. The rate of decay is governed by the filter coefficient, a1. (Note that if a1 1 the output will grow exponentially: this is general undesirable and such filters are called unstable.) Consider the impulse response of the second order IIR filter a1 = −1.8,a2 = 1, shown in figure 10.15a. Just as in the first order case, this filter has an infinite response, but this time takes the form of a sinusoid. This shows the power of the IIR filter - with only a few terms it can produce a quite complicated response. The filter a1 = −1.78,a2 = 0.9, has its impulse response shown in Figure 10.15b. We can see that it is a decaying sinusoid; we have in effect combined the characteristics of examples 10.14b and 10.15a. The key point to note is that in all cases, the input is the same, ie a unit impulse. The characteristics of the output are governed solely by the filter - after all the input has no periodicity or decay factor built in. If we choose a slightly different set
- 320. 296 Chapter 10. Signals and Filters -6 -4 -2 0 2 4 6 0 10 20 30 40 50 60 70 80 90 100 amplitude time -4 -3 -2 -1 0 1 2 3 4 0 10 20 30 40 50 60 70 80 90 100 amplitude time Figure 10.15 Impulse response for 2 IIR filters, showing a sinusoid and damped sinusoid response. of coefficients, we can create an output with a different period and decay factor. Different filters can produce a wide range of impulse responses; the decaying sinusoid pat- terns here only come from a specific type of second order IIR filter. In general though, the more terms in the filter, the more complicated the impulse response will be. The impulse response is important for a number of reasons. Firstly, it provides us with a simple way of examining the characteristics of the filter. Because the input only lasts for a single sample, the output is not “clut- tered” with the further interference from subsequent input samples. Impressionistically, exciting the filter with an impulse response can be thought of as “giving the filter a nudge and seeing what happens”. Secondly, and more importantly, we can use the principle of linearity and time invari- ance to describe the responses of more general signals by using the impulse response, explained next. 10.4.3 Filter convolution sum The response for a non-unit impulse is simply the unit impulse response scaled. So using the linearity principle, we can state that: 3δ[n] 7→ 3h[n] As the unit impulse is simply a single sample, a full input sequence x[n], can be characterised by a sequence of scaled unit impulses: x[n] = ∑w[k]δ[n−k] What are the values of w? As the delta function will set everything outside n−k to 0, it should be clear that w[n] = x[n] for a given n.Therefore, we can state: x[n] = K ∑ k=0 x[k]δ[n−k] (10.59)
- 321. Section 10.4. Digital Filters 297 This may seem a rather trivial result, but it shows that a signal can be fully described by a set of appropriately scaled impulse responses. The time invariance property of LTI systems tells us that a shifted input will give a shifted response. For the impulse function, the following is therefore true: δ[n] 7→ h[n] δ[n−1] 7→ h[n−1] δ[n−2] 7→ h[n−2] δ[n−k] 7→ h[n−k] for any integer k The linear property states that αδ[n] 7→ αh[n] for any α and so the following is also true: x[0]δ[n] 7→ x[0]h[n] x[1]δ[n−1] 7→ x[1]h[n−1] x[2]δ[n−2] 7→ x[2]h[n−2] x[k]δ[n−k] 7→ x[k]h[n−k] for any integer k We can use these two properties on equation 10.59 to find the output of the filter ∑ k x[k]δ[n−k] 7→ ∑ k x[k]h[n−k] which gives us: y[n] = ∑ k x[k]h[n−k] (10.60) which is known as the convolution sum. It shows that we can calculate the output for a filter for any input, provided we know the impulse response h[n]. 10.4.4 Filter transfer function We will now examine the frequency domain description or frequency response of the LTI filter by means of the z-transform. The IIR difference equation (10.55), expanded to a few terms, is: y[n] = b0x[n]+b1x[n−1]+...+bMx[n−M]+a1y[n−1]+...+aNy[n−N] By using the notation Z{.} to denote the z-transform, we can take the z-transform of both sides as follows: Y(z) = Z{b0x[n]+b1x[n−1]+...+bMx[n−M]+a1y[n−1]+...+aNy[n−N]} The coefficients are not affected by the z-transform, so using the linearity of the z-transform, this simplifies to: Y(z) = b0Z{x[n]}+b1Z{x[n−1]}+...+bn−MZ{x[n−M]}+a1Z{y[n−1]}+...+aNZ{y[n−N]}}
- 322. 298 Chapter 10. Signals and Filters The z-transform time delay property (equations 10.35 and 10.46) states that Z{x[n−k]} = X(z)z−k and so our equation becomes: Y(z) = b0X(z)+b1X(z)z−1 +...+bn−MX(z)zM +a1Y(z)z−1 +...+aNY(z)zN If we group the Y(z) terms on one side, and factor the common Y[z] and X[z], we get Y(z) = b0 +b1z−1 +...+bMzM 1−a1z−1 −...−aNzN X(z) The term is known as the transfer function of the filter H(z) = Y(z) X(z) = b0 +b1z−1 +...+bMzM 1−a1z−1 −...−aNzN (10.61) = M ∑ k=0 bk 1− L ∑ l=1 al (10.62) and fully defines the characteristics of the filter in the z-domain. From this, the frequency domain characteristics of the filter can easily be found (explained below). 10.4.5 The transfer function and the impulse response We have previously shown that the impulse response, which completely characterises the filter, allows us to calculate the time domain output for any input by means of the convolution sum: y[n] = ∞ ∑ k=−∞ h[k]x[n−k] (10.63) Section 10.3 showed how to derive the Fourier transform for a convolution and we can apply this to a digital signal by using the z-transform in the same way. Y(z) = Z n ∞ ∑ k=−∞ h[k]x[n−k] o = ∞ ∑ k=−∞ h[k] ∞ ∑ n=−∞ x[n−k]z−n = ∞ ∑ k=−∞ h[k]z−k X(z) (10.64) The ∑h[k]z−k term is simply the z-transform of h[n] defined in terms of k rather than n, so 10.64 simplifies to Y(z) = H(z)X(z) (10.65)
- 323. Section 10.5. Digital filter analysis and design 299 which is the same result as equation 10.62. This shows that the transfer function H(z) is the z-transform of the impulse response h[n]. This gives us the following two important properties: 1. a LTI filter is completely characterised by its impulse response, h[n] in the time domain. The z-transform of the impulse response is the transfer function, H(z), which completely characterises the filter in the z-domain. 2. The transfer function is a ratio of two polynomials, with the same coefficients as used in the difference equation. 10.5 DIGITAL FILTER ANALYSIS AND DESIGN Section 10.4 fully described the relationship between the time domain difference equations and the z-domain transfer function of a LTI filter. For first order filters, the coefficients are directly interpretable, for example as the rate of decay in an exponential. For higher order filters this becomes more difficult, and while the coefficients {al} and {bk} fully describe the filter, they are somewhat hard to interpret (for example, it was not obvious how the coefficients produced the waveforms in Figure 10.15). We can however use polynomial analysis to produce a more easily interpretable form of the transfer function. 10.5.1 Polynomial analysis: poles and zeros Consider a simple quadratic equation defined in terms of a real variable x: f(x) = 2x2 −6x+1 This equation can be factorised into f(x) = G(x−q1)(x−q2) and then the roots q1,q2 of the quadratic can be found by setting f(x) to 0 and solving for x (in this case the roots are x = 0.38 and x = 2.62). G is the gain, which simply scales the function. We can plot the polynomial as shown in Figure 10.16a. The roots of this expression (the points at which f(x) = 0) are marked on the graph. Although the roots only mark two points, they do in fact describe the general curve. Figure 10.16b shows the effect of varying the gain - the curve can become steeper or shallower, or even inverted if the gain is set to a negative number, but in all cases the general shape of the curve is the same. Because setting x to one of the roots results in f(x) = 0, we call these roots zeros. Now consider the same polynomial used as a denominator expression, where by convention we keep the gain as a numerator term: f(x) = 1 2x2 −6x+1 = G ′ (x− p1)(x− p2)
- 324. 300 Chapter 10. Signals and Filters -5 0 5 10 15 20 -2 -1 0 1 2 3 4 5 amplitude time (a) plot of 2x2 −6x+1 -20 -15 -10 -5 0 5 10 15 20 -2 -1 0 1 2 3 4 5 amplitude time (b) plot of g × (2x2 − 6x + 1) for different values of g Figure 10.16 Plot of polynomial. The points where the curve intersects the x-axis are called “zeros” -2 -1 0 1 2 3 4 5 -2 -1 0 1 2 3 4 5 amplitude time (a) plot of 1/(2x2 −6x+1) -2 -1 0 1 2 3 4 5 -2 -1 0 1 2 3 4 5 amplitude time (b) plot of |1/(2x2 −6x+1)| Figure 10.17 Plot of function with a polynomial as the denominator. The points where the function → ∞ are called “poles”. The plot of this is shown in Figure 10.17a. The shape of this curve is very different, but is again completely characterised by the roots plus a gain factor. This time values at the roots cause the function to “blow up”, x → p1, f(x) → ±∞. Such roots are called poles as they create a “pole-like” effect in the absolute (magnitude) version of the function, shown in 10.17b. We will now show how polynomial analysis can be applied to the transfer function. A polynomial defined in terms the complex variable z takes on just the same form as when defined
- 325. Section 10.5. Digital filter analysis and design 301 -1.5 -1 -0.5 0 0.5 1 1.5 imaginary -1.5 -1 -0.5 0 0.5 1 1.5 real 0 2 4 6 8 10 12 14 |H(z)| (a) z-domain, filter A -1.5 -1 -0.5 0 0.5 1 1.5 imaginary -1.5 -1 -0.5 0 0.5 1 1.5 real 0 2 4 6 8 10 12 14 |H(z)| (b) z-domain, filter B -1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1 imaginary real /tmp/tmpdfuJQM using 1:2 /tmp/tmpDsB2Ku using 1:2 (c) pole-zero plot, filter A -1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1 imaginary real /tmp/tmpdLQl-H using 1:2 /tmp/tmpxcpciY using 1:2 (d) pole zero plot, filter B Figure 10.18 z-domain and pole-zero plots for two filters, A and B. in terms of x. The z form is actually less misleading, because in general the roots will be complex (e.g. f(z) = z2 +z−0.5 has roots 0.5+0.5j and 0.5−0.5j.). The transfer function is defined in terms of negative powers of z - we can convert a normal polynomial to one in negative powers by multiplying by z−N. So a 2nd order polynomial is: H(z) = 1 z2 −a1z−a2 = z−2 1−a1z−1 −a2z−2 = G z−2 (1− p1z−1)(1− p2z−1) Figure 10.18a shows a 3-dimensional plot of the above transfer function with a1 = 1.0,a2 = −0.5. The numerator has two roots at z = 0, which form zeros. The roots of the denominator are
- 326. 302 Chapter 10. Signals and Filters complex (0.5+0.5j and 0.5−0.5j) and it can be seen that the poles do not now lie on the x-axis (which would only happen if they were completely real). As the poles and zeros completely characterise the filter (apart from the gain), we can dis- pense with the need to draw 3-dimensional graphs and simply plot where the poles and zeros lie on the z-plane. This is shown in Figure 10.18c. Figures 10.18b and 10.18d show the 3-d plot and pole-zero plot for a different 2nd order filter. This analysis can be extended to any digital LTI filter. As the general form transfer function, H(z) is defined as the ratio of two polynomials H(z) = b0 +b1z−1 +...+bMz−M 1−a1z−1 −...−aNz−N this can always be expressed in terms of its factors: H(z) = g (1−q1z−1)(1−q2z−1)...(1−qMz−1) (1− p1z−1)(1− p2z−1)...(1− pNz−1) from which pole-zero analysis can be performed. Furthermore, a factored filter can be split into smaller filters: H(z) = g 1−q1z−1 1− p1z−1 | {z } . 1−q2z−1 1− p2z−1 | {z } ... 1−qMz−1 1− pMz−1 | {z } = H1(z)H2(z)...HN(z) (10.66) which is often a useful step when constructing or analysing complicated filter. 10.5.2 Frequency Interpretation of z-domain transfer function Recall that we showed the relationship between the z-transform X(z) = ∞ ∑ n=−∞ x[n]z−n and the discrete-time Fourier transform by setting z = ejω̂: X(ejω̂ ) = ∞ ∑ −∞ x[n]e− jωn Given any z-domain expression, we can generate the equivalent frequency domain expression by simply replacing z with ejω. In section 10.2.6, we explained that the spectrum X(ejω̂) of a digital signal is periodic with period 2π for normalised frequency. This can be thought of as a function which sweeps out a circle of radius 1 in the z-domain. The frequency domain is in effect the particular ‘slice” on the z-domain that this circle describes.
- 327. Section 10.5. Digital filter analysis and design 303 -1.5 -1 -0.5 0 0.5 1 1.5 imaginary -1.5 -1 -0.5 0 0.5 1 1.5 real -1 0 1 2 3 4 |H(z)| (a) z-domain 0 2 4 6 8 10 -4 -3 -2 -1 0 1 2 3 4 amplitude normalise frequency (b) frequency domain Figure 10.19 Relationship between the z-domain and the frequency domain. The unit circle drawn on the z-domain diagram can be unfolded and straightened to produce the frequency response shown in figure (b). [ MORE: circle has yet to be drawn] Let us return to the 3-dimensional representation of the magnitude of H(z), for an example 2nd order filter, as shown in Figure 10.19a. This is as before, but the unit circle is now superim- posed. The value at the unit circle rises and falls with the shape of the plot, and when we “unfold” this, we get a 2-dimensional graph of magnitude against frequency as shown in Figure 10.19b. The phase spectrum can be generated in the same way, that is, by taking the values from the unit circle
- 328. 304 Chapter 10. Signals and Filters from the z-domain phase function. As previously mentioned, we can dispense with 3-d graphs because the pole-zero diagram fully characterises the transfer function. The frequency domain can easily be obtained from the z-domain by use of the poles and zeros. 10.5.3 Filter characteristics We will now use the above results on pole zero analysis and frequency domain interpretation to examine the properties of first order and second order filters, and show how these can be extended to the general case of all digital LTI filters. Consider again a first order IIR filter: h[n] = b0x[n]−a1y[n−1] H(z) = b1 1−a1z−1 H(ejω ) = b1 1−a1e− jω Figure 10.20, shows the time domain, z-domain pole zero plot and frequency domain magnitude spectrum for this filter with b0 = 1 and a1 varying over 0.8,0.7,0.6,0.4. Figure 10.20d shows a 3-d z-domain magnitude plot for a single value, a1 = 0.7. It is clear that the value of the coefficient a1 completely describes the shape of the three functions. For a single term polynomial, the root of 1−a1z−1 is simply a1. As the filter coefficient is real, so will the root, and hence the pole will always line on the real axis in the pole-zero plot. This type of filter is commonly called a resonator because it accurately models the types of resonances commonly found in nature. The peak in the the frequency response is known as a resonance because signals at or near that frequency will be amplified by the filter. For our purposes, it provides a good model of a single vocal tract formant. Resonances are normally described by three properties: amplitude - the height of the peak, frequency - where the peak lies in the spectrum, and bandwidth - a measure of the width or sharpness of the peak. The radius of the pole, r controls both the amplitude and the bandwidth of the resonance. As the value of a1 decreases, its pole moves towards along the real axis towards the origin, and hence the frequency response has a lower amplitude at ω̂ = 0. From the difference equations, we know that a first order IIR filter with a1 1 is simply a decaying exponential. Small values of a1 correspond to slow decay and narrow bandwidth, large values (still 1) a steeper decay and wider bandwidth. This is exactly what we would expect from the Fourier transform scaling property The standard measure of bandwidth is the 3db down point which is the width of the reso- nance at 1/ √ 2 down from the peak. For larger values of r = |a1|, two useful approximations of the bandwidth (in normalised frequency) are: B̂ ≈ −2ln(r) B̂ ≈ 2(1−r) √ r 0.5 r 1.0 (10.67)
- 329. Section 10.5. Digital filter analysis and design 305 -1.5 -1 -0.5 0 0.5 1 1.5 0 5 10 15 20 25 30 amplitude time (a) time domain impulse response 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 -4 -3 -2 -1 0 1 2 3 4 amplitude normalise frequency (b) magnitude spectrum of frequency response -1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1 amplitude time /tmp/tmpkwRBSH using 1:2 /tmp/tmpECGNt0 using 1:2 /tmp/tmpfrEsNU using 1:2 /tmp/tmp90NYI2 using 1:2 /tmp/tmpLR2LJB using 1:2 (c) pole zero plot -1.5 -1 -0.5 0 0.5 1 1.5 imaginary -1.5 -1 -0.5 0 0.5 1 1.5 real 0 5 10 15 20 25 30 |H(z)| (d) magnitude z-domain transfer function Figure 10.20 Plots of first order IIR filter, with a=0,8, 0,7, 0,6 and 0.4. As the length of decay increases, the frequency response becomes sharper. Because only a single coefficient is used, there will be one pole, which will always lie on the real-axis. As a1 → 1, the impulse response will have with no decay and the pole will lie 1.0. Because the pole lies on the unit circle, it will lie in the frequency response, and hence there will be an infinite value for frequency at this point in the spectrum. The first order filter creates a resonance pattern that already has a similar shape to a speech formant; the main difference of course is that its resonant frequency is at zero on the frequency axis. The resonance can easily be shifted from ω̂ = 0 by moving its pole off the real-axis, and this
- 330. 306 Chapter 10. Signals and Filters 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 -4 -3 -2 -1 0 1 2 3 4 amplitude normalise frequency (a) magnitude spectrum of transfer function -1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1 amplitude time /tmp/tmpQog_8w using 1:2 /tmp/tmp0vdv7f using 1:2 (b) pole-zero plot -1.5 -1 -0.5 0 0.5 1 1.5 imaginary -1.5 -1 -0.5 0 0.5 1 1.5 real 0 10 20 30 40 50 60 70 |H(z)| (c) magnitude z-domain transfer function Figure 10.21 Plots of first order IIR filter, with the single pole at 0.37 +0.58 j. As the pole is not on the real axis, the resonance occurs at a non-zero frequency. is shown in Figure 10.21. The effect of the pole position on the frequency response can be seen more clearly by expressing the pole in polar form, p1 = rejθ. The resonant frequency is given by θ, the angle of the pole, and the bandwidth is controlled as before by r. The first order resonator as just described has one serious drawback: because its pole is not on the real-axis, its coefficient a1 will be complex also, which is a violation of the condition that all the coefficients of the polynomial should be real. This problem can be countered by the use of complex conjugate pairs of poles. If we have a pole with the desired values rejθ, we simply create a second pole with value re− jθ. This will ensure the coefficients are always real, and as the pole has a negative value for θ, its resonance will occur in the negative frequency range, leaving
- 331. Section 10.5. Digital filter analysis and design 307 0 1 2 3 4 5 6 7 -4 -3 -2 -1 0 1 2 3 4 amplitude normalise frequency (a) pole-zero plot -1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1 amplitude time /tmp/tmpV_ny2Q using 1:2 /tmp/tmpYSkHQr using 1:2 /tmp/tmpmWEdeH using 1:2 /tmp/tmp5L1og7 using 1:2 /tmp/tmpwCVli3 using 1:2 (b) frequency response Figure 10.22 Plots of second order IIR filter for a series of poles. All have the same angle, but differ in distance from the unit circle. The poles in the pole-zero diagram are conjugate symmetric, meaning they are symmetric in the real-axis. Because of this the frequency response will have peaks at the same positive and negative values a single pole in the positive range. If we have two poles, we naturally now have a second order filter: h[n] = b0x[n]−a1y[n−1]−a2y[n−2] H(z) = b1 1−a1z−1 −a2z−2 H(z) = 1 (1− p1z1)(1− p2z1) H(ejω ) = b1 1−a1e− jω −a2e−2jω The full frequency response can be seen more clearly by expressing the poles in polar form: p1 = rejθ p2 = p∗ 1 = re− jθ
- 332. 308 Chapter 10. Signals and Filters Using this we can write the transfer function and frequency response as: H(z) = 1 (1−rejθz−1)(1−re− jθz−1) = 1 1−2rcos(θ)z−1 +r2z−2 H(ejω ) = 1 1−2rcos(θ)e− jω +r2e− j2w The following table gives four sets of example poles and coefficients for this filter, and the pole- zero plot and frequency responses for these are shown in Figure 10.22. r θ p1 p2 a1 a2 0.9 1.0 0.48 + 0.75j 0.48 - 0.75j 0.97 -0.81 0.8 1.0 0.43 + 0.67j 0.43 - 0.67j 0.86 -0.64 0.7 1.0 0.38 + 0.59j 0.38 - 0.59j 0.75 -0.48 0.6 1.0 0.32 + 0.51j 0.32 - 0.51j 0.65 -0.36 From the table and the pole-zero plot, the complex conjugate property of the poles can clearly be seen. This property gives rise to the symmetry in the frequency domain about the y-axis. It should also be clear that the relationship between the poles and coefficients is less obvious. These example responses were defined by keeping the angle θ constant, while varying the distance of the pole from the origin, r. It can be seen that as r approaches 1, and the position of the pole nears the unit circle, the frequency response becomes sharper. Exactly the same effect was observed with the first order filter. Moving the poles off the real-axis has the effect of moving the peaks away from the origin in the frequency response. For larger values of r ( 0.5) the resonant frequency is approximated by θ, as for the first order complex pole. For smaller values, the approximation breaks down because the skirts of each pole (that is, the rol-off area on either side of the actual resonance) interact and cause a shifting of frequency. Figure 10.23 and the table below show the effect of keeping r constant while varying the angle θ. As we might expect, the peaks now all have similar shape, but their location in the frequency response is different. r θ p1 p2 a1 a2 0.8 0.75 0.58+0.54j 0.58-0.54j 1.17, -0.64 0.8 1.0 0.43+0.67j 0.43-0.67j 0.86 -0.64 0.8 1.25 0.25+0.76j 0.25-0.76j 0.50 -0.64 0.8 1.5 0.056+0.78j 0.05-0.78j 0.11 -0.64
- 333. Section 10.5. Digital filter analysis and design 309 -1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1 amplitude time /tmp/tmpK2req4 using 1:2 /tmp/tmp33KyD9 using 1:2 /tmp/tmpohsE85 using 1:2 /tmp/tmpeS7T51 using 1:2 /tmp/tmpE1flb- using 1:2 (a) pole-zero plot 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 -4 -3 -2 -1 0 1 2 3 4 amplitude normalise frequency (b) frequency response Figure 10.23 Plots of second order IIR filter for a series of poles with the same radius but different angles. While position of the peaks varies with the angle of the poles. The bandwidth of the peaks is constant, but the amplitudes differ because as the poles near the frequency axis their skirts affect each other, giving each peak a lift. 0 1 2 3 4 5 6 7 -4 -3 -2 -1 0 1 2 3 4 amplitude normalise frequency (a) pole-zero plot 0 2 4 6 8 10 12 -4 -3 -2 -1 0 1 2 3 4 amplitude normalise frequency (b) frequency response Figure 10.24 Plots of second order IIR filter with poles and zeros for different pole positions. Figure (a) has a zero on the unit circle on the real axis. Figure (b) has a zero on the unit circle at the imaginary axis. In general it is the poles which define the most important characteristics of a resonator, namely resonance amplitude, frequency and bandwidth. Zeros do have a role to play and can be used to achieve other effects in the spectrum. Figure 10.24 shows the affect of using a term b1 = 1,
- 334. 310 Chapter 10. Signals and Filters which will place a zero at the origin, and b1 = −1 which will place a zero at the ends limit of the spectrum. Extra zeros can be used to generate notches or anti-resonances in the spectrum. We know that the vocal tract has multiple formants. Rather than developing more and more complicated models to relate formant parameters to transfer functions directly, we can instead make use of the factorisation of the polynomial to simplify the problem. Recall from equation 10.66 that any transfer function polynomial can be broken down into its factors. We can therefore build a transfer function of any order by combining simple first and second order filters: H(z) = H1.H2...HN (10.68) This greatly simplifies the construction of a complicated filter as we can make simple resonators in isolation using pole-zero analysis to set their parameters, and afterwards combine them by multiplication. It should be noted that poles which are close together will interact, and hence the final resonances of a system can not always be predicted from the constituent resonators. Nevertheless, this provides a powerful model for building an arbitrarily complicated filter. 10.5.4 Putting it all together We will now demonstrate some of the practical properties of the LTI filter. the previous section showed how we could build a model of a single formant. We will now extend this to generate a simple three formant model of a vowel. In discussing digital filters, we have mostly used nor- malised angular frequency ω̂. Recall that normal frequency (expressed in Hertz) can be converted to normalised angular frequency by multiplication by 2π/Fs, where Fs is the sampling rate. This gives the following formulas for estimating the pole radius and angle for a single formant of peak frequency f and bandwidth B. θ = 2πF Fs (10.69) r = e−2πB/Fs (10.70) We know from acoustic phonetics that typical formant values for an /ih/ vowel are 300Hz, 2200Hz, and 3000Hz. Formant bandwidths are harder to measure accurately, but less us assume a value of 250Hz for all three formants. Assuming a sampling frequency of Fs = 16000Hz, the following table shows how to calculate the poles from the formant frequencies and bandwidths. Formant Frequency (Hz) Bandwidth (Hz) r θ (normalised angular frequency) pole F1 300 250 0.95 0.12 0.963 + 0.116j F2 2200 250 0.95 0.86 0.619 + 0.719j F3 3000 250 0.95 1.17 0.370 + 0.874j
- 335. Section 10.5. Digital filter analysis and design 311 0 20 40 60 80 100 120 0 1000 2000 3000 4000 5000 6000 7000 amplitude frequency (Hz) (a) frequency response -1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1 imaginary real /tmp/tmpKzOhgH using 1:2 /tmp/tmp8uLEzm using 1:2 (b) pole-zero plot Figure 10.25 Three formant, 6 pole model of /ih/ vowel We construct a second order IIR filter for each formant, Fn: Hn(z) = 1 (1− pnz−1)(1− p∗ nz−1) where pn is the pole and p∗ n is complex conjugate of the pole for the formant n. We create the complete transfer function H/ih/ for the /ih/ vowel by simply multiplying each formant’s transfer function: H/ih/(z) = H1(z)H3(z)H3(z) = 1 (1− p1z−1)(1− p∗ 1z−1)(1− p2z−1)(1− p∗ 2z−1)(1− p3z−1)(1− p∗ 3z−1) Figure 10.25 shows the pole-zero plot and frequency response of the transfer function. From the plot, we can see that the formant frequencies do indeed appear at the approximately the desired locations. By definition, H gives the transfer function and frequency response for a unit impulse. In reality of course, the vocal tract input for vowels is the quasi-periodic glottal waveform. For demonstration purposes, we will examine the effect of the /ih/ filter on a square wave, which we will use as a (very) approximate glottal source. We can generate the output waveforms y[n] by using the difference equation, and find the frequency response of this vowel from H(ejω). The input and output in the time domain and frequency domain are shown in figure 10.26. If the transfer function does indeed accurately describe the frequency behaviour of the filter, we should expect the spectra of y[n], calculated by DFT to match H(ejω)X(ejω). We can see from figure 10.26 that indeed it does.
