Speech Feature Extraction and Data Visualisation

IT in Industry, vol. 3, no. 1, 2015 Published online 31-Mar-2015
ISSN (Print): 2204-0595
Copyright © Authors 16 ISSN (Online): 2203-1731
Speech Feature Extraction and Data Visualisation
Vowel recognition and phonology analysis of four Asian ESL accents
David Tien
School of Computing and Mathematics
Charles Stuart University
Bathurst, NSW, Australia
Yung. C. Liang and Ayushi Sisodiya
Department of Electrical and Computer Engineering
National University of Singapore
Singapore
Abstract—This paper presents a signal processing approach
to analyse and identify accent discriminative features of four
groups of English as a second language (ESL) speakers, including
Chinese, Indian, Japanese, and Korean. The features used for
speech recognition include pitch, stress, formant frequencies, the
Mel frequency coefficient, log frequency coefficient, and the
intensity and duration of vowels spoken. This paper presents our
study using the Matlab Speech Analysis Toolbox, and highlights
how data processing can be automated and results visualised.
The proposed algorithm achieved an average success rate of
57.3% in identifying vowels spoken in a speech by the four non-
native English speaker groups.
Keywords—speech recognition; feature extraction; data
visualisation;vowel recognition; phonology analysis
I. INTRODUCTION
In speech processing, pattern recognition is a very
important research area as it helps in recognizing similarities
among different speakers, and plays an indispensable role in
the design and development of recognition models and
automatic speech recognition (ASR) systems. Pattern
recognition consists of two major areas: feature extraction and
classification. All pattern recognition systems require a front
end signal processing system, which converts speech
waveforms to certain parametric representations, called
features. The extracted features are then analysed and
classified accordingly [1, 2]. Feature extraction techniques can
be broadly divided into temporal analysis and spectral
analysis. In temporal analysis, the speech waveform is used
directly for analysis; while in spectral analysis, spectral
representation of the speech signal is used instead. In our
study, temporal analysis techniques employed include pitch,
intensity, formant frequency, and log energy; whereas spectral
analysis techniques employed include the Mel-frequency
spectral and log-frequency spectral.
II. VOWEL RECOGNITION AND PHONOLOGY ANALYSIS
This paper proposes an easy to understand and user
friendly approach for speech feature extraction and data
visualization. Software in Matlab has been developed to
recognize features as spoken by four groups of non-native
English speakers (or English as Second Language, ESL) –
Chinese, Indian, Korean, and Japanese. These groups of
speakers have very distinctive accents.
A. Matlab – Speech Analysis Toolbox
The developed software makes the speech processing tasks
easy and accurate. The language toolbox provides the user
with a general envelop graph for different languages, and is
very useful for the development of speech recognition
applications. This work also plays a vital role as a catalyst for
future linguistic research.
The major functionalities of the software used in our work
include:
 Pitch Analysis
 Intensity Analysis
 Frequency Analysis
 Log Energy Analysis
 Log- Frequency Power Analysis
 Mel- Frequency Cepstral Analysis
 Vowel Identification
 Recording and maintaining of Speech Corpus of
different languages
 Language analysis of the speech features of different
languages
B. Formant Frequency for Vowel Identification
The formant frequency is considered a very important
aspect of the speech signal in the frequency domain. We used
formant frequencies in our work for vowel identification. We
will discuss the algorithm and its performance later in this
paper.
As explained by Prica and Ilić, the main difference
between vowels and consonants is that vowels resonate in the
throat. When a vowel is pronounced, the formants are exactly
the resonant frequencies of the vocal tract [3]. It is a natural
and obvious choice to use formant frequencies for vowel
identification.
The frequencies and levels of the first three formants of
vowels were measured. The statistical analysis of these
formant variables confirmed that the first three frequencies are
the most appropriate distinctive parameters for describing the
spectral differences among the vowel sounds. Maximum

