DIALECTAL VARIABILITY IN SPOKEN LANGUAGE: A COMPREHENSIVE SURVEY OF MODERN TECHNIQUES FOR LANGUAGE DIALECT IDENTIFICATION IN SPEECH SIGNALS

Jilin Daxue Xuebao (Gongxueban)/Journal of Jilin University (Engineering and Technology Edition)
ISSN: 1671-5497
E-Publication: Online Open Access
Vol: 42 Issue: 11-2023
DOI: 10.5281/zenodo.10203386
Nov 2023 | 459
DIALECTAL VARIABILITY IN SPOKEN LANGUAGE: A
COMPREHENSIVE SURVEY OF MODERN TECHNIQUES FOR
LANGUAGE DIALECT IDENTIFICATION IN SPEECH SIGNALS
Er. POONAM KUKANA
Department of Computer Science and Engineering, University School of Engineering & Technology, Rayat
Bahra University, Mohali, Punjab, India. Email: poonamkukana@gmail.com
Er. SUKHJINDER KAUR
Bahra University, Mohali, Punjab, India. Email: skaur29100@gmail.com
Dr. PUNEET SAPRA
Bahra University, Mohali, Punjab, India. Email: puneetsapra91@gmail.com
Er. CHIMAN SAINI
Bahra University, Mohali, Punjab, India. Email: chimansaini1994@gmail.com
Abstract
Main fundamental challenge for recent research work on speech based on science and technology is to
understand and model the user variants in Spoken Languages. Users have their style of speaking, reliant
on various factors, adding the dialect and accent of the speaker as well as the social and economic
background of the speaker and contextual attributes like degree of knowledge between the listener, speaker
and the position or rank of the speaking condition, from very normal to formal. In the past few decades, an
extensive progress has been seen in automatically verifying the language of a speaker offered a sample
speech. The main purpose of dialect verification is the recognition of a speaker’s region dialect, within a
pre-determined language, offered the acoustic signal alone. DR (Dialect Recognition) is a main issue in
particular, since even within the similar dialect and accent or register user change may occur. For
illustration, In Spontaneous speech, few speakers tend to exhibit more optimizing and alteration of function
words than others. The main issue of dialect recognition system has been viewed as challenging than that
of language classification or recognition due to the maximum similarity among dialects of the similar
language. While, dialects may differ in any dimensions of the linguistic spectrum such as syntactic, lexical,
morphological, phonological differences, these changes are likely to be more indirect across dialects than
those across languages such as Hindi, Punjabi and English etc.
Keywords: DR (Dialect Recognition), DI (Dialect Identification) ASR (Automatic Speech Recognition),
MFCC (Mel Frequency Cepstral coefficient), Linear Discriminant analysis (LDA), (LPC) Linear Prediction
Coefficient, Probable Linear Discriminate Analysis (PLDA), Relative spectra (RASTA) filtering, FFMP (feed
forward multilayer perceptron), CNN (Convolution Neural Network), RNN (Recurrent Neural Network).
I. INTRODUCTION
Speech is the method of expressing the ideas, expressions or thoughts. Speech is used
in the daily conversation. It is the kind of the conversation [1]. Speech plays an essential

ISSN: 1671-5497
Vol: 42 Issue: 11-2023
DOI: 10.5281/zenodo.10203386
Nov 2023 | 460
role in number of actions include socialization. Speech or Language is related to the
method of communication used by social beings [2]. It is the method of expressing the
ideas and thoughts through vocal sounds. A dialect is mainly a specific kind of the
language that specified to area or the human set [3]. It mainly has variations in
pronunciation, grammar, syntax, vocabulary. It is not easy to understand because dialect
is vocal by living people in specific location [4]. A speech is dialect with the defense forces.
The variation among the language and dialect may arrange the variation among speech
and dialect. The selection is known as speech that is connected for the identification of
edges of the countries, and person determines the speech and following consideration
[5].
Speech is one of the essential equipment for the communication among the social being
and surroundings. Consequently, an automatic recognition system is done for need of the
humans. Speech recognition system is used as distinct and regular models that are reliant
or non-reliant on humans [6]. Different methods manage a distinct acoustic method for
every word, association of the words that is related to different word speech recognition.
A speaker-based network needs that the user traces an instance of the word, sentence
or phrase previous to detection by the network [7].
The speech analysis is done through the speech input. Then the feature vector values
are determined. After that, the data processing is done from the selected feature vector.
The resulted speech recognition is achieved from the reference model. The process of
the automatic recognition system is described as;
Figure 1: Speech Recognition System [7]
It is done in training and recognition stage. During the training stage, the pre-processing
and recording of the speech is done where the feature extraction take place. In other
phase, acoustic study of the indefinite signal is done in recognition stage.