- 336. 312 Chapter 10. Signals and Filters -1.5 -1 -0.5 0 0.5 1 1.5 0 0.0050.010.0150.020.0250.030.0350.040.0450.05 amplitude time (a) input square wave -1.5 -1 -0.5 0 0.5 1 1.5 0 0.0050.010.0150.020.0250.030.0350.040.0450.05 amplitude time (b) output -4 -3 -2 -1 0 1 2 3 4 5 0 1000 2000 3000 4000 5000 6000 7000 8000 log magnitude frequency (Hz) (c) frequency response H(ejω) for the filter -1 0 1 2 3 4 5 6 0 1000 2000 3000 4000 5000 6000 7000 8000 log magnitude frequency (Hz) (d) spectrum X(ejω) of input -2 -1 0 1 2 3 4 5 6 0 1000 2000 3000 4000 5000 6000 7000 8000 log magnitude frequency (Hz) (e) DFT of output Figure 10.26 Operation of simulated /ih/ vowel filter on square wave. Figure (b) shows the time domain output for the square wave input (a). While the shape of the output is different, it has exactly the same period as the input. Figure (d) is the magnitude spectrum of the square wave. Figure (e) is the magnitude spectrum of the output, calculated by DFT. The harmonics are at the same intervals as the input spectrum, but each has had its amplitude changed. Figure (e) demonstrates that the effect of the filter is to multiply the input spectrum by the frequency response, Figure (c).
- 337. Section 10.6. Summary 313 Filtering a periodic signal demonstrates an important point of all LTI filters. While they am- plify and change the phase of components of an input signal, they do not change the fundamental period of the signal itself. In figure 10.26, it is clear that the output waveform has the same period as the input, and the harmonics have the same spacing in the input and output. Again, this clearly models speech behaviour - if we generate a voiced sound with our glottis, we can move our mouth into a number of configurations, each of which will produce a different sound, but in all cases the fundamental frequency of the sound is unchanged. 10.6 SUMMARY Periodic signals • A complex exponential, ejωt describes a sinusoidal waveform • A harmonic is any integer multiple of the fundamental frequency of a periodic waveform. • Any periodic waveform can be made from a sum of suitably scaled and phase shifted sinu- soids. This is the Fourier series. • Fourier analysis can be used to find the amplitudes and phases of the component harmonics in a periodic waveform. Frequency domain • A representation in the frequency domain is called a spectrum • Spectra are complex, and are usually represented by the magnitude spectrum and phase spectrum. • The ear is relatively insensitive to phase, and so the magnitude spectrum is normally suffi- cient to characterise the frequency domain representation of a signal. • For all real signals, the spectrum is conjugate symmetric. Digital signals • A digital signal can be obtained from a continuous signal by sampling at even intervals, given by the sampling period. • The sampling rate is the reciprocal of the sampling period. • A digital signal has an upper bound on the frequencies it can represent. This is called the Nyquist frequency and is exactly half the sample rate. • Use of normalised frequency, ω̂ = ω/Fs, avoids having to keep track of the sampling rate in digital signal processing expressions. • Spectra obtained from digital signals are periodic, with normalised period 2π.
- 338. 314 Chapter 10. Signals and Filters Transforms • The Fourier transform converts a continuous waveform into a continuous spectrum: X(jω) = Z ∞ −∞ x(t)e− jω dt • The discrete-time Fourier transform converts a digital waveform into a continuous spectrum: X(ejω̂ ) = ∞ ∑ −∞ x[n]e− jω̂n • The discrete Fourier transform converts a digital waveform into a discrete spectrum: X[k] = N−1 ∑ n=0 x[n]e− j 2π N kn • The z-transform converts a digital waveform into a continuous z-domain representation: X(z) = ∞ ∑ n=−∞ x[n]z−n • All transforms have the properties of linearity, scaling, time delay, frequency shift and con- volution. Digital filters • Digital, time-invariant linear filters operate on an input signal x[n] and produce an output signal y[n]. • A filter is fully defined by two sets of real coefficients {al} and {bk}. • The operation of the filter in the time domain is given by the difference equation. y[n] = M ∑ k=0 bkx[n−k]+ N ∑ l=1 aly[n−l] (10.71) • The time domain characteristics of a filter are fully captured by the response h[n] to a unit impulse δ[n]. • The operation of the filter in the z-domain is given by the transfer function. H(z) = b0 +b1z−1 +...+bMz−M 1−a1z−1 −...−aNz−N • The transfer function H(z) is the z-transform of the impulse response h[n].
- 339. Section 10.6. Summary 315 • The frequency response of the filter is given by setting z = ejω which gives: H(e− jω̂ ) = b0 +b1e− jω̂ +...+bMe− jMω̂ 1−a1e− jω̂ −...−aNe− jNω̂ • The z-domain transfer function can be factored into two polynomials, the roots of which are called poles and zeros H(z) = g (1−q1z−1)(1−q2z−1)...(1−qMz−1) (1− p1z−1)(1− p2z−1)...(1− pNz−1) • Poles have an easy to understand interpretation, and in many cases, formant frequency and bandwidth can be directly related to pole values. By suitable choice of pole, a frequency response with appropriate resonances can usually be constructed. • Complicated filters can by created by combining simpler filters. • Filters change the amplitude and phase of the components in the input. They do not change the fundamental period or fundamental frequency.
- 340. 11 ACOUSTIC MODELS OF SPEECH PRODUCTION The speech production process was qualitatively described in Chapter 7. There we showed that speech is produced by a source, such as the glottis, which is subsequently modified by the vocal tract acting as a filter. In this chapter, we turn our attention to developing a more formal quantita- tive model of speech production, using the techniques of signals and filters described in Chapter 10. 11.1 ACOUSTIC THEORY OF SPEECH PRODUCTION Such models often come under the heading of the Acoustic Theory of Speech Production, which refers both to the general field of research in mathematical speech production models, and to the book of that title by Fant [158]. While considerable previous work in this field had been done prior to its publication, this book was the first to bring together various strands of work and describe the whole process in a unified manner. Furthermore, Fant backed his study up with extensive empirical studies with x-rays and mechanical models to test and verify the speech production models being proposed. Since then, many refinements have been made to the model, as researchers have investigated trying to improve the accuracy and practicalities of these models. Here we focus the single most widely accepted model, but conclude the chapter with a discussion on variations on this. As with any modelling process, we have to reach a compromise between a model which accurately describes the phenomena in question, and one which is simple, effective and suited to practical needs. If we tried to capture every aspect of the vocal organs directly, we would have to account for every muscle movement, articulator shape and tissue absorption characteristic directly, and then determine how each of these affected how speech production. This would be a huge undertaking, and even if an accurate description of all these factors could be determined, the result may be too complicated for useful analysis. We therefore have to make quite a number of simplifying assumptions in order to obtain a usable model. 316
- 341. Section 11.1. Acoustic Theory of Speech Production 317 nasal cavity oral cavity tongue pharynx glottis larynx vocal folds hard palate soft palate lips nostril alveolar ridge teeth velum uvula (a) mid-sagittal drawing of vocal organs (b) Model of vocal organs with dis- crete components identified Figure 11.1 Diagram and model of vocal organs. 11.1.1 Components in the model Our first task is to build a model where the complex vocal apparatus is broken down into a small number of independent components. One way of doing this is shown in Figure 11.1b, where we have modelled the lungs, glottis, pharynx cavity, mouth cavity, nasal cavity, nostrils and lips as a set of discrete, connected systems. If we make the assumption that the entire system is linear (in the sense described in Section 10.4) we can then produce a model for each component separately, and determine the behaviour of the overall system from the appropriate combination of the components. While of course the shape of the vocal tract will be continuously varying in time when speaking, if we choose a sufficiently short time frame, we can consider the operation of the components to be constant over that short period time. This, coupled with the linear assumption then allows us to use the theory of linear time invariant (LTI) filters (Section 10.4) throughout. Hence we describe the pharynx cavity, mouth cavity and lip radiation as LTI filters, and so the speech production process can be stated as the operation of a series of z-domain transfer functions on the input.
- 342. 318 Chapter 11. Acoustic Models of Speech Production mass Figure 11.2 Simple mass/spring oscillator. If the mass is displaced from its original, at rest posi- tion, it will indefinitely oscillate horizontally. For example, in the case of vowels, speech is produced by the glottal source waveform travelling through the pharynx, and as the nasal cavity is shut off, the waveform progresses through the oral cavity and is radiated into the open air via the lips. Hence as filters connected in series are simply multiplied in the z-domain, we can write the system equation for vowels as: Y(z) = U(z)P(z)O(z)R(z) (11.1) Where U(z) is the glottal source, with P(z), O(z) and R(z) representing the transfer functions of the pharynx, the oral cavity and the lips respectively. As P(z), O(z) linearly combine, it is normal to define a single vocal tract transfer function V(z) = P(z)O(z), such that Equation 11.1 is written Y(z) = U(z)V(z)R(z) (11.2) which can be represented in the (discrete) time domain as: y[n] = u[n]⊗v[n]⊗r[n] (11.3) The configuration of the vocal tract governs which vowel sound is produced, and by studying this we can again an understanding of how the physical properties of air movement relate to the transfer function, and sounds produced. In a similar fashion, transfer functions can be determined for the various types of consonants. To determine the form these transfer functions take, we have to investigate the physics of sound, and this is dealt with next. 11.2 THE PHYSICS OF SOUND 11.2.1 Resonant systems We have informally observed that the vocal tract filter acts as a resonator; that is, it amplifies certain frequencies and attenuates others. How does this behaviour arise? The resonant nature of systems is often demonstrated by considering the motion of a mass and spring as shown in Figure 11.2. Imagine this system is set in motion by displacing the mass to
- 343. Section 11.2. The physics of sound 319 spring damper mass Figure 11.3 Mass/spring/damped system. The mass and spring combine as before to create oscil- lation, but the damper continually dissipates energy when the system is in motion, and will eventually cause the system to stop. the right and then letting go. The extended spring will try to return to its original position and in doing so will set the mass in motion. Initially, the speed of movement will be slow as the inertia of the mass must be overcome. As the spring reaches its original position, it will no longer exert any force, but as the mass is now moving, it will continue to move and pass the original position, moving to the left. As it does so, the spring will again be displaced from its original position, and the more this is done, the stronger the force will become, until eventually the mass stops and the spring pulls the mass towards the at rest position again. From Newton’s laws we also know that the system will continue moving until acted upon by an external force. The frequency at which the system oscillates is determined by the inertia of the mass and the compliance of the spring; resonant frequency = 1 2π r k m Hence a heavier mass will mean a slower oscillation and a stiffer spring will mean a quicker oscillation. The important point is that the system has a single resonant frequency determined by the physical properties of the mass and spring. The above equates to a source filter system where a single impulse has been given to the sys- tem; we saw in Chapter 10 that the impulse response from a system is a good way to characterise the system itself independent of the input. Now let us consider the case where a periodic forcing function is exerted on the mass, shown in Figure 11.3. Assuming the force is of sufficient strength, the mass will move with the periodicity of the driving (source) force. Again, this is what we would expect from Chapter 10, where we showed that the source and not the filter largely determines the output frequency of a system. Imagine now that we send sources of different frequencies into the system. We will find that at frequencies near to the resonant frequency the size of the oscillations will increase. In fact, at the resonant frequency itself, the oscillations will get larger and larger, and if no other factors are involve, will eventually become so large as to break the system1 Hence 1 This is what happens when things “shake themselves apart”, most famously the Tacoma Narrows bridge.
- 344. 320 Chapter 11. Acoustic Models of Speech Production damper mass spring forcing function Figure 11.4 Mass/spring/damped system driven by an external force. Here a periodic force is applied, and this will determine the frequency at which the system oscillates. As the driving frequency nears the natural resonant frequency (determined by the mass and spring) the size of the oscillations will increase. a mass-spring system with a driving force acts as a linear filter fed with a source. The above is of course somewhat idealised because we know that in general systems don’t shake themselves apart, and in the unforced case, when a mass and spring are set in motion, eventually the motion will stop. This is due to the third property known as resistance. In the mechanical world, resistance is opposition to velocity, and is often described as a damping effect on the system. Normally resistance is caused by friction in which the energy in the movement is converted into heat. As in general the heat cannot be converted back into motion energy, the effect of friction is to cause energy loss in the system, which will eventually bring the system to a halt. One of the important features of resistance is that usually its effect on the frequency of the resonance is negligible; that is, while the amount of resistance will determine how quickly the oscillations die out, it does not affect the frequency at which those oscillations occur. This is an important point for our model, as it often allows us to study the frequencies of resonances without needing to consider the damping effects. All these affects were shown in our examination of linear filters in Chapter 10. For exam- ple, Figure 10.15a shows the displacement of the mass against time for the undamped case, while Figure 10.15b shows the same system with damping. Figures 10.22 and 10.23 show the separa- tion of frequency and damping effects on a pole-zero plot. There, the angle of the pole or zero determined the frequency, whereas the radius of the pole determined the damping, with poles on the unit circle meaning that the damping was zero. In terms of a frequency response, the damping manifests itself as amplitude and bandwidth of the resonant peak, with high amplitude and narrow bandwidth arising from cases of low damping. The three factors of inertia (mass), resistance and capacitance (the reciprocal of spring com- pliance) are collectively known as impedance as intuitively, the system “impedes” the motion in the system.
- 345. Section 11.2. The physics of sound 321 fixed end Figure 11.5 If the held end of the rope is given a jerk, this creates an impulse, which will travel along the rope towards the wall. 11.2.2 Travelling waves The behaviour of the vocal tract is in many ways similar to the behaviour of the mass/spring/damped system just described. The input, “forcing function” is the glottal source, which is used to drive the system (the vocal tract) whose physical properties determine resonant characteristics, which then modify the input signal so as to amplify some frequencies and attenuate others. One crucial differ- ence however, is that the mass/spring/damped system is what is known as a lumped parameter system in that it comprises easily identifiable, discrete components such as the mass and spring. By contrast, the vocal tract is a distributed system, in which the properties of inertia, capacitance and resistance are evenly distributed throughout the system. We can formulate the properties of a distributed system by modelling it as a large number of small connected discrete systems; that is. Using calculus, we can determine the behaviour of a distributed system by considering the behaviour of the system as the size of the elements tends to zero. This steps involved in this are fully described in Rabiner and Schafer [368]; for our purposes though, we will simply take the results of this analysis and apply it to the problem of the sound propagation in the vocal tract. Before considering this directly, we will appeal again to an equivalent physical system, as this will let us examine an aspect of resonant behaviour that is central to our study, namely that of travelling waves. Consider first the easy to visualise example of waves travelling in a rope, in which one end is in a person’s hand and the other is tied to a wall. This is shown in Figures 11.5 and 11.6. If the person introduces some energy in the rope by giving the rope a jerk (i.e. an impulse), a wave will form at that end and travel down the rope. When the wave reaches the tied end at the wall, it will be reflected, and will travel backwards towards the person. When it reaches the held end, the wave will be reflected back again towards the wall and so on, until all the energy has eventually been lost from the system. The person can create a forcing function (i.e. source) input to the system by moving their hand in a repetitive fashion. In doing so, a series of pulses will form and travel forwards down the rope towards the wall, and these pulses will create a travelling wave, the frequency of which is dictated by the hand movement. When this forwards travelling wave is reflected, a backwards travelling wave will be created. The forward wave and backward wave will interfere - if the two have a peak at a given point, a large peak will occur, and if one has a negative peak while the other has a positive peak then these will cancel out leaving the rope at that point in its neutral
- 346. 322 Chapter 11. Acoustic Models of Speech Production Figure 11.6 The impulse in the rope travels towards the wall, where is is reflected out of phase with the original pulse. The reflected pulse travels backwards along the rope. position. In general, the observed shape of the rope will not form a simple pattern. However at some frequencies, the reflected wave will exactly reinforce the forward wave, and the rope will settle down into a fixed pattern. In such cases, the wave does not look like it is travelling at all, so this is termed a standing wave. The term is a little misleading because it might be taken to mean that there is no motion - this is not the case. It merely appears that the waves are not moving. Standing waves are only created at certain frequencies. At these frequencies, the wave is amplified, that is the overall displacement in the rope at a peak is greater than the input. Thus the rope acts as a resonator, amplifying certain frequencies and attenuating others. The frequencies at which the resonances occur are governed by the length of the rope and the speed of wave travel. The above demonstrates some of the fundamental properties of travelling waves. First, note that some of the behaviour is determined by the motion of the hand, while some is determined by
- 347. Section 11.2. The physics of sound 323 displacement rarefraction compression forcing function tube Figure 11.7 Longitudinal sound wave travelling along a tube. The wave travels by means of com- pression and rarefaction of the air particles. The displacement of a slice of air particles against position in the tube is shown below. the properties of the rope itself. For example, the frequency of the waves is determined purely by the rate of movement of the hand; the rope itself has no influence on this at all. In contrast, the rate at which the wave travels along the rope is not influenced by the hand at all; no matter how quickly or slowly the hand moves, the wave always travels at the same speed, and this speed is determined by the properties of the rope. This shows why it is appropriate to use the theory of signals and filters where an initial source (the wave created by the moving hand) has its behaviour modified by a filter (the rope). Furthermore, we should also clearly see the parallels between this and the behaviour of the glottis source and vocal tract filter in speech production. What then determines the behaviour of the rope? Firstly, we have the physical properties of the material of the rope; its inherent stiffness, its length, its mass and so on. Secondly, we have what are known as the boundary conditions, that is, the specifics of how the rope is terminated at either end. In the above example, one end of the rope was attached to a wall; there are other possibilities, such as the end of the rope being free - in such a case reflection also occurs, but in a different way from the fixed case. 11.2.3 Acoustic waves The properties of sound waves travelling in a tube are very similar to those of waves travelling in a rope. Instead of the wave travelling by means of moving the rope, a sound wave travels in air by a process of displacement of the particles of air. In the absence of any particular sound, the air has certain ambient properties, such as pressure and temperature. The effect of a sound source in the air causes the particles to move backwards and forwards and in doing so, the wave spreads from the source. This is shown in Figure 11.7, where a forcing function at one end of the tube causes a disturbance which travels along the tube, causing compressions, where the particles come closer together, and rarefaction where they move further apart. This is similar to the situation with the rope, in which the wave travels along the rope, but the rope itself is only temporarily displaced; the rope doesn’t end up travelling with the wave. One difference is that sound waves are longitudinal whereas waves in a rope are transverse. This means that rather than the particles of sound moving up and down relative to the direction
- 348. 324 Chapter 11. Acoustic Models of Speech Production of wave motion, they move backwards and forwards along the axis of wave motion. Despite this significant physical difference, the behaviour can be modelled in exactly the same way. As with the rope, the specifics of how the wave travels in the air are determined by the physical properties of the air, and in the case of standing waves by the boundary conditions. The physics of general sound wave propagation are complicated and form an extensive area of study. For our purposes however, we only need consider one specific type of sound propagation, that of sound propagation in a tube. Furthermore, we can assume that the tube has constant cross sectional area along its length, as this can generalise to cases where the tube varies in area. An important property of the air is its speed of sound, denoted by c. For a given pressure and temperature, this is constant, and while this can be calculated from more fundamental properties, it is easily measured empirically. A typical value is that a sound wave travels 340 metres in one second at room temperature and pressure. By speed of sound we mean the distance travelled by one part of the wave in unit time. Note the similarity with the rope; as the speed of sound propagation is constant, it doesn’t matter what the source does, all waves travel at exactly the same speed. In sound propagation in general, the signal or wave is manifested as particle velocity. This is the pattern of movement of air particles which makes up the sound wave. In tubes however, the mathematics is considerably simplified if we use a related quantity volume velocity, which is just the particle velocity multiplied by the area. As the area of the tube is constant, this doesn’t complicate our analysis. The air in the tube has an impedance Z, and the effect of this is to “impede” the volume velocity. This brings us to pressure, which can be thought of as the work required to move the air particles through a particular impedance. The fundamental equation linking the three is: P = ZU (11.4) Hence for a constant pressure a high impedance will mean a reduced particle velocity, or con- versely to maintain a constant particle velocity, more pressure will be needed if the impedance increases. This can easily be visualised; imagine pushing something in a pipe; if the pipe is un- constricted (nothing “impeding”, low impedance) a small amount of pressure (effort) will move something down the tube. If the tube is blocked, considerably more effort will be required to move the object. Acoustic impedance is formulated in a similar fashion to the mechanical impedance de- scribed in Section 11.2.1. First we have acoustic inductance, meaning that when a sound wave passes through the air, the inertia in the air impedes any acceleration of the air particles. Next we have a property called acoustic capacitance, which is equivalent to the effect of the spring in the mechanical system. This arises because th air is compressible, and the capacitance is a measure of how strongly the air resists this compression. Unlike inertia, this is a property which reacts to dis- placement, that is, distance moved, rather than acceleration. Finally we have acoustic resistance, which dampens the system and is responsible for energy loss.