ISSN (Print): 2204-0595
likelihood regions were computed and used to classify the
vowels. For speech recognition in particular, the formant
analysis is more practical because it is much simpler and can
be carried out in real time. The steps in our implementation
are listed below:
1. Read a speech file in wav format and store it into a
vector.
2. Measure the sample frequency and calculate the pace
of the speech according to which it divided the speech
into appropriate fragments for formant analysis.
3. Select the small divided signals one by one, and
further divide them into 10 equal blocks in the time
domain.
4. Select the block with maximum power content, which
is also a measure of the stress of vowel.
5. Normalize the selected block and shows the waveform
of the vowel segment.
6. Determine the frequencies at which the peaks in the
power spectral distribution occur.
7. Extract and compare the first three formants with the
segment extracted.
8. Calculate the Euclidean distance between the set of
frequencies obtained from the user, and each of the set
of frequencies corresponding to the vowels.
9. Use the minimum distance criterion to determine the
vowel.
The vowel pronunciations taken are listed below with their
Formant frequency which are calculated through the formant
frequency code:
 I ≈ IY = [255 2330 3000];
 I ≈ IH = [350 1975 2560];
 E ≈ EH = [560 1875 2550];
 A ≈ AE = [735 1625 2465];
 A ≈ AA = [760 1065 2550];
 O ≈ AO = [610 865 2540];
 U ≈ UW = [290 940 2180];
 U ≈ UH = [475 1070 2410];
 A ≈ AH = [640 1250 2610];
TABLE I. VOWELS POTENTIALLY SPOKEN BY NON-NATIVE ENGLISH
SPEAKER GROUPS
it’s a pe ace ful ex is te en ce
IY UH EH AE UH EH IY EH EH IH
IH AH IH IH UW AE IH UH UH IY
EH AA ~ EH ~ ~ EH ~ IH ~
~ ~ ~ ~ ~ ~ ~ ~ IY ~
C. Vowel Identification Results
We compared the similarities as well as differences in the
way vowels are spoken by each of the four groups of speakers.
We present the performance results of the proposed vowel
identification algorithm and findings of our phonology
analysis below. The sample speech used in the tests was,
“It’s a Peaceful Existence.”
In this sentence the possible vowel sounds to be recognized
are shown in Table 1. As non-native English speakers have
their own accents, we expected to recognise these vowels due
to potential inaccurate or incorrect pronunciations of the
speakers.
The algorithm mentioned above was run on all ten speech
samples for each of the four speaker groups, totalling forty
speech samples in the corpus. The success rate was calculated
using the matching percentage of results. One example for
each of the groups, after running the code, for the first vowel
“It’s” (IY, IH, EH) is given in Table 2 to Table 5. The success
rates for recognising ‘It’s’ as vowels are 83.3%, 86.6%,
33.3%, and 46.1% for Chinese, Indian, Japan, and Korean
speaker groups, respectively.
Similarly, the success rates of recognising all the vowels in
the sentence for all forty speech samples for each speaker
group are summarised in Table 6.
D. Observations of Individual Languages
From the above results, our algorithm is the most
successful in recognising the vowels ‘a’ and f‘u’l. The main
reason was that these syllables are spoken with the highest
stress in the sentence. Our algorithm is the least successful in
identifying the vowel sounds ‘ex’ and ‘pe’. The main reason
was that the Euclidean distances of these syllables were more
similar to the vowels immediately after (‘is’) or before (‘a’)
them, which misled the algorithm to choose the wrong vowels
instead. For the 40 speech samples the average success rates is
≈ 57.3%. As the number of speech samples increases, the
probability of successful identification by the algorithm also
increases, hence increasing the success rate.
TABLE II. RESULTS OF CHINESE SPEAKERS
Speaker Vowel Detected
Matching?
( yes/no)
Matching
Percentage
Chinese 1 IY Yes 100%
Chinese 2 EH Yes 100%
Chinese 3 IY, IY Yes 100%
Chinese 4 EH, EH Yes 100%
Chinese 6 No Vowel No 0%
Chinese 8 IH Yes 0%
Chinese 9 UH No 100%
Chinese 10 EH Yes 100%