ISSN: 1671-5497
Vol: 42 Issue: 11-2023
DOI: 10.5281/zenodo.10203386
Nov 2023 | 461
1. Feature Extraction: In this stage, noise robust evaluation is probable on the high
correlation coefficients. However, the minimum value of the noise speech signal
autocorrelation is related to the elimination of the major noise elements [8].
2. Acoustic Model: It is the main element of the automatic recognition system that is
responsible for the high load and performance of the network. It is established for
identifying the vocal fact. The establishment includes the usage of the audio database
of the speech and the text of the scripts. After that, compile the data into statistic
demonstration of speech that construct the words [9].
3. Lexical Method: The accent of every word in required language is provided where
lexicon is established. The different linking of the phones is determined by lexicon
method for providing the suitable words.
4. Language Approach: It is the main element operates on large number if the words,
comprising of large metrics and through accent vocabulary, created words series in
sentence.
II. DIALECT IDENTIFICATION IN THE LITERATURE
A. From a linguistic point of view
The most difficult part of speech recognition is dialect, which refers to the language
features of a given geographical population. The lack of databases and the time-
consuming analytical method continue to hamper research in dialect linguistics. As
technology has improved, a robust speech recognizer capable of dealing with unstable
settings such as noise and accent fluctuatsion has become an imperative requirement.
Dialects are not fixed; they fluctuate with location and time. In other words, dialects may
differ owing to language changes throughout time, just as they do with physical borders.
Today, we find that new generations employ terms and language that were not previously
available or used. As a result, languages evolve with time, as do their dialects.
Furthermore, each language has multiple dialects. The United States, for example, has
one dialect border that runs between the North and South along what is known as the
Mason-Dixon Line. American English, on the other hand, has a variety of dialect barrier.
There are several dialects in each of the fifty states, and each state may have a multiple
dialect.
1) DI Early Studies
Dialect research began in 1877, when George Wenker conducted a series of surveys to
define dialect areas Then Bailey made one of the initial attempts to describe the Midland
dialect and determine whether or not it exists. The study concluded that distinguishing
dialects should not be based on vocabulary because it varies depending on group or class
within the same geographical region. Davis and Houck were also attempting to determine
if the Midland area may be considered as a separate dialect region or not. Their research
was effective in extracting phonological and lexical data from 11 cities along a north-south
axis. The result was that the Midlands area cannot be called an intermediate region and
that there's a linear link between the dialects of the far South and the dialects of the