- 349. Section 11.2. The physics of sound 325 u+ u − Figure 11.8 A uniform tube with forwards travelling wave, u+ , and backwards travelling wave u− . In our uniform tube, the impedance is distributed evenly along the entire length of the tube rather than existing at a single point, as in a single mass or spring. Hence, the length of the tube is an important factor and the distributed nature means that determining the resonant characteristics is somewhat more complicated. As length is a factor a standard measure of impedance in such systems is call the characteristic impedance which is the impedance that would occur if the tube was infinitely long. This is termed Z0 and is given Z0 = ρc (11.5) where ρ is the density of the air. In air, the inductance is given by density divided by area ρ/A and capacitance is given by A/ρc2. For now, we shall assume that the resistance is air is in fact zero. We do this firstly because it helps simplify the derivation of the full model, but in fact this is further justified because the friction due to the air itself is relatively minor, and in reality other resistive factors (such as the contact with the tube walls) provide a greater (and more complicated) contribution to the overall energy loss. Tubes which have a zero resistance component are termed lossless tubes for obvious reasons. 11.2.4 Acoustic reflection Consider again the properties of properties of waves travelling in a rope. There we saw that at the termination of the rope, the travelling wave was reflected, and the reflected backwards travelling wave interfered with the forward wave to determine the behaviour of the complete system. Trav- elling sound waves move along the tube in a similar way. Depending on the boundary conditions at the end of the tube, some portion of the acoustic wave is reflected back, and this interferes with the wave travelling forwards. At specific frequencies, governed by the length of the tube and the speed of sound, the backwards and forwards waves will reinforce each other and cause resonances. In such a system, the volume velocity at a given point and time can be expressed as: u(x,t) = u+ (t −x)−u− (t +x) (11.6) where u+(t) is the forward travelling wave, and u−(t) is the backwards travelling wave. This is shown in Figure 11.8. The minus sign between the terms indicates that the two waves can cancel
- 350. 326 Chapter 11. Acoustic Models of Speech Production lungs lips glottis Figure 11.9 Vocal tract model composed of sequence of joined uniform tubes. each other out; when the forward wave meets an equal size volume velocity of the backward wave, the overall effect is to have no motion. Equation 11.5 allows us to state an expression for pressure in terms of the forward and backward volume velocities at a point in space and time in a similar way p(x,t) = ρc A u+ (t −x)+u− (t +x) (11.7) Here ρc/A is the characteristic impedance Z0 of the tube. Note that in this case, the terms are additive; an increase in pressure from the forward wave meeting an increase in pressure from the backward wave results in an even higher combined total pressure. If the area of the tube remains constant, the wave simply propagates through the tube. If however the area changes, then the impedance changes, and this causes reflection. The reflections set up standing waves and these cause resonances. In this way, the impedance patterns of the tube govern the resonant properties of the model2 11.3 VOWEL TUBE MODEL In vowel production, the glottal source creates a wave which passes through the pharyngeal cavity, then the oral cavity and finally radiates through the lips. The nasal cavity is blocked off, and so the pharyngeal cavity and mouth can be combined and modelled as a single tube. This tube varies in cross sectional area along its length, and it is the ability of a speaker to vary the configuration of this tube that gives rise to the different vowel sounds. Modelling a continuously varying tube is complicated, but we can approximate this by considering a series of short uniform tubes connected in series, shown in Figure 11.9. If the number of tubes is sufficiently large, a continuous tube can be modelled with arbitrary accuracy using this method (we return to the issue of just how many tubes are needed in Section 11.3.6). Recall that we described the overall process of vowel production as one where a signal from the glottis u[n] travels through and is modified by the vocal tract and lips to produce the speech waveform y[n]. In the z-domain, this is represented by Equation 11.2: Y(z) = U(z)V(z)R(z) 2 Note that the above expressions can be derived from a first principles solution of the wave equations and the proper- ties of air. This is beyond the scope of our study, but several good accounts are available [368] [364].
- 351. Section 11.3. Vowel Tube Model 327 where U(z) is the z-transform of the glottal source, V(z) is the vocal tract filter and R(z) is the radiation effect from the lips. Initially, we wish to find V(z) as the is the part of the expression that characterises the phonetic nature of the sound. From V(z) we can find the spectral envelope, formants and other linguistically useful information. To find V(z), we study the evolution of the volume velocity signal from the glottis to the lips. We will denote the volume velocity at the lips UL(z), and to avoid confusion, with denote the volume velocity glottal source as UG(z) rather than simply as U(z). Hence our expression is UL(z) = V(z)UG(z) (11.8) After this, we can determine the final speech signal Y(z) by applying the radiation characteristic R(z) to UL(z). Rather than attempt to find V(z) from the tube properties directly, we will use the tube model to find UL(z) in terms of UG(z), and divide these to find V(z). 11.3.1 Discrete time and distance In practice we wish to use our analysis on digital signals, so our first task is to convert the expres- sions for volume velocity and pressure in a tube into a digital form, so that we can perform the analysis independent of sampling rate. Recall from Section 10.2.1 that we can convert time t to a discrete representation time vari- able n by n = tFs where Fs is the sampling rate. As is standard in discrete signal processing, we will use normalised frequency (F̂) and normalised angular frequency (ω̂), which are calculated from frequency (F) and angular frequency (ω) by F̂ = ω̂ 2π = F Fs In addition, it is convenient to normalise our measure of distance from the start of the tube, x, with respect to the sampling frequency. It also helps simplify the analysis if we normalise this with respect to the speed of sound. As with time and frequency normalisation, this allows us to derive general expressions for the system independent of the sampling frequency used and the speed of sound in the medium. Hence normalised distance d is given by d = xFs c (11.9) Using this result we can state the volume velocity and pressure functions, given in Equations 11.6, 11.7 and 11.9, in discrete normalised time and distance as u[d,n] = u+ [n−d]−u− [n+d] (11.10a) p[d,n] = ρc A h u+ [n−d]−u− [n+d] i (11.10b)
- 352. 328 Chapter 11. Acoustic Models of Speech Production U [n − D ] k + k − U [n] − U [n] k + U [n] k+1 − U [n − D ] k+1 + U [n − D ] k+1 − U [n] k+1 + − + k − U [n + D ] Figure 11.10 Behaviour of forwards and backwards travelling waves at a tube junction. In the next section we will develop a general expression for the reflection and transmission of a wave at the junction of two tubes and in in the subsequent sections, show how this result can be used to determine the behaviour of the whole system. 11.3.2 Junction of two tubes If we consider a single tube k in our model, from Equations 11.6, 11.7 and 11.9 we can state the volume velocity and pressure functions in discrete normalised time and distance as uk[d,n] = u+ k [n−d]−u− k [n+d] (11.11a) pk[d,n] = ρc Ak h u+ k [n−d]−u− k [n+d] i (11.11b) Now we consider the junction between a tube k with normalised length Dk and its neighbouring tube k + 1, shown in Figure 11.10. As this tube is also uniform, it can also be described by Equations 11.11a and 11.11b. Joining two tubes creates a boundary condition at the junction, and it is this junction which determines how the sound propagates as a whole. In considering behaviour at a junction, we can use an important principle of physics which states that pressure and volume velocity cannot change instantaneously anywhere. So despite the sudden change in cross sectional area, the volume veloc- ity and pressure cannot abruptly change, and it follows from this that at the point of the junction the pressure and volume velocity must be equal. Hence uk[Dk,n] = uk+1[0,n] pk[Dk,n] = pk+1[0,n] Using these boundary conditions we can relate the forwards and backwards volume velocity equa- tions of tube k and tube k+1 at the junction. That is, at the end (d = Dk) of tube k and the start of tube k +1 (d = 0) u+ k [n−Dk]−u− k [n+Dk] = u+ k+1[n]−u− k+1[n] (11.12)
- 353. Section 11.3. Vowel Tube Model 329 U [n] U [n] U [n − D ] A A transmission coefficient reflection coefficient k + + k+1 − k+1 k k+1 Figure 11.11 Tube junction showing how wave u+ is made from reflection of U- in tube k+1, and transmission of u+ from tube k and likewise with the pressure: ρc Ak u+ k [n−Dk]+u− k [n+Dk] = ρc Ak+1 u+ k+1[n]+u− k+1[n] (11.13a) 1 Ak u+ k [n−Dk]+u− k [n+Dk] = 1 Ak+1 u+ k+1[n]+u− k+1[n] (11.13b) we can substitute these equations into one another and eliminate one of the volume velocity terms. For example, from Equation 11.12 we know that u− k [n + Dk] = u+ k+1[n] − u− k+1[n] + u+ k [n − Dk], which we can substitute into Equation 11.13b to give an expression for u+ k+1[n] u+ k+1[n] = h 2Ak+1 Ak +Ak+1 i u+ k [n−Dk]+ hAk+1 −Ak Ak +Ak+1 i u− k+1[n] (11.14) This equation is interesting in that it shows the make up of the forward travelling wave in tube k +1. As we would expect, this is partly made up of the forward travelling wave from tube k and partly from some of the backward travelling wave being reflected at the junction. The coefficient of u+ k in Equation 11.14 is the amount of u+ k that is transmitted into the next tube, k + 1, and is termed the transmission coefficient. The coefficient of u− k+1 is the amount of u− k+1 that is reflected back into tube k + 1 at the junction and is termed the reflection coefficient. As they are simply related, only one is needed to describe the junction and by convention reflection coefficients are used. The reflection coefficient is often denoted rk (meaning the reflection between tubes k and k +1) and from 11.14 is given by: rk = Ak+1 −Ak Ak +Ak+1 (11.15) Of particular importance is that fact that the boundary condition is completely described by the reflection coefficient, and this is only governed by the areas of the tubes; the speed of sound and
- 354. 330 Chapter 11. Acoustic Models of Speech Production density terms have been eliminated. We shall see that reflection coefficients are a simple and useful way to define the characteristics between two tubes and we will henceforth use them instead of tube areas. Using reflection coefficients, we can find expressions that relate the forward wave or the backward in one tube with the forward and negative waves in the other tube: u+ k [n−Dk] = 1 1+rk u+ k+1[n]+ rk 1+rk u− k+1[n] (11.16a) u− k [n+Dk] = −rk 1+rk u+ k+1[n]+ 1 1+rk u− k+1[n] (11.16b) 11.3.3 Special cases of junction The above result can be used to help set boundary conditions for the tube at the lips and the glottis. To see this, let us consider three special cases resulting from changing areas in the tubes. Firstly, if two connected tubes have the same cross sectional area we see from 11.15 that the reflection coefficient is 0 and the transmission coefficient is 1. That is, there is no reflection and all the wave is transmitted past the boundary. This is entirely what we would expect from our uniform tube model. Next, let us consider the situation where the tube is completely closed at one end. A closed termination can be modelled by having the second tube have an infinitely small area (that is, the size of the junction between the tubes disappears to nothing and hence creates a solid wall). The effect of this can be found from Equation 11.15. As the area of the second tube, Ak+1 → 0, the reflection coefficient rk → −1. Now consider a completely open tube. This can again be modelled with the tube junction model, but this time the area of the second tube is infinitely big. Hence, from Equation 11.15, as Ak+1 → ∞ the reflection coefficient rk → 1. Hence both a closed termination and a open termination completely reflect the wave. There is a crucial difference between the two situations however and this can be seen from the sign of the reflection coefficient. In the closed termination case, the tube ends in a solid wall and the reflection coefficient is -1. From Equation 11.16b we can see that in such a case the volume velocity will be 0. If we find the equivalent pressure expressions for Equation 11.16b, we see that the pressure terms add, and hence the pressure at this point is (ρc)/A1. Intuitively, we can explain this as follows. The solid wall prevents any particle movement and so the volume velocity at this point must be 0. The pressure however is at its maximum, as the wall stopping the motion is in effect an infinite impedance, meaning no matter how much pressure is applied, no movement will occur. In the open end situation, when the reflection coefficient is 1, we see from Equation 11.16b that the volume velocity is at a maximum and now the pressure is 0. Intuitively, this can be explained by the fact that now the impedance is 0, and hence no pressure is needed to move the air particles. Hence the pressure p(L,t) is 0, and the volume velocity is at its maximum.
- 355. Section 11.3. Vowel Tube Model 331 lips lungs r r G 1 glottis tube 1 tube 2 r L reflection coefficients open air Figure 11.12 Two tube model with terminations at the glottis and lips. This system has three reflection coefficients. In the middle we have r1 the only “real” reflection coefficient, whose value is given by the area Equation 11.15. The lips and glottis reflection coefficients are artificial values designed to ensure there are some losses from the system. Note that there is no backwards wave entering at the lips, and no forwards wave (save the source) entering at the glottis. 11.3.4 Two tube vocal tract model Here we will consider a model in which we have two connected tubes of different cross sectional areas. In this formulation, the tubes have areas A1 and A2, and normalised lengths D1 and D2, such that D1 + D2 = D, the total normalised length of the vocal tract. Using the results from Section 11.3.3, we represent the behaviour at the boundary by the reflection coefficient r1. The lips and glottis are modelled by two special tubes which have reflection coefficients rL and rG respectively. We can set these so as to produce different effects - one configuration is to have the glottis completely closed and the lips completely open; another is to have values for each coefficient being slightly less than this value so that some losses occur and so that some sound propagates from the lips. These lip and glottis tubes are special because we add an extra condition such that any tube entering these tubes is never reflected back again. Examining a two tube model in depth gives us a feel for how this system operates as a whole; following this section we will consider the behaviour of a single tube, the behaviour when we consider losses, and a final model where we have an arbitrary number of tubes. Our goal is to find the z-domain transfer function of the system, and we will find this by dividing the z-domain expressions for output at the lips by the input at the glottis. The best way to proceed is to define the volume velocity at the lips, use this to find the volume velocity at the junction between the two proper tubes, and use this to find the volume velocity at the glottis. To do this we use Equations 11.16b and 11.16b which relate the volume velocity in one tube to the
- 356. 332 Chapter 11. Acoustic Models of Speech Production volume velocity in the next. As our end point is a z-domain expression, we will make repeated use of the z-transforms of equations 11.16b which are U+ k (z) = zDk (1+rk) U+ k+1(z)− rkzDk (1+rk) U− k+1(z) (11.17a) U− k (z) = −rkz−Dk (1+rk) U+ k+1(z)+ z−Dk (1+rk) U− k+1(z) (11.17b) We start by defining UL(z) as the output at the lips. We now feed this into Equations 11.17b and 11.17b and use the fact that there is no backwards wave (U− L = 0) to set the volume velocities at junction k = 2 as: U+ 2 (z) = zD2 1+rL UL(z) (11.18a) U− 2 (z) = −rLz−D2 1+rL UL(z) (11.18b) By feeding these values into Equation 11.17b again, we can calculate the volume velocities U+ 1 and U− 1 at the junction k = 1. U+ 1 (z) = zD1 1+r1 .U2(z)− r1zD1 1+r1 U2(z) = zD1 .zD2 (1+r1)(1+rL) UL(z)+ r1rLzD1 z−D2 (1+r1)(1+rL) UL(z) U− 1 (z) = −r1z−D1 1+r1 .U2(z)+ z−D1 1+r1 U2(z) = r1z−D1 .zD2 (1+r1)(1+rL) UL(z)− z−D1 z−D2 (1+r1)(1+rL) UL(z) We have now reached the glottis, modelled by a special tube of length 0 and reflection coefficient rG. We take the length of this tube to be 0, and as z0 = 1 then Ug(z) = = 1 1+rG U+ 1 (z)− rG 1+rG U− 1 (z) To find U+ G we simply repeat the above procedure again. Note that we don’t need to know U− G as we are assuming this gets fully absorbed into the lungs. U+ G (z) = zD1 zD2 +r1rLzD1 z−D2 +r1rGz−D1 zD2 +rLrGz−D1 z−D2 (1+rG)(1+r1)(1+rL) UL(z)
- 357. Section 11.3. Vowel Tube Model 333 r = 1 L complete reflection complete reflection G r = −1 glottis lips open air Figure 11.13 Uniform tube model of vocal tract where the glottis end is closed and the lips end open By rearranging to get UL/UG and by dividing top and bottom by zD1 zD2 we get the standard form of the transfer function as a rational function in z−1: V(z) = UL UG = (1+rG)(1+r1)(1+rL)z−(D1+D2) 1+r1rLz−2D2 +r1rGz−2D1 +rLrGz−2(D1+D2) The terms containing r in the numerator sum to form the gain, G. The normalised tube lengths D1 and D2 must sum to the overall length D1+D2 = D, which gives V(z) = UL UG = Gz−D 1+r1rLz−2(D−D1) +r1rGz−2D1 +rLrGz−2D (11.19) Note that the lengths of the tubes D1,D2 must be integers; this is a constraint imposed by the discrete modelling of the vocal tract. The discrete modelling also dictates the number of poles: for a higher sampling rate requires more poles and hence the granularity of the tube lengths increases. 11.3.5 Single tube model A special case of the two tube model is the single tube or uniform tube model where the cross sectional area is constant along the entire length of the tube. This is a reasonable approximation of the vocal tract when producing the schwa vowel. If A1 = A2, then the refection coefficient r1 = 0 and hence Equation 11.19 simplifies to V(z) = Gz−D 1+rLrGz−2D (11.20) Recall that D is the length of the vocal tract in units of the normalised distance that we used to simplify the equation for discrete time analysis. By use of Equation 11.9, we can determine a
- 358. 334 Chapter 11. Acoustic Models of Speech Production value for D from empirically measured values. For instance, if we set the length of the vocal tract as 0.17m (the average for a male), the speed of sound in air as 340ms-1, and choosing a sampling rate of 10000Hz, Equation 11.9 gives the value of D as 5. Hence the transfer function becomes V(z) = Gz−5 1+rLrGz−10 (11.21) that is, an IIR filter with 5 zeros and 10 poles. It can be shown that a polynomial 1+rLrGz−10 can be factorised to a set of 10 complex roots, such that in polar terms the radius of each root is rLrG, and the angle is given by θn = π(2n−1) 10 (11.22) The numerator of the transfer function tells us that the output is delayed by a factor which is a function of the length of the tube. All these zeros lie at 0, and can be ignored for purposes of determining the frequency response of the transfer function. The poles are evenly spaced on the unit circle at intervals of π/5. The first is at π/10 = 0.314, which when converted to a real frequency value with Equation 10.29, gives 500Hz. Subsequent resonances occur every 1000Hz after that, i.e. 1500Hz, 2500Hz and so on. Recall that the losses in the system come from the wave exiting the vocal tract at the glottis and lips, and the degree to which this is done is controlled by the special reflection coefficients rL and rG. Let us first note, that since these are multiplied, it doesn’t affect the model as to where the losses actually occur, and in fact in some formulations the losses are taken to occur only at the glottis or only at the lips. Secondly, recall from Section 11.3.3, that the special case of complete closure at the glottis would produce a reflection coefficient of 1, and the special condition of complete openness at the lips would produce a reflection coefficient of -1. This is a situation where there are no losses at all in the system (unrealistic of course, for the simple fact that no sound would escape and hence a listener would never hear the speech.) For these special cases the product is rLrG is 1, which means that the denominator term is V(z) = Gz−5 1−z−10 It can be shown that when this is factorised, all the poles are spaced as before, but lie on the unit circle. This is of course exactly what we would expect from a lossless case. We can find the frequency response of the vocal tract by setting z = ejω̂. For the special lossless case then, V(ejω̂ ) = 2ej5ω̂ 1+ej10ω̂
- 359. Section 11.3. Vowel Tube Model 335 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 0 2 4 6 8 10 12 14 16 18 20 Figure 11.14 Magnitude spectrum for uniform tube, showing that the resonances are sharp spikes (because there is no damping) and that the resonances lie at even intervals along the spectrum. Making use of the inverse Euler expression, Equation 10.11 can also be written as: V(ejω̂ ) = 1 cos(5ω̂) (11.23) This frequency response is shown in Figure 11.14, and it can be seen that the formants are very sharp (no damping) and lie at evenly spaced intervals, as predicted by the transfer function. 11.3.6 Multi-tube vocal tract model We will now develop a general expression for the transfer function of a tube model for vowels, in which we have an arbitrary number of tubes connected together. We saw in Section 11.3.5 that the order of the final transfer function is determined by the normalised length of the tube D. This is turn is a function of sampling rate and vocal tract length and so longer vocal tracts or higher sampling rates will require higher order filters. Staying with the case of a 0.17m vocal tract with 340ms-1 speed of sound and 10000Hz sample rate, we find we need a 10th order transfer function as shown in the cases of the uniform and two tube models. The difference in the transfer function for these two models is simply in the inclusion of the extra non-zero coefficients in the denominator. Extending this principle, we find that the most powerful transfer function is one with a non-zero coefficient for every term in z−1. Such a transfer function can be created by a model with exactly 10 equal length tubes, each of which has length DN = 1/2. Importantly, for the given sample rate, no extra accuracy is achieved by increasing the number of tubes beyond this value.