ISSN (Print): 2204-0595
TABLE III. RESULTS OF INDIAN SPEAKERS
Matching?
( yes/no)
Matching
Percentage
Indian 1 IY Yes 100%
Indian 3 IY,IH Yes 100%
Indian 5 IY,IY,IY yes 100%
Indian 6 IY,IY yes 100%
Indian 7 IY, yes 100%
Indian 8 IY yes 100%
Indian 9 AE, IH Partial yes 50%
Indian 10 No vowel, IH Partial yes 50%
TABLE IV. RESULTS OF JAPANESE SPEAKERS
Matching?
( yes/no)
Matching
Percentage
Japanese 1 IY, No Vowel Partial Yes 50%
Japanese 2 UH, No Vowel No 0%
Japanese 3 UH, No vowel No 0%
Japanese 4 EH, IY Yes 100%
Japanese 5 AH, UH No 0%
Japanese 6 No vowel, UH No 0%
Japanese 7 IY Yes 100%
Japanese 8 UH, UH No 0%
Japanese 9 IY Yes 100%
Japanese 10 IY, UH Partial yes 50%
TABLE V. RESULTS OF KOREAN SPEAKERS
Matching?
( yes/no)
Matching
Percentage
Korean 1 IY yes 100%
Korean 2 No Vowel No 0%
Korean 3 EH, Yes 100%
Korean 4 IY Yes 100%
Korean 5 UH No 0%
Korean 6 IY,UH Partial Yes 50%
Korean 7 UH No 0%
Korean 8 No vowel, UH No 0%
Korean 9 IY, IY Yes 100%
Korean 10 UH No 0%
TABLE VI. SUCCESS RATES OF VOWEL IDENTIFICATION ALGORITHM
Speaker its a pe Ace Ful Ex is te en ce
Korean 46.1% 100% 10% 30% 75% 55% 43% 70% 70% 64%
Chinese 83.3% 86.2% 20% 64% 87% 34% 20% 60% 60% 20%
Japanese33.3% 60% 45.5% 56% 73% 0% 68% 60% 60% 68%
Indian 86.6% 80% 40% 60% 68% 46% 55% 30% 30% 56%
Average
Success
Rate
62.6%81.55%42.95%52.5%75.75%38.75%51.5%57.5%57.5%52%
TABLE VII. STRESS ENENGY LEVEL OF A CHINESE SPEAKER
Vowel Detected Stress Energy(db) Time of occurrence
EH 0.000454969 0.1 sec
EH 0.000611704 0.281429 sec
UW 0.000880056 0.46 sec
IY 0.000850723 0.644286 sec
IY 0.000387354 0.725714 sec
AH 0.000483724 1.00714 sec
IY 0.00086256 1.18857 sec
IH 0.000555674 1.37 sec
IY 9.24567e-005 1.55 sec
IY 0.000111215 1.732 sec
No vowel
UW 0.000183622 2.09571 sec
UW 0.00019356 2.27714 sec
UW 0.000205966 2.45857 sec
IY 0.000118864 2.64 sec
IY 0.000835543 2.72143 sec
IH 0.000378061 3.00286 sec
IH 7.88983e-005 3.18429 sec
UW 0.000179787 3.36571 sec
After looking at the overall results, we have further
analysed the following three aspects of each of the four
speaker groups:
1. Vowels identified, with stress levels and the times of
occurrence. An example of a Chinese speaker is
shown in Table 7.
2. Normalised signal graphs of the individual vowels.
The graphs of all four speaker groups are shown in
Fig. 1, 3, 5, and 7.
3. Stress energy graphs of the individual vowels. The
graphs of all four speaker groups are shown in Fig. 2,
4, 6, 8.