ISSN: 1671-5497
Vol: 42 Issue: 11-2023
DOI: 10.5281/zenodo.10203386
Nov 2023 | 462
Southern region. In contrast to the latter idea, Johnson demonstrated that integrating
phonological and lexical variables is inaccurate since it negatively impacts data patterns
and hence generates incorrect findings. Some terms were used to demonstrate that there
is a distinct difference in the dialects of the North and the Midlands, as well as the dialects
of the South and the Midlands, but not among the dialects of the North and the South.
The result was that the Midland area is distinct from both the North and South regions.
These were the initial steps in recognizing and categorizing dialects. The second stage
was to establish how much dialects correspond to each other after defining them. What
are the similarities and differences across dialects of an identical language, and via what
method can they be evaluated?
2) DI using Vowels
Peterson and Barney made one of the earliest efforts, focusing on the vowel spacing
feature. They discovered that perceptually distinct vowels occupy various areas in
formant space; also, the identical vowel produced by different persons appears in varied
spots in formant space. Finally, it was determined that participants' capacity to create and
interpret a particular vowel is influenced by their history. This work was noteworthy in the
field of recognition since it was the very first to emphasize the significance of dialects.
The one disadvantage of this investigation was the variability of individuals, which made
proper characterization of vowel spacing alterations difficult. As a result, Hillenbrand,
Clark, and Wheeler performed the same tests to achieve more precise findings for vowel
gap changes using a single homogenous class. More vowels and diphthongs have been
introduced to those covered in Peterson and Barney's study.
In addition, for each vowel, the spectral shifts, its duration, which stem and stable state
F1, F2 were measured. All of these efforts were made in order to deliver more accurate
and up-to-date data. Michigan was represented by the contestants. According to the data,
the individuals had identical vowel locations but with less vowel spacing in the vowel
space. The study concluded that lowering the distance between vowels had no effect on
their perception. Furthermore, our findings corroborated prior findings that F1 and F2
measures are insufficient to characterize vowel spacing.
Hajiwara duplicated the findings of Peterson, Barney, Clark, and Wheeler. The
contestants this time were from Southern California. Hajiwara was on the lookout for
formant modifications in this new dialect. Hajiwara's research discovered that a common
dialect like Southern California seldom generates a fully rounded vowel, which was the
major cause for having greater F2 for some vowels. This study discovered a new dialect
(South Californian) that may be compared to existing dialects. Furthermore, Hajiwara
thought that further study for different dialects within a single group should be done in the
future by expanding the number of of both genders’ participants.
3) DI using Consonants
Consonants, on the other hand, have been identified as dialect informational identifiers
due to their ability to disclose accents from other countries and social status. This
prompted William Labov to conduct sociolinguistic research on accent variations. As a

ISSN: 1671-5497
Vol: 42 Issue: 11-2023
DOI: 10.5281/zenodo.10203386
Nov 2023 | 463
new measuring metric that indicates personal origin, he employed rhoticity: the sound of
'r' when it occurs after a vowel (as in bar, sort, churn) and is known as postvocalic 'r'.
Rhoticity was seen as a low-status trait in the United Kingdom, but it was regarded as an
important characteristic of pronouncing in the United States. This made distinguishing
between American and British English simple. As a result, Labov conducted the first
sociolinguistic study, demonstrating that New York accent differences center around
postvocalic "r" use. He began with a brief poll to assess the dependability of the fresh
testing procedures. Labov did not interview speakers as previous studies had done;
instead, he wandered through three Manhattan department shops pretending to be a
consumer. He began by checking what things were on the 4th floor of each store, and
then he asking the sales associates where he could find these items. In each store, he
repeated the experiment. He chose the fourth level because it had two tokens of
postvocalic "r," and by appearing not to hear, he got every participant to say the two
syllables a second time, once impulsively and once attentively, resulting in four tokens
from each informant. The collected data revealed that Labov's theory was correct, and
that the postvocalic "r" fluctuates depending on social class, speaking style, and linguistic
context connected with each store's customers. Higher and lower-ranking employees
were readily identified in this study by evaluating the post-vocalic "r" pronunciation. Labov
demonstrated that the rhotic usage of the letter "r" signified social status and aspiration
from the as well as that it was more prevalent in younger speakers.
According to Labov's research, the New York accent changed towards the end of WWII.
Furthermore, varieties of American English lack the sound of the post-vocalic "r". This
approach was a simple and effective way to detect American accents. Through
sociolinguistic investigations, Labov was the first to develop new measurement methods
for identifying dialects.
To summaries, the past studies were experiments to assign the basic building pieces for
Dialect Identification (DI) research. The research was largely effective in demonstrating
that vowel gap and consonants are both crucial factors in categorizing dialects. Following
the discovery that acoustic properties varied across dialects, researchers conducted
investigations to uncover how these differences differ between regional dialects as well
as between various genders and ages. This prompted Byrd to investigate variances in
the TIMIT corpora while categories utterances based on area and gender. TIMIT contains
630 female and male speakers reflecting all eight American dialects. The study sought to
identify gender variations in glottalization, number of the flaps, palatalization, central
vowels, and speaking pace. The results revealed that the release stop, glottal stop, and
vowel contraction rates varied amongst dialects, but the number of flaps did not.
Furthermore, men tend to talk quicker than women, with fewer stressed vowels. However,
there was no substantial difference in dialects, and it was suggested that this might be
due to a lack of data for both genders. This demonstrates that there are several
parameters/features that might influence dialect variations. It is not a simple process to
extract these parameters. Furthermore, because this is a database-dependent problem,
it is hard to pinpoint the specific characteristic that is that causes a particular dialect shift.
A useful database must include individuals of all ages, both genders, and various