- 360. 336 Chapter 11. Acoustic Models of Speech Production The principle behind finding the N tube transfer function is the same as for the two tube model - start at the lips and apply Equations 11.18a and 11.18b for each tube in turn. This is tedious to perform in the manner used for the two tube system, but we can use matrix multiplication to speed up the process. Equations 11.17b and 11.17b can be expressed in matrix form as: Uk = QkUk+1 where Uk = U+ k (z) U− k (z) # and because D = 1/2, then Qk = z1/2 1+rk −rkz1/2 1+rk −rkz−1/2 1+rk z−1/2 1+rk # We can use the special reflection coefficients at the lips and glottis to model the terminating bound- ary conditions as before. In the matrix formulation, this gives UN+1 = UL(z) 0 # = 1 0 # UL(z) (11.24) for the lips and UG = h 1 1+rG 1rG 1+rG i U1(z) (11.25) for the glottis. We can use these boundary conditions to formulate an expression for the transfer function for an arbitrary number of tubes 1 V(z) = UG UL = h 1 1+rG 1rG 1+rG i N ∏ k=1 Qk 1 0 # which can be solved by iteratively calculating the terms. The result of this will always take the form: V(z) = 1+r0 2 z−N/2 N ∏ k=1 (1+rk) 1−a1z−1 −a2z−2...−aNz−N (11.26) which as we saw before, is simply a delay in the numerator, followed by an all-pole expression in the denominator. This is an important result, because it shows that a vocal tract without internal losses can be accurately modelled by an all-pole filter.3 3 Rabiner and Schafer [368] show that the above can be calculated more easily, and without direct matrix multiplication if we assume that the glottal impedance is infinite.
- 361. Section 11.3. Vowel Tube Model 337 It is possible to simplify the calculation by making the further assumption that glottis is a completely closed end, with infinite impedance and reflection coefficient 1. If we take AN(z) = N ∑ k=1 akz−k it is possible to show that in the case that rg = r0 = 1 that the following equation holds AN(z) = AN−1(z)+rkz−k Dk−1(z−1 ) (11.27) That is, the Nth order polynomial in z−1 can be determined from the N-1 order polynomial in z−1, z and the reflection coefficient rk. If we therefore start at N = 0 and set A0(z) = 1 we can calculate the value for A1, and then iteratively for all A(z) up to N. This is considerably quicker than carrying out the matrix multiplications explicitly. The real value in this however, will be shown in Chapter 12, when we consider the relationship between the all-pole tube model and the technique of linear prediction. 11.3.7 The all pole resonator model This general result is important as it shows that for any number of tubes, the result is always an all-pole filter. From Section 10.5.1, we know that this can be factorised into an expression containing individual poles, and these can often be interpreted directly as formant resonances. Here we see we have come full circle; in Section 10.5.3 we showed that a second order IIR filter could faithfully model the shape of a single formant, in Section 10.5.4 we showed that several of these could be cascaded to model a number of formants. In this chapter, we saw in Section 11.2.1 that a physical mass/spring/damped system can be modelled by the same Equations, and that a single mass, spring and damper create a resonance pattern of exactly the same type as a second order IIR formant model. In the preceding section, we have just shown that a connected tube model has an all-pole transfer function, with the number of poles equalling the number of tubes. From this result, it is clear that an all-pole tube model can faithfully create any formant pattern if the parameters to the tubes (i.e. the cross sectional areas) are set correctly. Furthermore, the transfer function for the tube model can be factorised so as to find the poles, and possibly identify individual formants. An important point to realise is that the formant patterns that arise from this all-pole model are a function of the whole model; while the transfer function can be factorised, it is not appro- priate to ascribe individual poles or formants to individual tubes in the model. The situation is more complex than this, and the formant patterns are created from the properties of all the tubes operating together.
- 362. 338 Chapter 11. Acoustic Models of Speech Production open phase return phase closed phase Figure 11.15 Plot of idealised glottal source, as volume velocity over time. The graph shows the open phase, where air is flowing through the glottis, the short return phase, where the vocal folds are snapping shut, and the closed phase, where the glottis is shut, and the volume velocity is zero. The start of the closed phase is the instant of glottal closure. 11.4 SOURCE AND RADIATION MODELS 11.4.1 Radiation So far, we have developed a transfer function V(z) which is defined as the ratio of the volume velocity of the lips over the volume velocity at the glottis. In practice however, when we mea- sure sound, we normally in fact measure pressure signals, as this is what most microphones are designed to respond to. Most microphones exist in the far-field, that is at some distance from the lips, and hence the signal is influenced by a radiation impedance from the lips. This can be modelled by another transfer function, R(z), which, from Equation 11.4 is defined as R(z) = P(z) UL(z) (11.28) Modelling radiation accurately soon becomes very complicated, so for our purposes, we use a very simple model that has been shown to do a reasonable job at approximating the radiation. This takes the form of a function which differentiates the signal, and this can be modelled by a FIR filter with a single zero: R(z) = 1−αz−1 (11.29) where α is a value less than but quite near to 1 (e.g. 0.97). This creates a high frequency boost of about 6dB per octave. 11.4.2 Glottal source We qualitatively described in Section 7.1.2 the process of how the glottis produces sound. In voiced speech, the vocal folds undergo a cycle of movement which gives rise to a periodic source
- 363. Section 11.4. Source and radiation models 339 sound. At the start of this cycle the vocal folds are closed. Pressure is exerted from the lungs which causes a build up in pressure beneath the folds until eventually the pressure forces open the folds. When the folds open, air moves through them and as it does do, the pressure beneath the folds reduces. Eventually the pressure is low enough so that the tension closes the vocal folds, after which the cycle repeats. The rate at which this cycle occurs determines the fundamental frequency of the source, and ranges from about 80Hz to 250Hz for a typical male, and 120Hz to 400Hz for a typical female or child. A plot of volume velocity against time for the glottal source is shown in Figure 11.15. Here we see the three main parts of the cycle: the open-phase where the glottis is opening due to the pressure from the lungs, the return phase where the glottis is closing due to the vocal tract tension, and the closed phase. Modelling the glottis is a tricky problem in that building a simple model seems relatively easy, but developing a highly accurate model that mimics glottal behaviour in all circumstances is extremely difficult and has not in any way been solved. We will return to the difficulties in accurate glottal modelling in Section 13.3.5, but for now we will demonstrate a simple model which at least gives us a flavour of how this works. An extensive study into glottal modelling is given in Flanagan [164], which describes various mass/spring/damped systems. These models can be somewhat difficult to model in discrete time systems, so instead we adopt models which simply generate a time domain function which has the properties described above. One such model [376] [368] is given by: u[n] = 1 2 1−cos(πn N1 0 ≤ n ≤ N1 cos(π(n−N1) 2N2 N1 ≤ n ≤ N2 0 otherwise (11.30) where N1 marks the end of the open phase and N2 marks the end of the return phase/start of the closed phase. Another quite widely used model is the Lijencrants-Fant model [159] (sometimes known as the LF model). This describes the glottal waveform derivative and is given by u[n] = 0 0 ≤ n To E0eα(n−T0)sin[Ωo(n−T0)] T0 ≤ n Te −E1[eβ(n−Te) −eβ(Tc−Te)] Te ≤ n Tc (11.31) where T0 is the instant of glottal opening, Te is the position of the negative minimum, Tc is the instant of glottal closure, α, β and Ω control the shape of the function, and E0 and E1 determine the relative heights of the positive and negative part of the curve. The LF function and its integral, which is the normal glottal flow function, is shown in figure 11.16. Quatieri [364] gives a third model, in which g(n) is created by two time reversed exponentially decaying sequences: u[n] = (β−n u[−n])∗(β−n u[−n]) (11.32) which have the nice property of having a simple z-transform
- 364. 340 Chapter 11. Acoustic Models of Speech Production 0 50 100 150 200 250 300 −15000 −10000 −5000 0 5000 Figure 11.16 Plot Lijencrants-Fant model for glottal flow and derivative glottal flow for a single period. U(z) = 1 (1−βz)2 (11.33) The most important control element in the glottal source is of course the rate at which the cycles occur. In Equation 11.30, this is determined by the positions of N1 and N2; in Equation 11.32 by T0, Te and T0; and in Equation 11.31 this is determined by the impulse function u[n]. In the last case, the glottal volume velocity function can be thought of as a low pass filtering of an impulse stream. From these expressions and empirical measurements, it is known that this low pass filter creates a roll off of about -12dB per octave. It is this, combined with the radiation effect, that gives all speech spectra their characteristic spectral slope. While Equation 11.32 gives a reasonable approximation of the glottal volume velocity signal, it does not model the secondary effects of jitter, shimmer and ripple as mentioned in Section 17. More realism can be added to the glottal signal by the addition of zeros into the transfer function, giving the expression: U(z) = M ∏ k=1 (1−ukz−1) (1−βz)2 (11.34) We have defined U(z) as a volume velocity signal, mainly for purposes of developing the vocal tract transfer function. While in reality, the radiation R(z) occurs after the operation of V(z), we aren’t restricted to this interpretation mathematically. As we shall see, it is often useful to combine U(z) and R(z) into a single expression. The effect of this is to have a system where the radiation characteristic is applied to the glottal flow waveform before it enters the vocal tract.
- 365. Section 11.5. Model refinements 341 This is equivalent to measuring the pressure waveform at the glottis, and if we adopt the radiation characteristic of Equation 11.29, the effect of this is to differentiate the glottal flow waveform. We call this signal the glottal flow derivative, denoted ṡ[n] and Ṡ(z) and and this is shown along with the normal glottal flow in Figure 11.16. The interesting feature of the glottal flow derivative is that it shows that the primary form of excitation into the vocal tract is a large negative impulse. This is useful in that the output of the vocal tract immediately after this should approximate well to the true impulse response of the system. 11.5 MODEL REFINEMENTS 11.5.1 Modelling the nasal Cavity When the velum is lowered during the production of nasals and nasalised vowels, sound enters via the velar gap, propagates through the nasal cavity and radiates through the nose. Hence for a more complete model, we have to add a component for the nasal cavity. This in itself is relatively straightforward to model; for a start, it is a static articulator, so doesn’t have anywhere near the complexity of shapes that occur in the oral cavity. By much the same techniques we employed above, we can construct an all-pole transfer function for the nasal cavity. The complication arises because of the behaviour of the oral cavity. For the case of vowels, the sound wave moved from the glottis through the oral cavity and radiated from the lips. In the case of nasalised vowels, the sound propagates from the glottis and through the pharynx as before, but at the velum some sound enters the nasal cavity while the remainder propagates through the mouth. The sound entering through the velum is filtered by the nasal cavity and radiates through the nose. The sound continuing through the oral cavity radiates through the lips in the manner described above. One complication to this is that the some sound in the oral cavity is reflected backwards and will enter the velar gap, and likewise some sound in the nasal cavity will be reflected backwards and will enter the oral cavity. If we however ignore this interaction for now, the system for nasalised vowels can be modelled as shown in Figure 11.17 with a simple splitting operation at the velum. Because we now have two filters operating in parallel, it is not possible to derive a single transfer function for the whole system. Rather we construct a transfer function for the oral cavity, a different one for the nasal cavity and then simply add the outputs of both to find the final signal. The case of nasal consonants is more complicated. As with nasalised vowels, the sound wave starts at the glottis moves through the pharynx and partly through the mouth. Some sound now travels through the velum, nasal cavity and radiates through the nose. The remainder continues travelling through the oral cavity, but instead of being radiated at the lips, now meets a closure at the lips. Hence a full reflection occurs, and so the sound travels backwards through the mouth, and some continues down towards the glottis and some enters the nasal cavity and radiates from the
- 366. 342 Chapter 11. Acoustic Models of Speech Production P(z) M(z) R(Z) U(z) U(z) L L (a) vowel and approximant P(z) M(z) U(z) U(z) L L N(z) R (Z) N R (Z) (b) nasalised vowel P(z) M(z) U(z) N(z) R (Z) N U (z) L (c) nasal) U (z) L M (z) V (z) B noise source U (z) R (z) L (d) unvoiced obstruent U (z) L M (z) V (z) B R (z) L source glottal G noise source U (z) W U (z) (e) voiced obstruent Figure 11.17 LTI filter models for different types of sound production.
- 367. Section 11.5. Model refinements 343 nose. In addition to this, there are of course the forwards and backwards reflections being caused all the time by the changing cross sectional areas of the nose and mouth. An equivalent way of looking at this is that the pharynx, nasal cavity and nose operate for nasals in the same the way as the pharynx, oral cavity and lips operate for vowels. In addition, there is now a side resonator of the front of the mouth. In phonetic terms the effect of this oral side cavity is very important; as the nasal cavity is static, it is in fact the ability of the tongue to create different oral side cavity configurations which is responsible for the nasal sounds /m/, /n/ and /ng/ all sounding different. Acoustically, the effect of the side branch is to “trap” some of the sound, and this creates anti-resonances. These can be modelled by the inclusion of zeros in the transfer function. As with the case of nasalised vowels, the parallel nature of the system means we can’t use a single transfer function; rather we have a system with an all-pole transfer function for the pharynx and back of the mouth, a splitting operation, an all-pole function for the nose and a pole and zero function for the oral cavity. 11.5.2 Source positions in the oral cavity Many consonant sounds, such as fricatives and stops have a sound source located in the oral cavity. This is created by the tongue nearing another surface (the roof of the mouth, the teeth etc) so as to cause a tight constriction. In the case of unvoiced obstruents, the glottis is open, and air flows through it until it reaches the constriction, at which point turbulence occurs. For voiced obstruents, the glottis behaves as with vowels, but again this air flow is made turbulent by the constriction. The difference in placement of the constriction is what gives rise to the difference between the various fricative and stop sounds of the same voiced. Hence /s/ has an alveolar construction whereas /sh/ has a palatal constriction. It is important to realise however that the constrictions themselves do not generate different sounding sources; it is the fact that the constrictions configure the vocal tract into different shaped cavities through which the source resonates that is responsible for the eventual difference in the sounds. The effect of a sound source in the middle of the vocal tract is to split the source such that some sound travels backwards towards the glottis while the remainder travels forwards towards the lips. The vocal tract is thus effectively split into a backward and forward cavity. The forward cavity acts a tube resonator, similar to the case of vowels but with fewer poles as the cavity is considerably shorter. The backwards cavity also acts as a further resonator. The backwards travelling source will be reflected by the changes in cross sectional area in the back cavity and at the glottis, creating a forward travelling wave which will pass through the constriction. Hence the back cavity has an important role in the determination of the eventual sound. This back cavity acts as a side resonator, just as with the oral cavity in the case of nasals. The effect is to trap sound and create anti- resonances. Hence the back cavity should be modelled with zeros as well as poles in its transfer function.
- 368. 344 Chapter 11. Acoustic Models of Speech Production 11.5.3 Models with Vocal Tract Losses The tube model described in Section 11.3 is termed “lossless” because no energy losses occur during the propagation of sound through the tube itself. Any energy losses which do occur happen through the lips and glottis, where the signal travels outwards but is not reflected back. Several studies have been made into modelling the losses which of course do occur in the tubes themselves. In terms of the discussion given in Section 11.2.3, this means that the impedance now is just composed of an inductive and capacitance component, but also has a resistive component. Rabiner and Schafer [368] give an extensive review of this issue and summarise that there are three main ways in which losses can occur. First, the passage of the wave will cause the walls of the vocal tract (e.g. the cheeks) to vibrate. This can be modelled by refinements to the transfer function. The effect is to raise the formant centre frequencies slightly, dampen the formants and especially the lower formants; we should expect this as it the mass of the vocal tract walls will prevent motion at the higher frequencies. The effects of friction and thermal conduction can also be modelled by adding resistive terms to the transfer function. The overall effect of these is the opposite of above; the formant centre frequencies are somewhat lowered, and the effects of this are more pronounced at higher frequencies. Apart from the overall contribution to damping, these effects tend to cancel each other out, and so the overall contribution of these resistive terms is quite small. As the degree of loss from the lips and glottis is considerably higher, these loss terms are often ignored altogether as they add considerable complication to the model with little gain in modelling accuracy. It should be noted that it is the particular form of the IIR expression that we derived that gives rise to a lossless tube; other IIR filter expressions can easily model losses. 11.5.4 Source and radiation effects Our model of the glottis and lips was extremely simple in that we used special reflection coeffi- cients rG and rL to generate reflections and losses (for values other that |1|). As we might expect, in reality the situation is much more complicated. Firstly, there is no such thing as an “ideal source” where we can simply add volume velocity into a system without any further effect. The source in fact is always to a greater or lesser degree coupled to the rest of the system, and has to be modelled if more accuracy is desired. In general this is modelled by having a real impedance connecting the source to the rest of the system. Again, assuming that sound can just propagate from the lips and “disappear” with no further effect is quite simplistic. A more accurate model can be created by modelling the radiation as a complex impedance connected as part of the system. In this chapter we have concentrated on sound propagation through a tube, and saw that even this, the simplest of all acoustic systems can be quite complicated. Another entire branch of the field of acoustics is devoted to sound propagation through three dimensional spaces, and this can be drawn on to produce a better model of radiation through the lips. As with the other refinements mentioned in this section, modelling both the source and lip losses with more realistic expressions
- 369. Section 11.6. Discussion 345 does increase the accuracy of the overall model, but does so at considerable modelling expense. 11.6 DISCUSSION In developing our model, we have attempted to balance the needs of realism with tractability. The all-pole vocal tract model that described in Section 11.3 will now be adopted for the remainder of the book as the model best suited to our purposes. In subsequent chapters, we shall in fact see that this model has some important properties that make its use particularly attractive. One of the key jobs in making assumptions in a model is to recognise that such assumptions have indeed been made. So long as this is done, and these assumptions are borne in mind, we should not run into too many problems when we rely on the model to deliver faithful analyses. With this in mind, we will now note the main assumptions we have made and discuss any potential problems that might arise because of these. The following list is in approximate increasing order of severity. Linear filter Throughout we have assumed that the system operates as a time invariant linear (LTI) filter of the type described in Chapter 10. While it is well known that there are many non-linear processes present in vocal tract sound propagation, in general the linear model provides a very good approximation to these. Time invariance During normal speech, the vocal tract is constantly moving, so long as the rate of change is very slow compared to the sampling rate, it is quite adequate to model the system as a sequence of piecewise stationary configurations. In Chapter12, we will discuss further the issue of quantising the time dimension so as to produce stationary “snapshots” of speech, but for now let us say that a good approximation to a continuously varying tract is to generate a new stationary vocal tract configuration about every 10ms. Straight tube One of the first assumptions we made was that we could model the vocal tract by a straight tube or sequence of straight tubes. The vocal tract is of course curved; very roughly it starts as a vertical tube at the glottis, takes a 90 degree turn at the top of the pharynx and then continues horizontally through the mouth. In fact it is one of the main principles of acoustic modelling that so long as any bend is not too severe, there is little difference between the propagation in straight or curved tubes. This can be seen clearly in music, where a modern trumpet is actually a quite long (10m) tube which has been wrapped many times to make its playing and carrying more convenient. The bends in the trumpet have little effect on the overall sound. Discrete tube model The number of tubes determines the granularity at which we model the real, continuously varying tube. As we saw in Section 11.3.6, the number of tubes is determined by the sampling rate, and so long as we have the correct number of tubes for the sampling rate being used, there will be no loss of accuracy. A rule of thumb is that we use one tube for every 1kHz sampling rate, so that 10KHZ requires 10 tubes and so on. For a true continuous
- 370. 346 Chapter 11. Acoustic Models of Speech Production signal (equivalent to an infinite sampling rate) we see that we require an infinite number of zero length tubes, ie a continuous tube. For all practical purposes, the discrete nature of the tube has no effect on accuracy. All pole modelling Only vowel and approximant sounds can be modelled with complete accuracy by all-pole transfer functions. We will see in Chapter 12 that the decision on whether to include zeros in the model really depends on the application to which the model is put, and mainly concerns tradeoffs between accuracy and computational tractability. Zeros in transfer functions can in many cases be modelled by the addition of extra poles. The poles can provide a basic model of anti-resonances, but can not model zero effects exactly. The use of poles to model zeros is often justified because the ear is most sensitive to the peak regions in the spectrum (naturally modelled by poles) and less sensitive to the anti-resonance regions. Hence using just poles can often generate the required spectral envelope. One problem however is that as poles are used for purposes other than their natural one (to model resonances) they become harder to interpret physically, and have knock on effects in say determining the number of tubes required, as explained above. Lossless tube The real losses in the vocal tract are slight, but as the difference between a pole being on the unit circle (no loss whatsoever) and slightly within the circle (say at a radius of 0.95) is quite noticeable, then any loss factors will affect the overall system. That said, so long as any ignored losses in the vocal tract are compensated for by increased loss at the lips and glottis, the effect should not be too great. If only the lips and glottis are used for controlling losses, the effect is that a single damping factor is applied to all the formants, and hence it is not possible to generate formants with independent bandwidths. Radiation characteristics The use of a reflection coefficient to model the radiation from the lips is really quite crude. In reality, radiation is very complicated and effected by the exact shape of the lips and head. In our simple example we effectively modelled the region outside the lips as an additional tube - for the complete reflection case this is an infinitely wide tube, that is, the vocal tract opens into a wall that extends infinitely far in all directions. For the cases of less that total reflection, an additional tube of finite area was used. As a real head is actually concave with respect to the direction of radiation, rather than convex in the additional tube case, we see that this in fact is a very crude model. Rabiner and Schafer, and Flanagan develop more sophisticated expressions for lip radiation [368] [164]. An important radiation effect that we have not touched upon is that the radiation determines the directional characteristics of the sound propagation. One commonly observed effect is that high frequencies are more narrowly directed, so that if a speaker is facing away from the listener, the high frequencies are not heard - the speech sounds “muffled”. Source filter separation This is one area where our model starts to break down significantly. Firstly we saw that modelling the glottis as a completely isolated source is quite unrealistic. In addition however, there are many physiological effects that make a complete source/filter separation unrealistic. For example, it is clear that when speaking with a high fundamental
- 371. Section 11.6. Discussion 347 frequency, the entire set of muscles in the throat and mouth are much tenser than in low pitched speech. Glottal source Our model of the source itself is perhaps the crudest and least accurate part of our model. We know for example that the shape of the volume velocity waveform changes with frequency - it is not the case that a fixed low pass filter will do. The model described above also says nothing about vocal effort. It is clear that shouting is not just the same as speaking with a louder source input - the entire characteristic of the glottal source has changed. Furthermore, the dynamics of glottal signals are very complicated. In particular, when real glottal signals are studied in the regions where voicing starts and stops, we see many irregular pitch periods, isolated periods of very short duration, and glottal signals that seem very noisy. The glottal model described in Section 11.4 models none of these affects. A significant amount of research has been conducted into building more sophisticated models, and progress has been made. However, many of these more sophisticated models are difficult to implement and control and hence the simple model explained above is often used. We shall return to this issue in Section 13.3.5, where we shall see that difficulties in modelling source characteristics well are in fact the most difficult problem in speech synthesis, to the extent that they have been one of the main driving factors in determining the path of speech synthesis in recent years. 11.6.1 Further reading Many books deal extensively with modelling the vocal tract acoustically, including Flanagan [164], Rabiner and Schafer [368] and of course Frant [158]. 11.6.2 Summary Model of the Vocal Organs • The vocal apparatus can be modelled as a set of discrete, interconnected components. • Each component functions as either a source component, which generates a basic sound, or a filter component, which modifies this sound. • The filter components are joined in various configurations (Figure 11.17) to form tubes. • For vowels, nasals and approximant, the source component is the glottis. The glottis is located at the end of the vocal tract tube. • Radiation occurs at the lips and nose, and this can also be modelled by a filter. • The filter components are the pharynx, oral cavity and nasal cavity. • All of the above are modelled as lossless tubes. • All of the above act as linear time invariant filters. • Each component tube can be modelled, without loss of accuracy, as a series of connected shorter uniform tubes. The number of uniform tubes required is a function of the required
- 372. 348 Chapter 11. Acoustic Models of Speech Production length of the main tube and the sampling rate. Lossless tube models • The sound wave generated by the source travels along the tube and in so doing is modified. • The tubes are assumed lossless, hence all the impedance terms are reactive, meaning that they only alter the frequency characteristics of the travelling wave, but do not cause energy losses. • The main source of this is reflection, which occurs when the tube changes in cross sectional area. • A tube which has a closed termination and a tube which has a completely open termination cause complete reflection. At a closed termination, the pressure is at a maximum and the volume velocity is zero; at an open termination the pressure is zero and the volume velocity is at a maximum. • At other changes in area, some part of the sound is transmitted and some is reflected back- wards. • The backwards waves interfere with the forwards waves, so that they reinforce at certain frequencies. This causes resonances and hence creates formants. • While each tube in itself is lossless, losses occur in the system because some sound escapes from the lips and glottis. Transfer function • By iterative application of the equations relating volume velocity in one tube to volume velocity in the next, we can determine a transfer function for the whole tube. • It is shown (Equation 11.26) that all such models produce a transfer function containing only poles (i.e. a polynomial expression in z−1 in the denominator). • We know, from Section 10.5.4, that such transfer functions naturally produce resonance patterns and hence formants. • The polynomial expression can be factorised to find the individual poles. • The values of the coefficients in the denominator polynomial are determined by the set of reflection coefficients in the tube. There is no simple relationship between the reflection coefficients and the poles, but they can be calculated if need be. • The bandwidths of the resonances are determined by the losses in the system, and as these only occur at the lips and glottis (or even just the lips or just the glottis), there is only one or two parameters in the system controlling all of these. • Modelling nasals and obstruents is more complicated because of the inclusion of side branches in the model. These “trap” the sound, and are best modelled with filters which also contain zeros.