ISSN (Print): 2204-0595
1) Chinese
 For Chinese speakers, stressed syllables tend not to
stand out, due to the fact that fully-stressed syllables
often occur together in the Chinese language.
 They pronounce the /UW/ /UH/ /EH/ for a long
duration (Vowel a).
 There is no distinction between /EH/ and /IY/
2) Indian
 The general pattern of the Indian speakers is that their
vowels are spoken clearly, and stress on /UW/ is high
compared to the other languages.
 The distinction between /IY/ and /UH/ is minimal.
 More vowel sounds were detected, indicating the
duration for spoken vowels is longer.
3) Japanese
 Japanese speakers tend to spend equal time on each
syllable when speaking English, due to the fact that
syllables occur at regular time intervals in the Japanese
language.
Fig. 1. Normalized Signals of the Vowels Identified of Chinese Speaker
Fig. 2. Stress Intensity Graph of the Vowels Identified of Chinese Speaker
Fig. 3. Normalized Signals of the VowelS Identified of Indian Speaker
Fig. 4. Stress Intensity Graph of the Vowels Identified of Indian Speaker
 Vowel A is pronounced as /EH/ and never like /UW/
 Similar to other speakers there is no distinction
between /EH/ and /IY/
4) Korean
 Koreans in general speak every character very fast, thus
resulting in sounding of vowels very similar.
 Stress level is high on /IY/
 There is no distinction between /UH/ and /IY/
 There is no distinction between /EH/ and /IY/
E. Phonlogy Analysis
Observations made on the four languages are presented
below. Our findings are generally consistent with existing
knowledge in linguistic research.
1) Chinese Speakers
 Chinese speakers find it difficult to differentiate
between the /l/ and /r/

ISSN (Print): 2204-0595
 Chinese speakers often cannot pronounce /t/ ,instead
they say it as a /θ/
 Chinese speakers often cannot differentiate between /s/
and / ʃ /
 In English, stressed syllables often stand out. In
Chinese, full syllables often occur together, so stressed
syllables often do not stand out [4]. When speaking in
English, Chinese speakers often do not put enough
emphasis on stressed syllables.
 Chinese speakers tend to have unnecessary aspiration
of plosives after /s/.
 Chinese speakers tend to speak the long paired vowels
and short pair vowels in a similar way, for example, /i/
in did and /i:/ deed are normally pronounced as the
same sound-long vowel /i/.
 Both /u/ and /u:/ in pull and pool are normally
pronounced as long vowels sound /u/.
 Chinese speakers tend to replace/h/ with /x/ in most
cases even though they sound very different.
2) Indian Speakers
 Indian speakers use /e/, /o/ complementarily with the
more common /i/, /u/.
 Indian speakers tend to pronounce /a/ in variation with
the rounded and more back /ɒ/.
 Indian speakers cannot make a distinction between
/ɒ/ and /ɔː/.
 Indian speakers have a more American style of talking.
They pronounce /a/ instead of rounded / ɒ/ or /ɔː/.
Fig. 5. Normalized Signals of the Vowels Identified of Japanese Speaker
Fig. 6. Stress Intensity Graph of the Vowels Identified of Japanese Speaker
Fig. 7. Normalized Signals of the Vowel Identified of Korean Speaker
Fig. 8. Stress Intensity Graph of the Vowels Identified of Korean Speaker
 Standard Hindi speakers have a difficulty in
differentiating between /v/ and /w/.
 Indian speakers have difficulty in pronouncing the
word <our>, saying [aː(r)] rather than [aʊə(r)]. They
tend to miss /u/ in their lingual.
 Indian speakers tend to unaspirate the
voiceless plosives /p/, /t/, /k/ . The ‘h’ (exhalation) is
generally missing.
 Indian speakers tend to allophonically change the
/s/ preceding alveolar /t/ to [ ʃ ] (<stop>/stɒp/ → /
ʃʈap/).
 Indian speakers tend to used /z/ and /dʒ/
interchangeably.
 Both /θ/and /ð/ are pronounced like /t/. Indians tend to
replace /pʰ/ with /f/.
3) Japanese Speakers
 For Japanese speakers, slicing phonetic information
remains, for example, ("Su" becoming "s"). In this case
the small fragment of phonetic sound ("u" sound) is
missing, leaving behind awkward vocal sounds.
 In Japanese [v V] /vʌ/ sound closer to the actual
pronunciation of [w a] /wa/