- 373. Section 11.6. Discussion 349 • As nasals and obstruents involved parallel expressions, they can not be expressed as a single transfer function.
- 374. 12 ANALYSIS OF SPEECH SIGNALS In this chapter we turn to the topic of speech analysis, which tackles the problem of deriving rep- resentations from recordings of real speech signals. This book is of course concerned with speech synthesis - and at first sight it may seem that the techniques for generating speech “bottom-up” as described in Chapters 10 and 11 may be sufficient for our purpose. As we shall see however, many techniques in speech synthesis actually rely on an analysis phase which captures key properties of real speech and then uses these to generate new speech signals. In addition, the various tech- niques here enable useful characterisation of real speech phenomena for purposes of visualisation or statistical analysis. Speech analysis then is the process of converting a speech signal to an alter- native representation that in some way better represents the information which we are interested in. We need to perform analysis because waveforms do not usually directly gives us the type of information we are interested in. Nearly all speech analysis is concerned with a three key problems. First, we wish to remove the influence of phase; second, we wish to perform source/filter separation, so that we can study the spectral envelope of sounds independent of the source that they are spoken with. Finally we often wish to transform these spectral envelopes and source signals into other representations, that are coded more efficiently, have certain robustness properties, or which more clearly show the linguistic information we require. All speech signals in the real world are continuous signals which describe the pattern of air pressure variation over time. These signals can be recorded with a variety of analogue means, but for computer analysis, we require our signals to be digitised such that the continuous signal is converted to a discrete signal. For our needs, we need not go into the details of analogue sound capture and recording as it is most likely that any reader will either come across signals that are already digitised, or that they will be able to “record” any acoustic signal directly into a digitised form by using a standard computer sound card, without the need for any interim analogue recording step1. 1 Many standard books explain the analogue to digital conversion process, for instance see Rabiner and Schafer [368]. 350
- 375. Section 12.1. Short term speech analysis 351 12.1 SHORT TERM SPEECH ANALYSIS A common starting point in speech analysis is to find the magnitude spectrum from a speech signal. We want this to be in a discrete form that is easy to calculate and store store in a computer, so we use the discrete Fourier transform (DFT) as our principle algorithm. In practice the fast Fourier transform (FFT) implementation is nearly always used due to its speed (see Section 10.2.4). 12.1.1 Windowing When we speak, the glottis and vocal tract are constantly changing. This is problematic for most of the techniques we introduced in Chapter 10 as these were designed to work on stationary signals. We can get around this problem by assuming that the speech signal is in fact stationary if consid- ered over a sufficiently short period of time. Therefore we model a complete speech waveform as a series of short term frames of speech, each of which we consider as a stationary time-invariant system. A frame of speech x[n] is obtained from the full waveform s[n] by multiplication by a window w[n] in the time domain: x[n] = w[n]s[n] (12.1) Three common windows are the rectangular window, hanning window and hamming window: rectangular w[n] = ( 1 0 ≤ n ≤ L−1 0 otherwise (12.2) hanning w[n] = ( 0.5−0.5cos(2πn L ) 0 ≤ n ≤ L−1 0 otherwise (12.3) hamming w[n] = ( 0.54−0.46cos(2πn L ) 0 ≤ n ≤ L−1 0 otherwise (12.4) Figure 12.1 shows a sinusoid after windowing by the three windows just described. The rectangu- lar window abruptly cuts off the signal at the window edges, while the hanning tapers the signal so that the discontinuities are absent. The hamming window is just a raised version of the hanning, and so has a similar shape accept for a slight discontinuity at the window edge. While an essential part of speech analysis, the windowing process unfortunately introduces some unwanted effects in the frequency domain. These are easily demonstrated by calculating the DFT of the sinusoid. The “true” frequency domain representation for this, calculated in continuous time with a Fourier transform, is a single delta peak at the frequency of the sinusoid. Figures 12.2 show the DFT and log DFT for the three window types. The DFT of the rectangular window
- 376. 352 Chapter 12. Analysis of Speech Signals -1 -0.5 0 0.5 1 0 100 200 300 400 500 ampl i t ude samples (a) Rectangular window -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 0 100 200 300 400 500 ampl i t ude samples (b) Hanning window -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 0 100 200 300 400 500 ampl i t ude samples (c) Hamming window Figure 12.1 Effect of windowing in the time domain shows a spike at the sinusoid frequency, but also shows prominent energy on either side. This effect is much reduced in the hanning window. In the log version of this, it is clear there there is a wider main lobe than in the rectangular case - this is the key feature as it means that more energy is allowed to pass through at this point in comparison with the neighbouring frequencies. The hamming window is very similar to the hanning, but has the additional advantage that the side lobes immediately neighbouring the main lobe are more suppressed. Figure 12.3 shows the effect of the windowing operation on a square wave signal and shows that the harmonics are preserved far more clearly in the case where the Hamming window is used. 12.1.2 Short term spectral representations Using windowing followed by a DFT, we can generate short term spectra from speech wave- forms. The DFT spectrum is complex and and can be represented by its real and imaginary parts or its magnitude and phase parts. As explained in section 10.1.5 the ear is not sensitive to phase information in speech, and so the magnitude spectrum is the most suitable frequency domain rep- resentation. The ear interprets sound amplitude in an approximately logarithmic fashion - so a doubling in sound only produces an additive increase in perceived loudness. Because of this, it is usual to represent amplitude logarithmically, most commonly on the decibel scale. By convention, we normally look at the log power spectrum, that is the log of the square of the magnitude spec- trum. These operations produce a representation of the spectrum which attempts to match human perception: because phase is not perceived, we use the power spectrum, and because our response to signal level is logarithmic we use the log power spectrum.
- 377. Section 12.1. Short term speech analysis 353 0 20 40 60 80 100 120 140 0 500 1000 1500 2000 2500 3000 3500 4000 log magnitude frequency (Hz) (a) Rectangular window -300 -250 -200 -150 -100 -50 0 50 0 500 1000 1500 2000 2500 3000 3500 4000 log magnitude frequency (Hz) (b) Rectangular window 0 10 20 30 40 50 60 70 0 500 1000 1500 2000 2500 3000 3500 4000 log magnitude frequency (Hz) (c) Hanning window -300 -250 -200 -150 -100 -50 0 50 0 500 1000 1500 2000 2500 3000 3500 4000 log magnitude frequency (Hz) (d) Hanning window 0 10 20 30 40 50 60 70 0 500 1000 1500 2000 2500 3000 3500 4000 log magnitude frequency (Hz) (e) Hamming window -300 -250 -200 -150 -100 -50 0 50 0 500 1000 1500 2000 2500 3000 3500 4000 log magnitude frequency (Hz) (f) Hamming window Figure 12.2 Effect of windowing in the frequency domain, with magnitude spectra on the left hand side and log magnitude spectra on the right hand side. Figures 12.4 and 12.5 show some typical power spectra for voiced and unvoiced speech. For the voiced sounds, we can clearly see the harmonics as the series of spikes. They are evenly spaced, and the fundamental frequency of the the source, the glottis, can be estimated by taking the reciprocal of the distance between the harmonics. In addition, the formants can clearly be seen as the more general peaks in the spectrum. The general shape of the spectrum ignoring the harmonics is often referred to as the envelope.
- 378. 354 Chapter 12. Analysis of Speech Signals 0 20 40 60 80 100 120 140 160 0 1000 2000 3000 4000 5000 6000 7000 8000 log magnitude frequency (Hz) (a) Magnitude spectrum Hamming window -3 -2 -1 0 1 2 3 4 5 6 0 1000 2000 3000 4000 5000 6000 7000 8000 log magnitude frequency (Hz) (b) Log magnitude spectrum Hamming window 0 50 100 150 200 250 300 350 0 1000 2000 3000 4000 5000 6000 7000 8000 log magnitude frequency (Hz) (c) Magnitude spectrum Rectangular window -3 -2 -1 0 1 2 3 4 5 6 0 1000 2000 3000 4000 5000 6000 7000 8000 log magnitude frequency (Hz) (d) Log magnitude spectrum Rectangular window Figure 12.3 Comparison of rectangular and hamming window on a square wave of fundamental frequency 400Hz and DFT length 512. Both spectra show significant deviation from the ideal of a single spike for each harmonic, but it is clear that the general pattern is that of a square wave, ie only odd harmonics which are exponentially decaying. More distortion occurs for the weaker harmonics as their own energy is swamped by the side lobes of the other harmonics. This effect is much worse for the rectangular window case for which the higher harmonics are barely discernible. 12.1.3 Frame lengths and shifts While the spectra of the hamming and hanning windowed sinusoid give a truer frequency repre- sentation than the rectangular window, the spike has a noticeable width, and is not the ideal delta function that the continuous Fourier transform predicts. The width of the spike is an inverse func- tion of window length - as the window lengthens the spike becomes narrower, as demonstrated in Figure 12.6. This of course is predicted by the scaling property (see Section 10.3.3). While longer windows give produce a more accurate spectrum, the amount of speech required to generate a (say) 2096 length window is such that the frame-based stationary assumption will have been violated, as the vocal tract shape and glottal period may have changed substantially within this time. Hence we encounter the problem referred to as the time-frequency trade-off in speech. The speech spectrum is constantly varying as we move from saying one sound to another. In fact it is a rare case when two consecutive frames of speech share exactly the same frequency characteristics. From the definition of the DFT we know that to achieve a high frequency resolution we need a
- 379. Section 12.1. Short term speech analysis 355 -20000 -15000 -10000 -5000 0 5000 10000 15000 20000 25000 0 100 200 300 400 500 ampl i t ude samples (a) Rectangular window -20000 -15000 -10000 -5000 0 5000 10000 15000 20000 0 100 200 300 400 500 ampl i t ude samples (b) Hanning window 0 50000 100000 150000 200000 250000 300000 350000 400000 450000 500000 0 1000 2000 3000 4000 5000 6000 7000 8000 log magnitude frequency (Hz) (c) Magnitude spectrum 0 2 4 6 8 10 12 14 0 1000 2000 3000 4000 5000 6000 7000 8000 log magnitude frequency (Hz) (d) Log Magnitude Spectrum Figure 12.4 Signal and spectra for /ih/ vowel with N=32 -5000 -4000 -3000 -2000 -1000 0 1000 2000 3000 4000 0 100 200 300 400 500 ampl i t ude samples (a) Rectangular window -4000 -3000 -2000 -1000 0 1000 2000 3000 4000 0 100 200 300 400 500 ampl i t ude samples (b) Hanning window 0 20000 40000 60000 80000 100000 120000 0 1000 2000 3000 4000 5000 6000 7000 8000 log magnitude frequency (Hz) (c) Magnitude spectrum 0 2 4 6 8 10 12 0 1000 2000 3000 4000 5000 6000 7000 8000 log magnitude frequency (Hz) (d) Log Magnitude Spectrum Figure 12.5 Signal and spectra for /s/ vowel with N=32
- 380. 356 Chapter 12. Analysis of Speech Signals 0 20 40 60 80 100 120 140 0 500 1000 1500 2000 2500 3000 3500 4000 l o g magni t ude frequency (Hz) (a) 512 length window 0 50 100 150 200 250 300 0 500 1000 1500 2000 2500 3000 3500 4000 l o g magni t ude frequency (Hz) (b) 1024 length window 0 100 200 300 400 500 600 0 500 1000 1500 2000 2500 3000 3500 4000 l o g magni t ude frequency (Hz) (c) 2048 length window Figure 12.6 Effect of frame length of width of spike large number of waveform samples. But as the frame length increases, we include speech from different vocal tract or glottis configurations. For example if our window is very long we might have several different phonemes in the window, all with completely different characteristics. If on the other hand, we use a very short window, in which we are sure to be capturing only a single speech sound, we will only have a few terms from the DFT resulting in too course a representation of the spectrum. Hence there is a tradeoff. If we require information about the fundamental frequency of the signal, it is important that the frame length be long enough to include one or more periods of speech. Figure 12.4 shows waveforms and spectra for a signal for a variety of window lengths. In addition to the length of window, we also have to consider how often we should calculate the spectrum. The frame shift is defined as the distance between two consecutive frames. Again, there is no single correct value for this. In some applications (e.g. when calculating a spectrogram, described below) a very short frame shift (e.g. 1ms) is desirable as we wish to represent every possible transient effect in the evolution of the spectrum. For other uses, this is considered overkill, and as spectra take time to calculate and space to store, a longer frame shift of about 10ms is often used. In general we decide frame shifts based on the rate of evolution of the vocal tract. A good rule of thumb is that the vocal tract can be considered stationary for the duration of a single period, and so values in this range are often chosen. Typical values of frame length and frame shift are about 25ms and 10ms respectively. These can be expressed in numbers of samples by multiplication by the sample rate being used, and frame lengths are often chosen to be powers of two (128, 256 etc) because of the use of the FFT. In speech synthesis, it is common to perform pitch-synchronous analysis, where the frame
- 381. Section 12.1. Short term speech analysis 357 -20000 -15000 -10000 -5000 0 5000 10000 15000 0 100 200 300 400 500 ampl i t ude samples (a) 64 sample frame 3 4 5 6 7 8 9 10 11 0 1000 2000 3000 4000 5000 6000 7000 8000 log magnitude frequency (Hz) (b) log magnitude spectrum of 32 sample frame -20000 -15000 -10000 -5000 0 5000 10000 15000 20000 0 100 200 300 400 500 ampl i t ude samples (c) 128 sample frame -2 0 2 4 6 8 10 12 14 0 1000 2000 3000 4000 5000 6000 7000 8000 log magnitude frequency (Hz) (d) Log Magnitude Spectrum -20000 -15000 -10000 -5000 0 5000 10000 15000 20000 25000 0 100 200 300 400 500 ampl i t ude samples (e) 256 sample frame 0 2 4 6 8 10 12 14 0 1000 2000 3000 4000 5000 6000 7000 8000 log magnitude frequency (Hz) (f) Log Magnitude Spectrum -20000 -15000 -10000 -5000 0 5000 10000 15000 20000 25000 0 100 200 300 400 500 ampl i t ude samples (g) 512 sample frame 0 2 4 6 8 10 12 14 0 1000 2000 3000 4000 5000 6000 7000 8000 log magnitude frequency (Hz) (h) Log Magnitude Spectrum -20000 -15000 -10000 -5000 0 5000 10000 15000 20000 25000 0 100 200 300 400 500 ampl i t ude samples (i) 1024 sample frame 0 2 4 6 8 10 12 14 0 1000 2000 3000 4000 5000 6000 7000 8000 log magnitude frequency (Hz) (j) Log Magnitude Spectrum Figure 12.7 Signal and spectra for a vowel with different window lengths. For the shorter frames, only a course spectrum can be discerned. The harmonics are clearest for 512 length case; above this the fundamental frequency changes across the window mean that there is no single harmonic spacing.
- 382. 358 Chapter 12. Analysis of Speech Signals Figure 12.8 Wide band spectrogram Figure 12.9 Narrow band spectrogram is centred around the pitch period. As pitch is generally changing, this makes the frame shift variable. Pitch-synchronous analysis has the advantage that each frame represents the output of the same process, that is, the excitation of the vocal tract with a glottal pulse. In unvoiced sections, the frame rate is calculated at even intervals. Of course, for pitch-synchronous analysis to work, we must know where the pitch periods actually are: this is not a trivial problem and will be addressed further in section 12.7.2. Note that fixed frame shift analysis is sometimes referred to as pitch-asynchronous analysis. Figure 12.3 shows the waveform and power spectrum for 5 different window lengths. To some extent, all capture the envelope of the spectrum. For window lengths of less than one period, it is impossible to resolve the fundamental frequency and so no harmonics are present. As the win- dow length increases, the harmonics can clearly be seen. At the longest window length, we have a very good frequency resolution, but because so much time-domain waveform is analysed, the position of the vocal tract and pitch have changed over the analysis window, leaving the harmonics and envelope to represent an average over this time rather than a single snapshot.