ISSN (Print): 2204-0595
 English has a variety of fricatives and affricates which
are more widely distributed than in Japanese [5]. The
Japanese consonantal system does not have /f/, /v/, /θ /,
/ð/, / ʃ /, / ʒ /, /ʧ/, and /ʤ/.
 Similar to Chinese speakers, the Japanese speakers tend
to say the /r/ sound as /l/.
 For Japanese speakers, the time it takes to say a
sentence depends on how many syllables in the
sentence, but not on how many stressed syllables, as it
should be spoken properly [5]. This affects the way that
Japanese speakers speak English.
 In Japanese, /ʃ/ /ʒ/ and /ʧ/ /ʤ/ do not usually appear as
distinct phonemes. When /s/ /z/ and /t/ /d/ appear
before the vowels /I/ and /U/, Japanese speakers tend to
allophonically pronounce /ʃ/ /ʒ/ and /ʧ/ /ʤ/ instead.
4) Korean Speakers
 Japanese and Korean seem like allophones and thus
have a pitch difference. Japanese tend to speak /t/ less
sharply as Koreans.
 Some English vowels are not available in Korean, such
as /i/ /v/ / æ/ /ej/ /ow/ /ɔ/ /ə/ / ʌ/ /a/. Thus the
pronunciation of Koreans becomes unclear in English.
 Korean speakers tend to pronounce /m, n/ without the
nasal effect.
 The phoneme /ŋ/ appears frequently between vowels.
 Koreans tend to say /p/ as /b/, /t/ as /d/, / tɕ/ as /dʑ/, and
/k/ as /g/.
 Koreans tend to pronounce /n/ and /l/ similar to /l,
hence both /nl/ and /ln/ are pronounced as [lː].
F. Collection of Speech Corpus
To execute the project the most vital part of the software
was the selection of speech corpus. To get reliable results
having a vast range of speech samples is extremely important.
And more the range the standardization of the samples
actually contributes positively to the results. These samples
can later be used for testing of other speech software and in
conducting linguistic studies. These samples helped in
comparing the similarities and differences in the above
mentioned analysis of languages. The collection of samples
was saved as .wav files. The methodology of sample
collection is given below:
1. The sentence chosen for the analysis was ‘It’s a
Peaceful Existence’. The reason this sentence was
chosen was because it consisted of very distinctive
vowels which could certainly help in confirming the
vowel identification method. Also in addition it has
alphabets which can be used as distinguishing factors
for different speaker groups.
2. Forty students from the National Univeristy of
Singapore, ten from each of the four nationalities
(China, India, Japan, Korea), were recruited to
produce speech samples.
3. All speech samples were checked for discrepancies,
and were saved as .wav files and filed in folders
according to the native language of the speaker.
4. The sample locations can be stored in an SQL
database, and linked and accessed by the codes.
The above analysis can be used for several applications
such as speech recognition, linguistics research, and studying
ESL. It can also help understand how the speaker’s own native
language affect their pronunciations, and therefore, address
their issues more directly.
III. CONCLUSION AND FUTURE WORK
In this paper, we have presented the results of using
MATLAB’s Speech Analysis Toolbox for speech feature
extraction and data visualization. There are two main areas in
this study. In the first part of this work, the formant frequency
algorithm was implemented to identify vowels and a success
rate of 57.3% was achieved. We anticipate better results with
more speech samples in the future. In the second part, we
performed a phonological analysis of four Asian languages:
Indian, Chinese, Korean, and Japanese. We have provided a
general cover graph for every feature of each of the languages,
and demonstrated the differences and similarities of the four
languages.
This work can be extended in a number of areas. Firstly,
more speech features such as Perceptual Linear Prediction
(PLP), Relative Spectral PLP can be implemented. Secondly,
more languages can be included in the study. Thirdly,
automatic recognition systems can be developed based on
feature extraction for different purposes. For example, an
accent recognition system can be developed using speech
extraction and language analysis. Finally, visualization can be
improved so the results can be understood better and easier.
REFERENCES
[1] A. Biem, S. Katagari, and B. H. Juang, “Discriminatiove Feature
Extraction for Speech Recognition,” in Proc. 1993 IEEE-SP Workshop,
pp. 392-401.
[2] L. R. Rabiner and R. W. Schafer, Theory and Application of Digital
Speech Processing, 1st
ed. Pearson, 2010.
[3] B. Prica and S. Ilić, “Recognition of Vowels in Continuous Speech by
Using Formants,” Facta universitatis-series: Electronics and Energetics,
vol. 23, no. 3, 2010, pp. 379-393.
[4] K. Brown, Enclopedia of Language and Linguistics, 2nd
ed. Elsevier,
2010.
[5] K. Ohata, “Phonological Differences between Japanese and English:
Several Potentially Problematic Areas of Pronunciation for Japanese
ESL/EFL Learners,” Asian EFL J. vol. 6, no. 4, 2004, pp. 1-19
[6] A. Spanias, T. Painter, and V. Atti, Audio Signal Processing and
Coding. Hoboken, NJ: John Wiley & Sons, 2006.
[7] D. O’shaughnessy, Speech Communication: Human and Machine.
India:University Press, 1987.
[8] L. Deng and D. O’shaughnessy, Speech Processing: a Dynamic and
Optimization-Oriented Approach. New York, NY: Marcel Dekker Inc,
2003.
[9] J. Baker, L. Deng, J. Glass, S. Khudanpur, C. H. Lee, N. Morgan, and D.
O’Shaughnessy, “Developments and Directions in Speech Recognition
and Understanding, Part 1,” Signal Processing Magazine, vol. 26, no. 3,
pp. 75-80, 2009.
[10] Encyclopedia Britannica. (2014). Pitch [Online]. Available:
http://www.britannica.com/EBchecked/topic/1357164/pitch