- 383. Section 12.1. Short term speech analysis 359 12.1.4 The spectrogram A spectrum only shows the frequency characteristics for a single frame at a single point in time - often this is referred to as a spectral slice. We often wish to observe the evolution of the spectrum over time - this can be displayed as a spectrogram. A spectrogram is a 3-dimensional plot as shown in Figure 12.8. The horizontal axis denotes time, the vertical axis frequency, and the level of darkness represents amplitude - the darker the plot the higher the amplitude as this point. The spectrogram is formed by calculating the DFT at successive points along the waveform. For a high definition spectrogram, the frame shift is normally small, say 1ms. The choice of window length is important to the appearance of the spectrogram - Figure 12.9 shows a narrow band spectrogram which was calculated with a DFT length of 400; Figure 12.8 shows a wide band spectrogram calculated with a DFT length of 40. The narrow-band spectrogram has a detailed spectrum, and so the harmonics can clearly be seen as horizontal stripes in the voiced sounds. It should be clear that the distance between adjacent stripes for a particular time value is constant, as we would expect of harmonics. In fact an estimate of the fundamental frequency can be taken by measuring the inter-harmonic distance. The wide-band spectrogram has a very short DFT and so the frequency information is course and the harmonics are not discernible. The formants however, clearly are - these are seen as the wide horizontal stripes. As the window length is short, the time-resolution is high, and as such several frames are needed to cover a single period. The effect of this can be seen in the fine vertical streaks: frames which contain a glottal impulse will be dark, those which do not will be lighter. The distance between successive dark streaks can be used to gain an estimate of the pitch. As the frames are much longer in the narrow-band case, it is not possible to discern pitch this way. The narrow-band and wide band spectrograms clearly demonstrate the time-frequency tradeoff. In general, the wide band spectrogram is used for finding information about phones, as it shows the formant structure more clearly. Most speech analysis computer packages have spectrogram display software, and from these it is easy to visualise and examine speech. The reader should be made aware however that many of these packages perform a degree of graphical post-processing, which brings out the salient features in the spectrogram. This is often achieved by setting a saturation levels below which specify cut offs beyond which everything is either white or black. It is important to be aware of these post processing steps because they can give rise to different looking spectrograms, and if caught unaware, one might mistakenly conclude that two speech signals are more dissimilar than they really are. 12.1.5 Auditory scales It is well known that human sensitivity to the frequency scale is not linear; for instance we know that musical relationships are clearly logarithmic. Studies into the low-level perception of sounds have resulted in a number of auditory scales which define a new frequency range that is more in
- 384. 360 Chapter 12. Analysis of Speech Signals 0 50 100 150 200 250 300 −12 −10 −8 −6 −4 −2 0 2 (a) Filter bank analysis with evenly spaced bins 0 50 100 150 200 250 300 −12 −10 −8 −6 −4 −2 0 2 (b) Filter bank analysis with bins spaced accord- ing to the mel-scale Figure 12.10 Filter bank analysis on magnitude spectra line with the human sensitivity to sounds at different frequencies. The mel-scale was the product of experiments with sinusoids in which subjects were re- quired to divide frequency ranges into sections. From this, a new scale was defined in which one mel equalled one thousandth of the pitch of a 1KHz tone [415]. The mapping from linear frequency to this scale is given by: m = 2595log10(1+ f/700) (12.5) Further studies based on loudness (i.e. the percept of amplitude) found that a more accurate representation was one that had a linear response for lower frequencies, becoming logarithmic for higher frequencies. One popular scale based on this is the Bark-scale [522]: B = 13arctan(0.76f/1000)+3.5arctan((f/7500)2 ) (12.6) A third popular scale is the equivalent rectangular bandwidth or ERB scale, which is measured from the ability to detect sinusoids in the presence of noise. From this auditory percep- tion is described in terms of an ”equivalent rectangular bandwidth” (ERB) as a function of centre frequency [316]. B = 6.23×10−6 f2 +9.33910−2 f +28.52 (12.7) 12.2 FILTER BANK ANALYSIS We will now turn to the important problem of source-filter separation. In general, we wish to do this because the two components of the speech signal have quite different and independent linguis- tic functions. The source controls the pitch, which is the acoustic correlate of intonation, while the filter controls the spectral envelope and formant positions, which determine which phones are be- ing produced. There are three popular techniques for performing source-filter separation. First we
- 385. Section 12.3. The Cepstrum 361 will examine filter bank analysis in this section, before turning to cepstral analysis and linear prediction in the next sections. If we visually examine a short length DFT such as that in Figure 12.4, or in fact any spec- trogram, we can see the general shape of the spectral envelope. That is, if we “squint” and ignore the harmonics, in many cases the shape of spectral envelope is discernible independent of the par- ticular positions of the harmonics. When doing this we are effectively averaging out the influence of individual harmonics and creating a blurring effect and this is the principle behind filter bank analysis. Filter bank analysis is performed by first creating a series of bins each centred on a particular frequency. Time domain filter-bank analysis can be performed by creating a band pass filter (either FIR or IIR) which allows frequencies at or near the bin frequency to pass, but attenuates all other frequencies to zero. After the filtering, the amount of energy is calculated and assigned to that bin. Once this has been done for every bin, we have a representation of how energy varies according to frequency. The effect of the filtering operation is to blur the effect of individual harmonics, such that the final representation is largely that of the vocal tract and not the source. Alternatively, we can perform filter bank analysis on spectra. Here, we take a magnitude spectra and multiply it by a windowing function centred on a bin frequency. This has the effect of setting all energy outside the window to zero, while allowing us to measure the energy inside the window. This is show in Figure 12.10a. Often the bins are not spaced at equal frequencies, but defined logarithmically or according to one of the frequency scales of Section 12.1.5. This is shown in Figure 12.10b. Filter bank analysis is a simple and robust technique for finding the spectral envelope, but it is only a partial solution to source filter separation. First, the amount of “blurring” required to eliminate the harmonics may be too severe and may eliminate some of the detail we require; alternatively, and especially with high pitched speech, we can find that peaks in the envelope fall between two harmonics and therefore are ignored. Secondly, filter bank analysis only gives us the spectral envelope derived from the vocal tract transfer function; it does not give us the source. For many applications (e.g. ASR) this is not a problem, but in TTS we generally require both components of the signal, and so other techniques are normally used. 12.3 THE CEPSTRUM 12.3.1 Cepstrum definition The cepstrum is (most commonly) defined as the inverse DFT of the log magnitude of the DFT of a signal: c[n] = F −1 log| F { x[n] } | (12.8)
- 386. 362 Chapter 12. Analysis of Speech Signals 0 20 40 60 80 100 120 0 1000 2000 3000 4000 5000 6000 7000 8000 amplitude normalise frequency (a) Magnitude spectrum and envelope 0 2 4 6 8 10 12 14 0 1000 2000 3000 4000 5000 6000 7000 8000 amplitude normalise frequency (b) log Magnitude spectrum and envelope -300 -200 -100 0 100 200 300 400 500 600 700 0 50 100 150 200 250 amplitude samples (c) Cepstrum Figure 12.11 Steps involved in calculating the cepstrum. For demonstration purposes, an estimate of the spectral envelope has been overlaid on the two DFT spectra. The “periodicity” of the harmonics in the log spectrum can clearly be seen. this gives rise to the spike at point 120 in the cepstrum. The low values (30) in the cepstrum describe the envelope. It should be clear that these low values are separated from the spike - because of this it is a simple task to separate the source and filter. where F is the DFT and F −1 is the inverse DFT. For a windowed frame of speech y[n] the cepstrum is therefore c[n] = N−1 ∑ n=0 log N−1 ∑ n=0 x[n]e− j 2π N kn ej 2π N kn (12.9) Figure 12.11 shows the process of calculating the DFT, log, and inverse DFT on a single frame of speech. We will now look at how this operation performs source/filter separation. 12.3.2 Treating the magnitude spectrum as a signal Consider the magnitude spectrum of a periodic signal, such as that shown in Figure 12.11. As we have proved and seen in practice, this spectrum will contain harmonics at evenly spaced intervals. Because of the windowing effects, these harmonics will not be delta function spikes, but will be somewhat more “rounded”. For most signals, the amplitude of the harmonics tails off quite quickly
- 387. Section 12.3. The Cepstrum 363 as frequency increases. We can reduce this effect by compressing the spectrum with respect to amplitude; this can easily be achieved by calculating the log spectrum. It should be clear in the many examples of log spectra in this chapter that the relative amplitudes of the harmonics are much closer in the log spectra than the absolute spectra. Figure 12.5b shows this effect. Now, forget for a moment what the log spectrum actually represents, and consider it a wave- form instead. In we were told Figure 12.11b was a waveform, we would informally describe this as some sort of quasi-periodic signal (after it, it has a repeating pattern of harmonics), with a time evolving “amplitude effect” which makes some “periods” higher than others. It is clear that the rate at which the “periods” change is much greater than the evolution of the “amplitude effect”. If we wished to separate these, we would naturally turn to our standard tool for analysing the frequency content of waveforms, namely the Fourier transform. If we calculate the DFT of our “waveform” we get a new representation, which we will designate the cepstrum. This is shown in Figure 12.11c. As we would expect, the periodicity of the “waveform” is shown by a large spike at about position 120. Because we haven’t performed any windowing operations which cause blurring, we would expect this to be close to a delta function. Again, as we would expect, because the “periods” were not sinusoidal, this has harmonics at multiples of the main spike. The “amplitude effect” will be present in the cepstrum also, but as this is varying much more slowly than the “periods”, it will be in the lower range of the cepstrum. As the “amplitude effect” is the spectral envelope and the spikes represent the harmonics (ie the pitch), we see that these operations have produced a representation of the original signal in which the two components lie at different positions. A simple filter can now separate them. 12.3.3 Cepstral analysis as deconvolution Recall from Section 11.1 that Equation 11.2 described the speech production process as y[n] = u[n]⊗v[n]⊗r[n] where where u[n] is the glottal or other source, v[n] is the vocal tract filter and r[n] is the radiation. In the time domain, these expressions are convolved and so are not easily separable. If we don’t know the form of the vocal tract filter, we can’t perform any explicit inverse filtering operation to separate these components. We can however perform separation as follows. Let us subsume the radiation filter r[n] into v[n] to give a single filter v′[n]. We can then write the time domain expression of Equation 11.2 as y[n] = u[n]⊗v′ [n] (12.10) Taking the Fourier transform of Equation 12.10 gives Y(ejω ) = U(ejω )V′ (ejω )
- 388. 364 Chapter 12. Analysis of Speech Signals -6 -4 -2 0 2 4 6 8 0 1000 2000 3000 4000 5000 6000 7000 8000 log magnitude frequency (Hz) Figure 12.12 Series of spectra calculated from cepstra. The 512 point cepstrum has all its values after a cut off point K set to 0 and then a 512 point DFT is applied to generate a spectrum. The figure shows this spectrum for 4 values of K, 10, 20, 30, 50. (In the figure each has been graphically separated for clarity, with K = 10 as the bottom spectrum and K = 50 as the top. In reality all lie in the same amplitude range.) As the number of coefficients increases, so does the level of spectral detail, until effects of harmonics start appearing. The choice of optimal cut-off point depends on whether spectral precision or elimination of harmonics is the top priority. If we now take the log of the magnitude of this we get log(|Y(ejω )|) = log(|U(ejω |)V′ (|ejω )|) = log(|U(ejω )|)+log(|V′ (ejω )|) in which the source spectrum and filter spectrum are now just added together. We can now return to the time domain by taking an inverse DFT, giving us c[n] = cu[n]+cv[n] (12.11) that is, a time domain representation created by addition of source and filter. Unlike Equation 12.10, the source and filter are now added and not convoluted in the time domain. 12.3.4 Cepstral analysis discussion If we take the DFT of a signal and then take the inverse DFT of that, we of course get back to where we started. The cepstrum calculation differs in two important respects. First we are only using the magnitude of the spectrum - in effect we are throwing away the phase information. The inverse DFT of a magnitude spectrum is already very different from the inverse DFT of a normal (complex) spectrum. The log operation scales the harmonics which emphasises the “periodicity” of the harmonics. It also ensures that the cepstrum is the sum of the source and filter components and not their convolution. The cepstrum is useful as it splits the signal into an envelope, given by the first few coeffi- cients, and a source, given by the spike. Subsequent analysis usually throws away one of these
- 389. Section 12.4. Linear prediction analysis 365 parts: if we are interested in the vocal tract we use only the low coefficients, if we are interested in the pitch and behaviour of the glottis we study the peak. We can demonstrate this by calculating a spectrum in which the spike is eliminated - this is done by simply setting a cut off point K above which all the cepstral coefficients are set to zero. From this, we can create a log magnitude spec- trum by applying a DFT on the modified cepstrum. Figure 12.12 shows this for 4 different values of K. The figure shows that a reasonably accurate spectral envelope can be generated from about 30 coefficients, but the number chosen really depends on the degree of precision required in the spectrum. For many cases where the spectral envelope is required, it is advantageous to keep the low part of the cepstrum as it is, and not convert it back to a frequency domain spectrum. The low cepstral coefficients form a very compact representation of the envelope, and have the highly de- sirable statistical modelling property of being independent so that only their means and variances, and not their covariances, need be stored to provide an accurate statistical distribution. It is for this reason that cepstral are the feature representation of choice in most speech recognition systems. Despite this, cepstra are quite difficult to visually interpret, so we often find that while systems use cepstra internally, we still use spectra when we wish to visually inspect a signal. The higher part of the cepstrum contains the pitch information, and is often used as the basis for pitch detection algorithms, explained in section 12.7.1 12.4 LINEAR PREDICTION ANALYSIS Linear Prediction (LP) is another technique for source filter separation, in which we use the techniques of LTI filters to perform an explicit separation of source and filter. In LP we adopt a simple system for speech production, where we have an input source x[n], which passed through a linear time invariant filter h[n], to give the output speech signal y[n]. In the time domain this is: y[n] = x[n]⊗h[n] (12.12) The convolution makes it difficult to separate the source and filter in the time domain, but this is easily achieved if we transform our system to the z-domain H(z) = X(z) Y(z) (12.13) In Section 12.12, we showed that the lossless tube model was a reasonable approximation for the vocal tract during the production of a vowel. If we assume for now that H(z) can therefore be represented by an all-pole filter, we can write Y(z) = X(z)H(z) (12.14) = 1 1− P ∑ k=1 akz−k .X(z) (12.15)
- 390. 366 Chapter 12. Analysis of Speech Signals where {ak} are the filter coefficients. By taking the inverse z-transform2 we see that in the time domain this becomes: y[n] = P ∑ k=1 aky[n−k]+x[n] (12.16) The importance of this equation is that it states that the value of the signal y at time n can be determined by a linear function of the previous values of y, apart from the addition of the current value of the input x[n]. Because of this, the technique is known as linear prediction. In essence, this states that an output signal is closely approximated by a linear sum of the previous samples: ỹ[n] = P ∑ k=1 aky[n−k] (12.17) and this is identical to the difference Equation 12.16 except for the input term. In linear pred- ication, the input x[n] is usually referred to as the error, and is often written e[n] to show this. This may seem a little harsh, but the concept comes from the fact that y[n] is the actual signal and ˜ y[n] is its approximation by the model, the only difference being the contribution from x[n]. As this accounts for the difference between the predicted and actual signal it is called the error. We now turn to the issue of how to automatically find the predictor coefficients of H(z) for a frame of speech. 12.4.1 Finding the coefficients: the covariance method The goal of linear prediction is to find the set of coefficients {ak} which generate the smallest error e[n], over the course of the frame of speech. We will now examine two closely related techniques for finding the ak coefficients based on error minimisation. We can sate equation 12.16 in terms of the error for a single sample: e[n] = y[n]− P ∑ k=1 aky[n−k] (12.18) 2 We can most easily prove this in reverse; if y[n] = P ∑ k=1 aky[n−k]+x[n] then taking the z-transform of both sides results in Y(z) = n=∞ ∑ n=−∞ P ∑ k=1 aky[n−k] z−n +E(z) = P ∑ k=1 ak n=∞ ∑ n=−∞ y[n−k]z−n +E(z) Using the dummy substitution m = n−k gives = P ∑ k=1 ak m=∞ ∑ m=−∞ y[m]z−m z−k +E(z) = P ∑ k=1 akY(z)z−k +E(z) rearranging the Y(z) terms gives the all-pole transfer func- tion of Equation 12.15.
- 391. Section 12.4. Linear prediction analysis 367 We wish to find the values {ak} which minimise the squared error, summed over the frame of speech. Using E to denote the sum squared error we can state: E = N−1 ∑ n=0 e2 [n] = N−1 ∑ n=0 y[n]− P ∑ k=1 aky[n−k] !2 We can find the minimum of E by differentiating with respect to each coefficient ak, and setting to zero: δE δaj = 0 = N−1 ∑ n=0 2 y[n]− P ∑ k=1 aky[n−k] y[n− j] ! for j = 1,2,3...P = −2 N−1 ∑ n=0 y[n]y[n− j]+2 N−1 ∑ n=0 P ∑ k=1 aky[n−k]y[n− j] which gives N−1 ∑ n=0 y[n− j]y[n] = P ∑ k=1 ak N−1 ∑ n=0 y[n− j]y[n−k] for j = 1,2,3...P (12.19) If we define φ(k, j) as φ(j,k) = N−1 ∑ n=0 y[n− j]y[n−k] (12.20) Equations 12.19 can be written more succinctly as: φ(j,0) = P ∑ k=1 φ(j,k)ak (12.21) the full form of which is φ(1,0) = a1φ(1,1)+a2φ(1,2)+...+aPφ(1,P) φ(2,0) = a1φ(2,1)+a2φ(2,2)+...+aPφ(2,P) φ(3,0) = a1φ(3,1)+a2φ(3,2)+...+aPφ(3,P) . . . φ(P,0) = a1φ(P,1)+a2φ(P,2)+...+aPφ(P,P) We can easily calculate all the quantities φ(1,0),φ(2,0), and so it should be clear that this is a set of P equations with P unknown terms, namely the ak coefficients. This can be written in matrix form as
- 392. 368 Chapter 12. Analysis of Speech Signals φ(1,0) φ(2,0) φ(3,0) . . . φ(P,0) = a1 a2 a3 . . . aP φ(1,1) φ(1,2) ··· φ(1,P) φ(2,1) φ(2,2) ··· φ(2,P) φ(3,1) φ(3,2) ··· φ(3,P) . . . φ(P,1) φ(P,2) ··· φ(P,P) (12.22) and expressed as φa = ψ We can find the value of a (that is, the coefficients ak) by solving this equation. This can be done simply, by solving for a1 in terms of the other coefficients and substituting again, but this is extremely inefficient. Because the square matrix φ is symmetric a much faster technique known as Cholskey decomposition can be used [302] [295]. This method is called the covariance method (for somewhat obscure reasons). It works by calculating the error in the region of samples from n = 0 to n = N −1, but in fact uses samples in he region n = −P to n = N −1 to do so. This is because the method requires a full calculation of the error at time n = 0, and to do so, it must use some previous samples to do so. The calculation over many frames can be found by either simply using the main signal itself, and moving the start position n to the designated frame start each time; or by creating a separate windowed frame of speech, of length N + P samples, and performing the calculation starting P samples into the frame. The choice is mainly determined by implementation issues (e.g. a system may have already performed the windowing operation before we start). No special windowing functions are required with this method: either we can perform the calculations directly on the main waveform, or a rectangular window can be used if windowing must be performed. 12.4.2 Autocorrelation Method A slightly different method know as the autocorrelation method can also be used, and as this has many practical advantages over the covariance method, it is often the method most used in practice. Recall from Section 10.3.9 that the autocorrelation function of a signal is defined as: R[n] = ∞ ∑ m=−∞ y[m]y[n−m] (12.23) This is similar in form to Equation 12.20, and if we can calculate the autocorrelation, we can make use of important properties of this signal. To do so, in effect we have to calculate the squared error from −∞ to ∞. This differs from the covariance method, where we just consider the speech samples from a specific range. To perform the calculation from −∞ to ∞, we window the waveform using a hanning, hamming or other window, which has the effect of setting all values outside the range 0 = n N to 0.
- 393. Section 12.4. Linear prediction analysis 369 While we are calculating the error from −∞ to ∞, this will always be 0 before the windowed area, as all the samples are 0 and hence multiplying them by the filter coefficients will just produce 0. This will also be the case significantly after the window. Importantly however, for P samples after the window, the error will not be 0 as it will still be influenced by the last few samples at the end of the region. Hence the calculation of the error from −∞ to ∞ can be rewritten as φ(j,k) = ∞ ∑ −∞ y[n− j]y[n−k] = N−1+P ∑ n=0 y[n− j]y[n−k] which can be rewritten as φ(j,k) = N−1−( j−k) ∑ n=0 y[n]y[n+ j −k] (12.24) As this is a function of one independent variable j−k, rather than the two of 12.24, we can rewrite it as φ(j,k) = N−1−k ∑ n=0 y[n]y[n+k] (12.25) As the limits of this expression equate to −∞ to ∞, this is the same as the autocorrelation function 12.23, and we can write: R(j −k) = φ(j,k) Now we can rewrite the matrix form of 12.21 in terms of the autocorrelation function: φ(j,0) = P ∑ k=1 φ(j,k)ak (12.26) R(j) = P ∑ k=1 R(j −k)ak (12.27) This significance of all of this is that the autocorrelation function (of any signal) is even, that is R(j −k) = R(j +k) = R(| j −k|), and this means that when we now write the matrix in Equation 12.22 in terms of R: R(1) R(2) R(3) . . . R(P) = a1 a2 a3 . . . aP R(0) R(1) R(2) ··· R(P−1) R(1) R(0) R(1) ··· R(P−2) R(2) R(1) R(0) ··· R(P−3) . . . R(P−1) R(P−2) R(P−3) ··· R(0) (12.28)
- 394. 370 Chapter 12. Analysis of Speech Signals Unlike the square matrix in Equation 12.22, this matrix is symmetric and all the elements in its diagonals are symmetric. This is known as a Toeplitz matrix, and because of the properties just mentioned, it is significantly easier to invert. 12.4.3 Levinson Durbin Recursion The matrix given in Equation 12.24 can of course be solved by any matrix inversion technique. Such techniques can be slow however (usually of the order of p3 where p is the dimensionality of the matrix) and hence faster techniques have been developed to find the values of {ak} from the autocorrelation functions R(k). In particular, it can be shown that the Levinson-Durbin recursion technique can solve Equation 12.28 in p2 time. For our purposes, analysis speed is really not an issue, and so we will forgo a detailed discussion of this and other related techniques. However, a brief overview of the technique is interesting in that it sheds light on the relationship between linear prediction and the all-pole tube model discussed in Chapter 11. The Levinson-Durbin recursion operates by considering a set of initialisation conditions, and using these to find the coefficients of a first order polynomial (that is, just a single term in z−1.) which minimise the mean squared error. From this we find the coefficients of a second order polynomial, third order and so on, using the previous polynomial and minimisation of the error, until we have reached a polynomial of the required filter order P. The algorithm initialises by setting the error and predictor coefficients of a (contrived) 0 order polynomial as: a0 0 = 0 E0 = R(0) Next we define a weighting factor for the ith model as: ki = R(i)− i−1 ∑ j=1 a i− j j R(i− j) Ei−1 using this, we find the predictor coefficients for the ith model: ai i = ki (12.29) ai j = ai−1 j −kiai−1 i− j 1 ≤ j ≤ i−1 (12.30) and from this, we can update the minimum mean squared prediction error for this step as: Ei = (1−k2 i )Ei−1 These steps are repeated until i = p where we have a polynomial and hence set of predictor coef- ficients of the required order. We have just seen how the minimisation of error over a window can be used to estimate the linear prediction coefficients. As these are in fact the filter coefficients that define the transfer function of the vocal tract, we can use these in a number of ways to generate other useful representations.
- 395. Section 12.5. Spectral envelope and vocal tract representations 371 12.5 SPECTRAL ENVELOPE AND VOCAL TRACT REPRESENTATIONS The preceding sections showed the basic techniques of source filter separation using first cepstral then linear prediction analysis. We now turn to the issue of using these techniques to generate a variety of representations, each of which by some means describes the spectral envelope of the speech. 12.5.1 Linear prediction spectra The main motivation in LP analysis is that it provides an automatic and explicit means for sepa- rating the source and filter components in speech. Recall the basic equation (11.2) of the vocal tract, Y(z) = U(z)V(z)R(z) Here Y(z) is the speech, U(z) is the source, V(z) is the vocal tract and R(z) is the radiation. Ideally, the transfer function H(z) found by linear prediction analysis would be V(z), the vocal tract transfer function. In the course of doing this, we could then find U(z) and R(z). In reality, in general H(z) is a close approximation to V(z) but is not exactly the same. The main reason for this is that LP minimisation criterion means that the algorithm attempts to find the lowest error for the whole system, not just the vocal tract component. In fact, H(z) is properly expressed as H(z) = G(z)V(z)R(z) for voiced speech, where G(z) represents the glottal transfer function, and H(z) = V(z)R(z) for unvoiced speech. The distinction between the linear prediction coefficient representing H(z) instead of V(z) has most significance for source modelling, and so we will delay a full discussion of this until Section 12.6. For most purposes, the fact that these H(z) models more than just the vocal tract isn’t a big problem, the upshot is that some additional poles are included in the transfer function, but the effect of these can oftne be ignored, and a good estimate of the vocal tract envelope can be found by converting H(z) directly to the frequency domain. If a purer estimate of V(z) is required, one of the techniques described in Section 12.6 can be used instead. The transfer function H(z) that we have been estimating can be converted to the frequency domain by simply setting z = ejωt (see Section 10.5.2) to give H(ejωt). Doing this gives us the spectral envelope of the linear prediction filter, which apart from the extra poles, we can interpret as the vocal tract spectral envelope. A plot of this LP envelope overlayed on the DFT is shown in Figure 12.13a. It is quite clear that the LP spectrum is indeed a good estimate of the spectral envelope. Figures 12.14 show spectral envelopes for a range of sounds. Here we can see a much noted “weakness” of LP envelope estimation; which is that the spectrum can appear very “peaky”. This
- 396. 372 Chapter 12. Analysis of Speech Signals 0 1000 2000 3000 4000 5000 6000 7000 8000 −4 −3 −2 −1 0 1 2 3 4 5 (a) Linear prediction spectrum superimposed on DFT 0 1000 2000 3000 4000 5000 6000 7000 8000 −4 −2 0 2 4 6 8 10 12 Frequency Amplitude (b) Linear prediction spectra for analysis of order 4,6,8,10,12,14,16 and 18. In reality the plots lie on top of one another, but each has been vertically separated for purposes of clarity. The bottom plot is from LP analysis of order 4, the top plot order 18. Figure 12.13 Examples of linear prediction spectra
- 397. Section 12.5. Spectral envelope and vocal tract representations 373 arises because in many speech sounds, particularly vowels, the poles lie close to, but not on the unit circle. At such locations, even a very small difference in their distance from the unit circle can result in a much sharper formant peaks. For example, for a sampling rate of 10000Hz, a pole with radius 0.96 will produce a formant bandwidth of 130Hz, while a pole of 0.98 produces a formant of bandwidth 65Hz, and so any slight error in determining the radius of the pole can produce a spectrum with exaggerated formants. This problem does not occur with pole frequencies, and so the LP envelope can normally be relied upon to find the positions of resonances quite accurately. Recall that in Chapter 11 we showed the number of tubes required is directly related to the sampling rate, length of tube and speed of sound (see Section 11.3.5) . For a speed of sound of 340ms−1 and vocal tract length of 17cm, we showed that 10 tube sections were required for a sampling rate of 10000Hz. Assuming these values for the speed of sound and vocal tract length, we see that we require one tube section per 1000Hz sampling rate. As each tube section gives rise to one pole, we conclude that V(z) should be a 10th order all-pole filter for a 10000Hz sampling rate, 16th order filter for a 16000Hz sampling rate and so on. The source (G(z)) and radiation (R(z)) filters are considerably simpler, and can be adequately modelled by 4 to 6 additional poles for both. It is highly advantageous to keep the number of poles even, and so an overall value of 14 poles or 16 poles is normally adopted for a 10000Hz sample rate, and 20 or 22 poles for a 16000Hz sample rate. Figure 12.13b shows the spectra predicted from linear prediction filters for a range of orders. It is quite clear that as the order increases, so does the detail in the spectrum. It is interesting to compare this to Figure 12.12 which shows the same investigation but with cepstra. It is clear that the cepstral spectra are much more heavily influenced by the harmonics as the order of analysis increases. 12.5.2 Transfer function poles As with any filter, we can find the poles by finding the roots of the transfer function polynomial. 12.5.3 Reflection coefficients Consider again the all-pole transfer function we have been using all along V(z) = G A(z) where A(z) is A(z) = 1− P ∑ k=1 akz−k (12.31) Recall that the key equation in the Levinson-Durbin recursion is the step that relates a predictor coefficient aj to the quantity k and the previous predictor coefficient aj−1 given in Equation 12.30.
- 398. 374 Chapter 12. Analysis of Speech Signals 0 1000 2000 3000 4000 5000 6000 7000 8000 −3 −2 −1 0 1 2 3 4 5 Frequency Amplitude (a) /ow/ 0 1000 2000 3000 4000 5000 6000 7000 8000 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 Frequency Amplitude (b) /s/ 0 1000 2000 3000 4000 5000 6000 7000 8000 −3 −2 −1 0 1 2 3 4 5 Frequency Amplitude (c) /eh/ 0 1000 2000 3000 4000 5000 6000 7000 8000 −3 −2 −1 0 1 2 3 4 Frequency Amplitude (d) /iy/ 0 1000 2000 3000 4000 5000 6000 7000 8000 −2 −1 0 1 2 3 4 5 Frequency Amplitude (e) /ae/ 0 1000 2000 3000 4000 5000 6000 7000 8000 −3 −2 −1 0 1 2 3 4 5 Frequency Amplitude (f) /ae/ Figure 12.14 Examples of spectra calculated from LP coefficients for a range of different sounds. Note how all the spectra are absent from any harmonic influence, and that all seem to describe the spectral envelope of the speech. Note the characteristic formant patterns and roll-off for all vowels, and compare this with the spectra for /s/. The same vowel is present in both examples (e) and (f). Note that while the formant positions occur in the same positions in both, the amplitudes and bandwidths are quite different, especially in the fourth formant, which is barely noticeable in example (e).
- 399. Section 12.5. Spectral envelope and vocal tract representations 375 ai i = ai−1 j −k2ai−1 i− j (12.32) Let us now examine what happens when we substitute Equation 12.31 into equation 12.32. For demonstration we shall choose the case for a third order filter, in which case we have three expressions: a3 3 = k3 a3 2 = a2 2 −k3a2 1 a3 1 = a2 1 −k3a2 2 The expansion of A3(z) is A3 (z) = 1−a3 1z−1 −a3 2z−2 −a3 3z−3 (12.33) and if we substitute equations for 12.31 into this we get A3(z) = 1−[a2 1 −k3a2 2]z−1 −[a2 2 −k3a2 1]z−2 −k3z−3 = 1−a2 1z−1 −a2 2z−2 +k3a2 2z−1 +k3a2 1z−2 −k3z−3 The first terms, 1−a2 1z−1 −a2 2z−2, are of course just the expansion of the second order filter A2(z) A3 (z) = A2 (z)−k3[a2 2z−1 +a2 1z−2 +z−3 ] (12.34) The other terms are similar to the polynomial expansion, except that the coefficients are in the reverse order to normal. If we multiply each term by z3 we can express these terms as a polynomial in z rather than z−1 which gives: A3 (z) = A2 (z)−k3z−3 [a2 2z2 +a2 1z+1] A3 (z) = A2 (z)−k3z−3 A2 (z−1 ) (12.35) In a similar fashion, the above relation can shown for any order of filter. The significance of all this can be seen if we compare Equation 12.35 with Equation 11.27. Recall that Equation 11.27 was used to find the transfer function for the tube model from the reflection coefficients, in the special case where losses only occurred at the lips. The equations are in fact identical if we set rk = −ki. This means that as a by product of Levinson-Durbin recursion, we can in fact easily determine the reflection coefficients which would give rise to the tube model having the same transfer function to that of the LP model. This is a particularly nice result; not only have we shown that the tube model is all-pole and that the LP model is all pole; we have now in fact shown that the two models are in fact equivalent and hence we can find the various parameters of the tube directly from LP analysis. Reflection coefficients are a popular choice of LP representation because they are easily found, they are robust to quantisation error and can in fact be used directly without having to
- 400. 376 Chapter 12. Analysis of Speech Signals be converted into predictor coefficients (see Rabiner and Schafer [368] for how this is done). Furthermore, and perhaps most significantly for our purposes, they have a physical interpretation, which means that they are amenable to interpolation. For example, if we had two different vocal tract configurations and wished to generate a set of vocal tract configurations between these, we could do by interpolating the reflection coefficients, in the knowledge that by doing so we are creating a set of reasonable vocal tract shapes. This process could not be done with the predictor coefficients themselves, as this would very likely lead to unnatural or unstable filters. 12.5.4 Log area ratios Recall from our tube model, that the reflection coefficients were originally defined in terms of the ratio of the areas between two adjacent tubes. Equation 11.15 stated that the amount of the forward travelling wave that was reflected back into the tube was given as: rk = Ak+1 −Ak Ak+1 +Ak Rearranging this gives: Ak Ak+1 = 1−rk 1+rk (12.36) Hence it is possible to find the ratio of two adjoining tubes from the reflection coefficients. It is not however possible to find the absolute values of the areas of the tubes (this is independent, just as the gain is independent of the filter coefficients). A common way of representing 12.36 is in terms of log area ratios, which are defined as: gk = ln Ak Ak+1 = 1−rk 1+rk (12.37) log area ratios come in useful in a variety of ways, for example as a means of smoothing across joins in a manner that reflects physical reality. 12.5.5 Line spectrum frequencies Recall from the tube model of Section 11.3 that losses in the model are only caused by wave propagation from the ends of the tube; no losses occur within the tube itself. Recall also from section 11.5.3 that the resistance factors which cause the losses have virtually no effect on the location of the resonant peaks. In using the Levinson-Durbin solution to the LP equations, we have simplified the loss situation further by assuming that losses only occur from one end of the tube. Line spectrum frequencies (LSFs) are found from an adaption of the tube model where we eliminate the losses from the glottis. This can be thought of as a process of creating a perfectly lossless tube and calculating its transfer function. Because the resistive factors have a minor impact of the locations of the formants in the frequency domain, this modelling assumption helps us find good estimates for the formant positions.
- 401. Section 12.5. Spectral envelope and vocal tract representations 377 0 50 100 150 200 250 0 1000 2000 3000 4000 5000 6000 7000 8000 Figure 12.15 An example speech wave of /... ii th .../ If the tube is completely lossless, all the poles will lie on the unit circle, the formants will not be damped and hence have infinite amplitude. The spectrum of this will therefore be extremely “spiky” and it is this property that gives rise to the name, in that the spectrum appears as a pattern of lines rather than a smooth function. We start the LSF analysis by constructing two new transfer functions for the lossless tube. We do this by adding a new additional tube at the glottis, which reflects the backwards travelling wave (whose escape would otherwise cause the loss) into the tube again. As explained in Section 11.3.3 , this can be achieved by either having a completely closed or completely open termination. In the closed case, the impedance is infinite, which can be modelled by a reflection coefficient value 1, by contrast the open case can be modelled by a coefficient value -1. Recall that we showed in section 12.5.3 that the operation of the Levinson-Durbin calculation was in fact equivalent to the determination of the transfer function for the tube model from its reflection coefficients. In both cases, we made use of an expression where we can formulate a new transfer function of order P+1, from an existing transfer function of order P and a new reflection coefficient rp+1: Ap+1 (z) = Ap(z)+rp+1zP+1 A(z−1 ) We can therefore use this to construct the transfer function for the LSF problem, where we wish to simulate a completely lossless system by addition of one extra tube and reflection coefficient. We
- 402. 378 Chapter 12. Analysis of Speech Signals do this for both the open and closed case, which gives us two new transfer functions, by convention called P(z) and Q(z): P(z) = Ap(z)−zP+1 A(z−1 ) (12.38) Q(z) = Ap(z)+zP+1 A(z−1 ) (12.39) As with any polynomial, these can be factorised to find the roots, but when we do so in this case, we find that because the system is lossless then all the roots lie on the unit circle. If we solve for P(z) and Q(z) we find that the roots are interleaved on the unit circle. In essence, the LSF representation converts the linear prediction parameters from a set of P/2 pairs of complex conjugate poles which contain frequency and bandwidth information, into a set of P/2 pairs of fre- quencies, one of which represents the closed termination and the other the open termination. Note that while the LSF tube is lossless, the loss information in the original LP model is still present. In the process of generating A(z) back from P(z) and Q(z), the extra ±1 reflection coefficient is removed and hence the loss and the bandwidths can be fully recovered. The LSF representation is attractive for our purposes because it is in the frequency domain and hence amenable to many types of analysis. This is true of the original linear prediction poles as well of course, we can convert these to a set of frequencies by taking their angle from the pole- zero plot. The LSFs have shown to be superior however in that they can be robustly estimated and interpolated, whereas the original pole frequencies can be quite sensitive to noise and quan- tisation. A final property of the LSFs is that when two occur close together, the overall value of the transfer function H(z) = 1/a(z) is high. This implies that the original poles responsible for these frequencies form a formant, and are not simply poles “making up the shape” of the spectral envelope. 12.5.6 Linear prediction cepstrum The cepstral representation has benefits beside its implicit separation of the source and filter. One often cited property is that the components of a cepstral vector are to a large extent uncorrelated, which means that accurate statistical models of such vectors can be built using only means and variances and so covariance terms are not needed (see Section 15.1.3). Because of this and other properties, it is often desirable to transform the LP coefficients into cepstral coefficients. Assuming that we wish to generate a cepstrum of the same order as the LP filter, this can be performed by the following recursion: ck = ln(G) k = 0 ak + 1 k k−1 ∑ i=1 iciak−1 1 k ≤ p (12.40)
- 403. Section 12.5. Spectral envelope and vocal tract representations 379 12.5.7 Mel-scaled cepstrum A very popular representation in speech recognition is the mel-frequency cepstral coefficient or MFCC. This is one of the few popular represenations that does not use linear prediction. This is formed by first performing a DFT on a frame of speech, then performing a filter bank analysis (see Section 12.2) in which the frequency bin locations are defined to lie on the mel-scale. This is set up to give say 20-30 coefficients. These are then transformed to the cepstral domain by the discrete cosine transform (we use this rather than the DFT as we only require the real part to be calculated): ci = r 2 N N ∑ j=1 mj cos πi N (j −0.5) (12.41) It is common to ignore the higher cepstral coefficients, and often in ASR only the bottom 12 MFCCs are used. This representation is very popular in ASR as it firstly has the basic desirable properties that the coefficients are largely independent, allowing probability densities to be mod- elled with diagonal covariance matrices (see Section 15.1.3). Secondly, the mel-scaling has been shown to offer better discrimination between phones, which is an obvious help in recognition. 12.5.8 Perceptual linear prediction It is possible to perform a similar operation with LP coefficients. In the normal calculation of these, spectral representations aren’t used and so scaling the frequency domain (as in the case of mel-scaled cepstrum) isn’t possible. Recall however that in the autocorrelation technique of LP, that we used the set of autocorrelation functions to find the predictor coefficients. In Section 10.3.9 we showed that the power spectrum is in fact the Fourier transform of the autocorrelation function, and hence the autocorrelation function can be found from the inverse transform of the power spectrum. So instead of calculating the autocorrelation function in the time domain, we can calculate it by first finding the DFT of a frame, squaring it, and then performing an inverse DFT. In this operation it is possible to scale the spectrum as described above. The final LP coefficients will therefore be scaled with respect to the Mel or Bark scale, such that more poles are used for lower frequencies than higher ones. 12.5.9 Formant tracking The above representations are all what can be termed as “bottom-up” representations in that they are all found from the original speech frame by means of a mathematical function or process. We can also consider other representations where we move away from these techniques and move towards representations which we believe more accurately represent the properties of the signal that are most perceptually important. In Chapter 7, we showed that formants are considered one of
- 404. 380 Chapter 12. Analysis of Speech Signals Figure 12.16 Example spectrogram with formants superimposed, as calculated by a formant track- ing algorithm. the most important representations in traditional phonetics, but in addition they are used in formant synthesis, explained later in Section 13.2, and they have also been used as cost functions in unit selection, as explained in Section 16. In traditional algorithms, which operated by a mixture of signal processing and rules, a three stage approach is generally followed. First the speech is pre-processed, for example to downsam- ple the speech to the frequency range where the formants occur; second candidate formants are found, and finally a heuristic step is applied. that determines the formants from the most likely candidates. As we have just seen, in many cases the poles which result from LP analysis can be considered to give rise to formants, and so it is natural that many formant tracking algorithms use LP poles as their initial candidates. An alternative is to find peaks in either the DFT or LP spec- trum and assign these as candidates. Yet another technique is that of multiple centroid analysis, based on the properties which define the overall shape of the spectrum [116]. Often the algorithms fix the number of formants to be found (usually three), and the post processing stage uses dynamic time warping to find the best formant paths (formant tracks) across the course of the utterance. More recently, statistical processing has been brought to bear on this problem. In essence, formant tracking is no different from any statistical problem in which we have evidence extracted from features, and constraints determined from a knowledge of the problem. Given such an ap- proach, it is possible to put this in statistical terms in which the candidates form a likelihood, and the known dynamics of the formants (e.g. their number and how much they can move from frame to frame) can form a prior. With this, a Viterbi search can be performed which finds the most likely formants for a sentence in a single step. Algorithms of this kind make use of Hidden Markov Models [4], nonlinear prediction algorithms [129], and graphical models [297]. These statistical algorithms are significantly more accurate and robust than their heuristic counterparts. In the early days of speech research, formants held the centre stage as the representation for speech. Today however, it is extremely rare to find formants used in speech recognition, synthesis or other systems. This is partly because robust formant trackers have always proved hard to develop, and so researchers moved towards cepstra and LP representations that could be robustly found from speech, even if their form didn’t explicitly show the type of information that
- 405. Section 12.6. Source representations 381 was thought most important. Furthermore, formants are really only a powerful representation for vowel sounds; they are not particularly good discriminators of most consonants. Hence even if a perfect formant tracker was available, other types of representations would be needed for some consonant sounds. 12.6 SOURCE REPRESENTATIONS We now turn to the topic of finding and analysing the source signal from the speech waveform. Of course linear prediction attempts to do just this, but just as the LP transfer function doesn’t always equate to the vocal tract transfer function, neither does x[n] equate to the source signal. This is basically because in LP analysis H(z) over-extends itself due to the minimisation criteria, and so subsumes some of the characteristics of the true glottal source. With a little effort however, we can in fact extract a good representation of the source by a number of techniques which we will now examine. As with all our other modelling problems, we have a range of trade-offs between modelling complexity and ease of use, and we will examine a few of these now. 12.6.1 Residual signals In the LP formulation of Equation 12.16 y[n] = P ∑ k=1 aky[n−k]+x[n] we termed x[n] the error signal, because it is the difference between the predicted and actual signal. Hence for a single frame of speech, if we perform LP analysis and note the values of x[n] we will have a complete error signal for that frame. Alternatively, we can find this by a process of inverse filtering. From simple rearrangement of Equation 12.16, we can express x[n] as follows: x[n] = y[n]− P ∑ k=1 aky[n−k] = P ∑ k=0 aky[n−k] where a0 = 1 Here we have expressed x[n] purely in terms of the coefficients {ak} and the current and previous values of y[n]: if we compare this to equation 10.54, we see that this is just a standard FIR filter. Hence if we have the speech signal and the coefficients for a single frame, we can find x[n] by a standard FIR filter operation. Apart from a gain normalisation, this will be identical to the error signal x[n]. The inverse filtering technique is useful because often we are required to find the error signal after the LP analysis has been performed, or for cases (explained below) where we wish to
- 406. 382 Chapter 12. Analysis of Speech Signals find the residual for a different interval than that used for analysis. If the residual signal is then fed through the normal LP (IIR) filter, the original speech is reconstructed. This process is shown in Figure 12.17. If we apply the inverse filtering technique to each frame in the entire speech waveform, we create a sequence of residual signals, one for each frame. A single residual waveform for the whole utterance can be made from these residual frames; but as the details of this are somewhat subtle, a description of this is left until section 12.6.5. How do we now find the glottal source signal from the residual x[n]? Recall that in the z-domain, the main filter expression is Y(z) = U(z)V(z)R(z) where Y(z) is the speech output, U(z) is the source, V(Z) is the vocal tract and R(z) the radiation. Recall from section 11.4, that we can model this source in two main ways. Firstly, we can model it explicitly as a time domain function, which has a number of parameters determining the length of the opening phase, the length of the closed phase and so on. The Lijencrants-Fant (LF) [159] model of this is shown in Figure 11.16. Alternatively, we can model it as a sequence of impulses which are passed through a glottal filter G(z). From experimentation and explicit modelling, we know that this glottal filter will contain poles and zeros, but will be of considerably lower order than the vocal tract filter. Here we will examine the filter model, as a proper understanding of this helps in the determination of the equivalent parameters for the time domain model. Using I(z) to represent the impulse sequence, we can therefore express U(z) as: U(z) = G(Z)I(z) (12.42) U(z) = M ∏ k=0 bG k z−k 1− N ∏ l=1 aG l z−l I(z) (12.43) where {bG k } and {aG l } represent the coefficients of the glottal filter G(z). Recall from section 11.4 that the radiation can be approximated as a differentiator, which can be expressed as a single zero FIR filter. Because the filter coefficients of the G(z)V(z)R(z) will combine, the result will be that the numerator will have the zeros of the glottal source and radiation, while on the denominator, the poles of the glottal filter will combine with the poles of the vocal tract. The problem is that the result of LP analysis will give us a single expression H(z) for the glottal and vocal tract poles, without being able to show which poles belong to which filter. Furthermore, the fact that we are attempting to model a system with poles and zeros with an all-pole analysis further complicates the matter. To perform more explicit source/filter separation, we need to find a way of cancelling the effect of the zeros and separating the poles into their component filters.
- 407. Section 12.6. Source representations 383 0 50 100 150 200 250 300 350 400 −0.04 −0.02 0 0.02 0.04 0.06 (a) One windowed frame of voiced speech 0 50 100 150 200 250 300 350 400 −0.01 −0.005 0 0.005 0.01 (b) The residual for the frame of speech, calculated with an inverse filter of the predictor coefficients. Note that the error is highest a) at the start of the frame, where there are no proper previous values on which to base the predication, and b) at the instant of glottal closure. The prediction error is highest at this point because it is here that the glottal derivative flow is at its highest. 0 50 100 150 200 250 300 350 400 −0.04 −0.02 0 0.02 0.04 0.06 (c) Reconstructed waveform, created by passed the residual through filter. This is virtually identical to the original speech Figure 12.17 Comparison of residual with original and reconstructed waveforms.
- 408. 384 Chapter 12. Analysis of Speech Signals 0 200 400 600 800 1000 1200 −3 −2.5 −2 −1.5 −1 −0.5 0 0.5 x 10 4 open phase closed phase Figure 12.18 Model of glottal flow, model of glottal flow derivative, and real speech waveform over time . 12.6.2 Closed-phase analysis Figure 12.18 shows a diagram of a glottal flow input, the glottal flow derivative and the output speech. From this, we see that during each pitch period there is an interval when the glottal flow is zero because the glottis is shut. As there is no contribution from the glottis here, if we estimate H(z) over this interval, we will be able to equate this to a combined V(z) and R(z) expression, free of influence from G(z). Calculation of the LP coefficients over this interval is known as closed-phase analysis. To perform closed phase analysis properly, we need to determine when the glottis shuts; these points are called the instants of glottal closure and can be found automatically by the epoch detection algorithms described below in Section 12.7.1. A further benefit to performing closed phase analysis is that during these intervals, the closed glottis obeys the model assumptions of
- 409. Section 12.6. Source representations 385 being completely reflective and having an infinite impedance; by contrast, during the open phase there are losses at the glottis, but worse there will definitely be interaction with the sub-glottal system (that is, the assumption made in section 11.3.3 where we assume that no signal is reflected back from the lungs will not hold). So given that G(z) has negligible effect during the closed phase, all we need now do is remove the effect of R(z) to find V(z). From experimentation and knowledge of radiation properties, we can approximate R(z) as a differentiator of the signal. This can be modelled by a first order FIR filter of the form: y[n] = x[n]−αx[n−1] Hence the effect of this can be removed by multiplying the signal with the inverse equivalent of this, which is an IIR filter, acting as an integrator: x[n] = y[n]+αx[n−1] For reasons of stability, it is best to have α slightly less than the “true” integrating/differentiating value of 1; values in the range 0.95 to 0.99 are often used. This process is known as pre-emphasis. Now we have a reasonably accurate expression for the vocal tract transfer function, we can inverse filter the speech to find the glottal flow signal. The trick here is to inverse filter the signal over a whole pitch period (or longer) rather than just the interval over which the LP coefficients were calculated. By doing this, we inverse filter over the region where the glottis is open, and from this we can find the form of the input, glottal signal. This process is shown in Figure 12.19, along with the shapes of the glottal flow and glottal flow derivative from the Lijencrants-Fant model. The figure clearly shows that the glottal flow derivative is highly similar to the residual signal, especially with regard to the large negative spike at the instant of glottal closure. The figure also shows that the assumptions made about the impulse excitation in the basic LP model are also quite justified; the excitation in the closed phase comes from the large negative going spike in the glottal derivative signal can be approximated by an impulse quite easily. In many cases, especially in female or other high pitched speech, the length of the closed phase can be very small, perhaps only 20 or 30 samples. In the autocorrelation method, the initial samples of the residual are dominated by the errors caused by calculating the residual from the zero signal before the window. The high (and erratic) error in the residual can be seen in the first few samples of the residual in Figure 12.17b. For short analysis windows this can lead to a residual dominated by these terms, and for this reason, covariance analysis is most commonly adopted for closed phase analysis. It should be noted that in many real-life cases, the pre-emphasis term is not applied, and hence the calculated residual is in fact a representation of the glottal flow derivative. This is the case for most speech synthesis purposes, and henceforth we will assume that all residuals are calculated in this way, unless otherwise stated.