ISSN (Print): 2204-0595
[11] M. P. Kesarkar, “Feature Extraction for speech Recognition,” M. Tech.
Credit Seminar Report, Electronic Systems Group, EE. Dept, IIT
Bombay, Nov. 2003.
[12] B. S. Atal and S. L. Hanauer, “Speech Analysis and Synthesis by Linear
Prediction of the Speech Wave,” J. Acoustical Society of America, vol.
50, no. 2B, 1971, pp. 637-655.
[13] T. L. Nwe, S. W. Foo, and C. R. De Silva, “Detection of Stress and
Emotion in Speech Using Traditional and FFT Based Log Energy
Features,” in Proc. 4th
Int. Conf. Information, Communications and
signal Processing, Singapore, 2003, pp. 1619-1623.
[14] A. V. Oppenheim, R. W. Schafer, and J. R. Buck, Discrete-Time Signal
Processing. Upper Saddle River, NJ: Prentice Hall, 1999.
[15] Punskaya, E, Basics of Digital Filters [Online]. Available:
http://freebooks6.org/3f3-4-basics-of-digital-filters-university-of-
cambridge-w7878/
[16] L. R. Rabiner and R. W. Schafer, Digital Processing of Speech Signals,
Englewood Cliffs, NJ: Prentice-Hall, 1978.
[17] S. W. Smith, The Scientist and Engineer's Guide to Digital Signal
Processing, California Technical Publishing, 1997.
[18] C. E. Shannon and W. Weaver, The Mathematical Theory of
Communication, Urbana: the University of Illinois,1964.
[19] U. Zölzer, Digital Audio Signal Processing. John Wiley and Sons, 2008.
[20] D. P. W. Ellis. (2008, October 28). An Introduction to Signal Processing
for Speech [Online]. Available: a
http://www.ee.columbia.edu/~dpwe/pubs/Ellis10-introspeech.pdf
[21] J. Cernocky and V. Hubeika. (2009), Speech Signal Processing –
introduction. [Online]. Available: a
http://www.fit.vutbr.cz/~ihubeika/ZRE/lect/01_prog_intro_2008-
09_en.pdf

Speech Feature Extraction and Data Visualisation

More Related Content

What's hot (18)

Similar to Speech Feature Extraction and Data Visualisation (20)

More from ITIIIndustries (20)

Recently uploaded (20)

Speech Feature Extraction and Data Visualisation