- 410. 386 Chapter 12. Analysis of Speech Signals 0 200 400 600 800 1000 1200 −3.5 −3 −2.5 −2 −1.5 −1 −0.5 0 0.5 x 10 4 Figure 12.19 From top to bottom, model of glottal flow, model of glottal flow derivative, real speech waveform and real residual . 12.6.3 Open-phase analysis It is not always possible or desirable to used closed-phase analysis. Sometimes other constraints in an application mean a fixed frame spacing is required, and sometimes it is simply not possible to find the instant of glottal closure with any degree of accuracy. What then happens if the LP analysis is performed over an arbitrary section of speech? The main upshot of this is that the LP analysis will attempt to model the effect of the zeros in the glottal transfer function, and in doing so include the effect of the zeros in its poles. From the basic Maclaurin expansion of ∞ ∑ n=0 xn = 1+x+x2 +x3 +.... = 1 1−x
- 411. Section 12.6. Source representations 387 we can show that ∞ ∑ n=0 bn z−1 = z z−b = 1 1−bz1 inverting this, we get: 1 ∞ ∑ n=0 bnz−1 = 1−bz1 which shows that a single zero can be exactly modelled by an infinite number of poles. It can be shown that so long as the number of poles ≫ than the number of zeros, a reasonable approximation of the effect of zeros can in fact be modelled by the poles. In doing do however, we are in effect modelling these zeros with all the poles in the transfer function, and so the poles will all be at slightly different locations than in the case of closed-phase analysis. Luckily, this doesn’t seem to have too great an influence on the frequency response of the LP transfer function, but it does have a knock on effect when the LP filter is used to find the residual. As we have modelled the glottal zeros with poles, the residual now does not closely resemble the glottal derivative, rather it is significantly “noisier”, and “spikier”. This means, that first, the spectrum of the residual will be flat, and that secondly the main excitation will be in the form of isolated impulses rather than the shape of the glottal derivative function of closed phase analysis. This can be seen by comparing the open phase residual of Figure 12.17 with the closed-phase residual of 12.19. In general, fixed frame LP analysis is the norm, and so the open phase residual signals such as that shown in Figure 12.17 are most commonly observed. Note that this ability to model zeros with poles also allows us to provide a means to represent vocal tract transfer functions for nasals and obstruents with all-pole functions. While the accuracy of these all-pole expressions is not as high as for vowels, the convenience in being able to use a single method and expression for all speech sounds usually outweighs any issues arising from the inability to model vocal tract zeros properly. It is worth making a final point about residuals when they are used as sources. Recall from the original LP equation that as well as the residual being the outcome when the speech is inverse filtered, it is also simply the error in the LP analysis. Hence if a residual is used to excite the filter from which it came, the result should be a perfect reconstruction of the original signal. As we shall see in Section 14, there are many implementation issues to do with frame edges and so on, but it is worth noting that no matter how crude any of the assumptions we have made with regard to LP analysis in general, a powerful feature of the model is that the speech can be exactly reconstructed if excited with the correct input. 12.6.4 Impulse/noise models Linear prediction was in fact primarily developed for use in speech coding applications. As we have just seen, performing LP analysis allows us to deconvolve the signal into a source and filter, which can then be used to reconstruct the original signal. Simply separating the source and filter
- 412. 388 Chapter 12. Analysis of Speech Signals doesn’t actually save any space, the trick comes in replacing the residual by a much simpler (and more economical) source signal. In what we shall henceforth call classical linear prediction the source is modelled by two systems; one for voiced sounds and one for obstruent sounds. In Figure 12.19, we saw that the in- tegrated residual signal was a reasonably close approximation to the glottal flow derivative signal. The primary feature in such a signal is the negative spike caused by the glottal closure. As this is reasonably close to an impulse, the classical LP model simply uses a single impulse as the excita- tion for voiced sounds. To do this properly, one impulse is generated for every pitch period, and so in analysis, an estimation of the pitch period has to be found and stored/transmitted. All obstruent sounds have a frication source, generated by white noise in our model from section 11.5.2. In the classical LP system, all we need store/transmit is the fact that we have an obstruent sound, upon which a noise source is used during reconstruction. For voiced obstruents, both the impulse and noise source are used. The impulse/noise model provides a very efficient means of encoding the residual, but does so at some cost, the effects of which can clearly be heard in the reconstructed signal as a buzz or metallic sound. Apart from the space saving, having an impulse excited system has the additional benefit in that we are free to generate the impulses when we like; in other words, we are not limited to generating the impulses at the same spacing as in the analysis. This provides a means of changing the pitch of a signal and in fact forms the basis of the synthesis technique described in Section 13.3. 12.6.5 Parameterisation of glottal flow signal Performing the above analysis allows us to find a reasonable estimate of the glottal flow or glottal flow derivative. In fact, we can go one stage further and parameterise this by further analysis. One technique described by Quatieri [364] involves fitting the Lijencrants-Fant model to the residual by means of least squares minimisation. In philosophy, this is of course similar to linear prediction analysis itself, where we attempt to find the set of predictor coefficients which give the lowest error. In practice, the LF model is non-linear and so the LP technique is not available to us. Instead, numerical techniques such as the Gauss-Newton method are used to iterate to a set of optimum model parameters. Dutoit [148] describes a different technique where the glottal source parameters are actually found at the same time as the predictor coefficients. While we can find the optimal parameters for any given frame for a model such as LF, we never actually achieve a perfect fit. In fact, we can determine a sample by sample difference between the model and the residual. This of course, is another “residual” or “error” signal, which now represents the modelling error after the LP model and the LF model have been taken into account. One interesting study by Plumpe et al [356] describes how this signal is in fact highly individual to speakers and can be used as the basis of speaker recognition or verification.
- 413. Section 12.7. Pitch and epoch detection 389 12.7 PITCH AND EPOCH DETECTION 12.7.1 Pitch detection The process of finding the fundamental frequency (F0) or pitch from a signal is known as pitch detection or pitch tracking. Recall from Chapter 7 that fundamental frequency is defined as the frequency at which the glottis vibrates during voiced speech, whereas pitch is the perceived rate of periodicity in the speech signal by the listener. In general, we ignore this distinction, mainly because in most cases any difference between these two values is less than the known accuracy of any detection algorithm. One point that is important though is that F0 tracks are in fact used for many quite different purposes. In some of the signal processing operations described below, an accurate frame by frame value is required; however in intonational analysis, we are usually only interested in the general shape of the track, and so small perturbations are often smoothed for such needs. As with formant tracking, traditional pitch detection algorithms PDAs often have three stages; a pre-processing stage, where for instance the speech is downsampled or low-pass filtered, a bottom up candidate finding step, and a decision making step whereby the best candidates are picked. A wide variety of techniques have been proposed for pitch tracking, including: • Using the cepstrum. As cepstral analysis generates a representation where the pitch infor- mation appears as a spike, it is possible to find this spike and determine the pitch from that. • Autocorrelation function. In the time domain, successive pitch periods are usually similar to one another. As the autocorrelation functions measure how similar a signal is to itself when shifted, it will have a high value when a a pitch period is most closely overlapped with its neighbour. We can hence find the period by finding these points where the autocorrelation function is at a maximum. In a similar fashion, the normalised cross correlation function can be used. • Frequency domain. As the harmonics in the DFT are spaced at multiples of the period, these can be used to find the period. One strength in this technique is that there are multiple harmonics and hence more than one measurement can be made. A drawback is that they can often be hard to measure accurately, especially at higher frequencies. • Residual analysis. As the residual is mostly free from vocal tract influence, it represents the source, and from this the period can be seen. In practice this technique is rarely used as the artefacts in LP analysis often make the determination of the period quite difficult. The classic work on pitch detection is that of Hess [202] who gives a thorough overview of all the PDA techniques just mentioned. With PDAs, the basic techniques only go so far and an important component of the algorithm is the use of many heuristic or empirically determined factors. Bagshaw [28] conducted an objective test on 6 popular techniques and claims that the time domain super resolution pitch determination algorithm [309] is the best. Other well known pitch-
- 414. 390 Chapter 12. Analysis of Speech Signals 0 0.5 1 1.5 2 2.5 3 0 50 100 150 200 250 300 (a) “raw” output from PDA 0 0.5 1 1.5 2 2.5 3 0 50 100 150 200 250 300 (b) “raw” output, but plotted with symbols rather than lines 0 0.5 1 1.5 2 2.5 3 0 50 100 150 200 250 300 (c) Smoothed and interpolated output from PDA Figure 12.20 Different ways of displaying F0 plots against time
- 415. Section 12.7. Pitch and epoch detection 391 trackers include Talkin’s get f0 and RAPT algorithms [431] (these were not included in Bagshaw’s test, but are generally acknowledged to be very accurate). As with formant tracking, traditional PDAs suffer somewhat from their heuristic rather than probabilistic formulation. Recently, more interest in probabilistic techniques has arisen [141], and for instance Li et al [283] cite accuracy results for a probabilistic technique which is significantly more accurate than get f0. 12.7.2 Epoch detection: finding the instant of glottal closure We have just seen that closed phase linear prediction requires that we analyse each pitch period separately. This type of speech analysis is called pitch synchronous analysis and can only be performed if we are in fact able to find and isolate individual periods of speech. We do this by means of an pitch marking orepoch detection algorithm (EDA). The idea behind an EDA is to locate a single instant in each pitch period that serves as an “anchor” for further analysis. These positions are often known as pitch marks, pitch epochs or simply epochs. In general, they can refer to any reference point, but are often described in terms of salient positions on the glottal flow signal, such as the peak of the flow for each pitch period. For many types of analysis (such as TD-PSOLA described in Section 14.2) it doesn’t really matter where the anchor point is chosen, so long as it is consistent from period to period. Often a time- domain reference point such as the peak of the highest excursion in each period is used, or the trough of the lowest excursion. That said, many analysis techniques focus on one particular point known as the instant of glottal closure (IGC). This is manifested as a large negative going spike in the glottal flow derivative, as shown in Figure 12.21. If we view this as the input to the vocal tract (which we do be including the radiation term), we see that it closely resembles a large negative impulse. Such an impulse excites the filter in such a way that we can examine the filter properties without interference from the ongoing effects of the glottis. As we saw above, this is the basis on which close-phase LP analysis is performed. Locating the instant of glottal closure and epoch detection in general is widely acknowledged as a thorny problem. In the past, most pitch synchronous applications could use a rough estimate of the ICG, as the technique was not particularly sensitive to determining these instants exactly. Indeed, the closed phase LP technique described above does not need a particularly accurate esti- mate - the window used by the technique does not have to start exactly at the ICG. With the more widespread adoption of pitch synchronous speech synthesis techniques (described in Chapter 14) it becomes more important to locate the exact ICG, or some other anchor point, more exactly. Dutoit [148] gives a review of some of the difficulties involved, and describes a process common in the early 1990s whereby researchers would actually mark epochs by hand. This was seen as being an acceptable burden when speech databases were small and when the only alternative was to use other much poorer sounding synthesis techniques. Dutoit describes some of the difficulties involved in this, most specifically the fact that because of various phase effects, it is not always the largest negative peak in the waveform which corresponds to the IGC. Furthermore, it is known that phase distortion can make time domain marking of waveforms virtually impossible for either
- 416. 392 Chapter 12. Analysis of Speech Signals 0 2000 4000 6000 8000 10000 12000 14000 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 Figure 12.21 Comparison of (from top to bottom) waveform, laryngograph signal, integrated resid- ual and residual. machine or human. These difficulties led many to a radical alternative, which was to measure the glottal be- haviour by means other than from the speech waveform. A common techniques was to use a device known as an electroglottograph or a laryngograph. This consists of two metal discs which are placed on each side of the protruding part of the larynx (the “adam’s apple”). A very high frequency alternating current is passed through the system, such that the current passes from one disc, through the larynx, to the other disc, and from this the electrical impedance and flow of current across the larynx can be measured. If the glottis is open, the impedance is high and the current is low; if the glottis is shut the impedance is low and the current is high. Laryngograph sig- nals (known as Lx signals) do not however generate the glottal flow signal itself as the relationship between the electrical impedance and flow is not simply related. The Lx signal is however quite simple, and from it, it is usually possible to find the instants when the glottis shuts (the IGC) and opens. Figures 12.21 and 12.22 show a speech waveform, an Lx signal and a LP derived estimate of the glottal flow derivative. The Lx signal clearly shows each pitch period, and is particularly clear at showing exactly when the transition between voiced and unvoiced sounds occurs, which is one of the most difficult areas to perform accurate epoch detection. Epoch detection by means of Lx signals has been quite widely used in TTS in the last 15 years, but it is rare to find explicit mention of this. Acero [3] does however provide a good account of the process of Lx use for TTS. There are two main problems with Lx use. Firstly, the technique doesn’t work well for some
- 417. Section 12.7. Pitch and epoch detection 393 6000 6200 6400 6600 6800 7000 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 instant of glottal closure instant of glottal opening Figure 12.22 Fine grained comparison of (from top to bottom) waveform, laryngograph signal, integrated residual and residual. The vertical lines show the instants of glottal closure and glottal opening. Here we can see that while the general periodicity of the form of the LX and residual glottal flow are similar, the functions are not simply related in shape. Note that when the glottis is open, the LX signal is at a minimum because the current flow is at a minimum. The integrated residual signal (c) approximates the glottal volume velocity. Ideally the closed phase in this signal should be flatter, but the many modelling errors combine to give this part of the waveform a gradual upward slope. A small sharpening in the gradient of this can be seen at the instant of glottal opening. speakers due to physiological reasons. Secondly, it is simply awkward to have to record all the speech data with both a microphone and the Lx device. It is not particularly comfortable (although it is completely safe) and some speakers, particularly highly paid voice talents, are not happy with its use. Thankfully, a number of fully automatic techniques now do exist that can extract the IGCs from the speech waveform directly. Many of these work on the residual, and it is known that the negative pulse of the residual can give quite an accurate estimation of the IGC if care is taken with the analysis and the signal does not suffer from phase distortion. Recently several other techniques have been developed for automatic extraction from speech. Brookes [68] describes a technique which uses the glottal energy flow, and claims that this is significantly more accurate than residual analysis. Vedhius describes a general technique that uses dynamic programming to constrain the epoch search, and shows that this technique can be extended to other points of interest besides the IGC [469].
- 418. 394 Chapter 12. Analysis of Speech Signals 12.8 DISCUSSION For those approaching these topics for the first time, some of the formulations, reasoning and sheer number of representations can make the speech analysis subject seem quite daunting. It is in fact virtually impossible to learn this subject purely from the mathematics: speech analysis is a practical subject and requires practical study, and by far the best means to study this is by interac- tively playing with real speech signals. Luckily a number of packages exist which allow intimate examination of speech signals. Using these, it is possible to examine magnitude spectra, linear prediction envelopes, spectrograms of different types and so on. In fact, a good understanding of the speech representations described here can be gained from interactive study alone without any understanding of the equations involved (although this of course is not recommended). Speech analysis is a broad topic and is used in many applications. Historically, most of the techniques described here were first developed for use in speech coding, where the aim is to trans- mit a signal down a wire as efficiently as possible. With any general type of data, there is often considerable redundancy, and by studying the patterns in which the data naturally occurs we can study this redundancy and eliminate it. These general techniques only go so far, and it soon be- came apparent that special purpose coders designed specifically for speech signals could be much more efficient, delivering more than a factor of 10 compression rate. In fact, linear prediction was developed for just this purpose. Linear predictive coding (LPC) as it is known, separates the source and filter for each frame, but only sends the filter coefficients, normally encoded as reflec- tion coefficients or line spectral frequencies. The residual is not transmitted, rather an estimate of the pitch is sent for voiced signals, and a flag stating whether the source is noisy (ie is an obstruent sound). When the speech is to be reconstructed, the pitch information is used to generate an im- pulse sequence for voiced signals, and white noise is added if the noise flag it set. The quality of LPC speech is considerably worse than the original, but it has proved a popular choice, especially for situations with user acceptance and requirements for very low transmission rates (usually mili- tary). More recently, LPC coders with more sophisticated residual reconstruction techniques have been used in normal commercial applications. Automatic speech recognition systems use speech analysis as their first stage. All ASR systems are essentially pattern recognisers, and so work well with when presented with input features with certain properties. These include that the representation should be compact, so that the number of parameters to be learned is low; and this shares obvious common ground with the speech coding applications just mentioned. A good example is that many ASR systems use multi- variate Gaussians to model phone distributions (see Chapter 15). For a vector with 10 components, doing so fully requires a set of 10 means and a set of 10 by 10 (100) covariances. If however the components are statistically independent of each other, we can ignore most of the covariance terms and model the distribution with 10 means and just 10 variances. In addition, it is desirable that the representations should be invariant for a single class; for this reason, recognition is never performed on the waveform itself, as the phase aspect makes signals with the same envelope look
- 419. Section 12.8. Discussion 395 quite different. Currently, mel-scale cepstral coefficients, and perceptual linear prediction coefficients trans- formed into cepstral coefficients, are popular choices for the above reasons. Specifically they are chosen because they are robust to noise, can be modelled with diagonal covariance, and with the aid of the perceptual scaling are more discriminative than would otherwise be. From a speech synthesis point of view, these points are worth making, not because the same requirements exist for synthesis, but rather to make the reader aware that the reason MFCCs and PLPs are so often used in ASR systems is for the above reasons, and not because they are intrinsically better in any general purpose sort of way. This also helps explain why there are so many speech representations in the first place; each has strengths in certain areas, and will be used as the application demands. In fact, as we shall see in Chapter 16, the application requirements which make, say, MFCCs so suitable for speech recognition are almost entirely absent for our purposes. We shall leave a dis- cussion as to what representations really are suited for speech synthesis purposes until Chapter 16. 12.8.1 Further reading There are several good books on speech analysis, most of which adopt an application neutral attitude. The classic work is Rabiner and Schafer [368] and while written several decades ago, this book remains an excellent reference for all the fundamentals of speech analysis. It assumes the reader is familiar with basic signal processing, and so can be a little difficult for those with no prior experience. Given an understanding of the topics covered in Chapter 10 of this book, the reader should have little difficulty. An excellent more recent book is Quatieri [364]. The subject matter doesn’t differ much from that of Rabiner and Schafer, but the book covers many of the techniques developed since. Other good references include Furui [168], Huang et al [224], Kleijn and Paliwal [256], Deller and Proakis [128] and Gold and Morgan [176]. 12.8.2 Summary • Speech analysis is in general concerned with three main problems: 1. eliminate phase; the magnitude DFT does this 2. separate source and filter; this can be done with cepstral analysis or linear prediction 3. transform the representation into a space which has more desirable properties; log mag- nitude spectra follow the ear’s dynamic range; mel-scaled cepstra scales according to the frequency sensitivity to the ear. Log area ratios are amenable to simple interpolation and line spectral frequencies show the formant patterns robustly. • The continuously varying signal is divided by a process called windowing into a sequence of stationary smaller signals called frames. • Windowing affects the signal and so a perfect representation of the original waveform is not possible.
- 420. 396 Chapter 12. Analysis of Speech Signals • The most common spectral representation is achieved by applying a discrete Fourier trans- form (DFT) to the frame of speech. • Successive application of DFTs can be used to generate a spectrogram. • The cepstrum is the inverse Fourier transform of the log magnitude spectrum. It separates variation in frequency across the range and so implicitly separates source and filter compo- nents in the spectrum. • Linear prediction performs source/filter separation by assuming an IIR system represents the filter. This allows the filter coefficients to be found by a process of minimising the error predicted from the IIR filter. • The linear prediction equations can be solved by the autocorrelation or covariance methods. Each involves solving the equations by means of efficient matrix inversion. • The direct linear prediction, IIR, coefficients are inherently unrobust with regard to quanti- sation and interpolation. A set of derived representations, including reflection coefficients, log area ratios and line spectral frequencies avoid these problems. • The source signal, called the residual, can be found by inverse filtering the speech signal with the linear prediction coefficients. • With appropriate care during analysis and filtering, this commonalities between this residual and the glottal flow derivative signal can be shown.
- 421. 13 SYNTHESIS TECHNIQUES BASED ON VOCAL TRACT MODELS In this chapter we introduce the three main synthesis techniques which dominated the field up until the late 1980s, collectively known as first generation techniques. While these techniques are used less today, it is still useful to explore them, as apart from simple historical interest, they give us an understanding of why today’s systems are configured the way they are. As an example, we need to know why today’s dominant technique of unit selection is used rather than the more basic approach which would be to generate speech waveforms “from scratch”. Furthermore, modern techniques have only been made possible by vast increases in processing power and memory, so in fact for applications which require small footprints and low processing cost, the techniques explained here remain quite competitive. 13.1 SYNTHESIS SPECIFICATION: THE INPUT TO THE SYNTHESISER First generation techniques usually require a quite detailed, low-level description of what is to be spoken. For purposes of explanation, we will take this to be a phonetic representation for the verbal component, along with a time for each phone, and an F0 contour for the whole sentence. The phones will have been generated by a combination of lexical lookup, LTS rules and post- lexical processing (see Chapter 8), while the timing and F0 contour will have been generated by a classical prosody algorithm of the type described in Chapter 9. It is often convenient to place this information in a new structure called a synthesis specification. Hence the specification is the input to the synthesiser and the waveform is the output. Initially, we will assume a simple model where the specification is a list S S S = = = s s s1, , ,s s s2, , ,. . .. . .. . ., , ,s s sN , in which each item s s sj contains the phoneme identity, a duration, and one or more F0 values. Recall from Chapter 9, that the F0 contour is necessarily dependent on the timing information, as the F0 generator must produce an F0 contour that matches the length of the utterance. One option is to specify the F0 contour on a frame by frame basis, so that if we had say a 50ms duration phoneme and a 10ms frame shift, we would have 5 F0 values in the specification. A 70ms duration phoneme would have 7 frames and so on: 397
- 422. 398 Chapter 13. Synthesis Techniques Based on Vocal Tract Models [F0 ] = 121,123,127,128,126 (5 frame example) [F0 ] = 121,123,127,128,126,121,118 (7 frame example) and so a example of a complete specification item would look like s s sj = PHONEME /ah/ DURATION 70 F0 125,123,119,117,115,115,114 Having different numbers of F0 values per item can be unwieldy, so someti