ISSN: 1671-5497
Vol: 42 Issue: 11-2023
DOI: 10.5281/zenodo.10203386
Nov 2023 | 464
affiliations. This useful database, however, is still missing other hidden characteristics that
change with dialect, like moods and physical state.
4) DI using Acoustic and Phonetic Features
Clopper acknowledged and recognized six American dialects in some successful
attempts. The acoustic aspects of a dialect were researched initially, followed by the
perceptual features of listeners. In this paper, several words that differ between dialects
were chosen and their acoustic-phonetic features were investigated. Before proceeding
with the overview of both articles, it is useful establishing certain key parameters.
Fricatives are consonants generated by pushing air through a small channel formed by
the close proximity of two articulators. In the instance of [f], the articulators are the lower
lip and the upper teeth. Frication refers to turbulent wind. Another crucial criterion is vowel
back-ness, which refers to the location of the tongue during vowel articulation in relation
to the back of the mouth. However, vowels are classified as rear or front based on the
frequency of the second formant (F2) rather than the actual articulation. The higher the
F2 value, the closer the vowel is to the front; the lower the F2 value, the further back the
vowel is. Di- pthongization, also known as vowel breaking, is the transformation of a
mono-phthong into a di-phthong or tri-phthong. Di-phthongization refers to the
transformation into a di-phthong. Nagy and Zhang were interested in novel metrics such
as r- fullness to distinguish between rhotic and non-rhotic dialects, vowel brightness to
check for "r" insertion in a word, and fricative voicing and length to determine if individuals
pronounce words with fricatives. Experiments were conducted after evaluating and
determining the above listed factors to determine which qualities listeners utilize for
recognizing a dialect. According to the findings of these investigations, listeners usually
employ four characteristics in perceptions and dialect identification. R-fullness,
blackness, vowel brightness, and di-pthongs were the four criteria. Listeners were also
able to distinguish the dialects of the South and New England from others. The capacity
of an individual listening to recognize a dialect is determined by where he or she has lived
throughout his or her life.
Later, in another study, the goal was to provide a thorough characterization of major
American English dialects utilizing the previously extracted acoustic-phonetic elements.
The unexpected finding here was the variation of traits across the same dialect. It is hard
to generalize characteristic variation on a specific dialect group. Unexpected
characteristics among Southern males included fronting and rising of the "u," as well as
merged vowels on occasion. Finally, it was concluded that there exist phonological
distinctions across dialects and that we can determine a dialect based on well-chosen
groupings of strongly correlated lexical sets. Human perceptions of dialects use the same
processes, but occasionally fails owing to dialect variance within a single group.
5) DI using Words and Lexical Sets
Wells chose several terms and defined them as sets with an unchanging pronunciation
pattern, so supporting the same premise that certain words are ideal dialect identifiers. A
set of keywords was created, each with its own lexical set. A lexical set is a group of
words that all have the same vowel pronunciation. During studies, it was discovered that

ISSN: 1671-5497
Vol: 42 Issue: 11-2023
DOI: 10.5281/zenodo.10203386
Nov 2023 | 465
two lexical sets blended inside an accent. As a result, the finding was that a dialect may
be distinguished using certain mergers and lexical set pronunciation.
Zhang and Nagy investigated the reciprocal information between distinct lexical sets in
order to classify dialects based on phonological characteristics. This approach was
evaluated on 168 binary parameters defining the pronunciation of English vowels and
consonants for speakers from 35 different nations. This approach yielded clusters that
were comparable to those created by typical clustering methods. A comparison of
clustering approaches was also performed, and the variations in clustering effects were
investigated. This study was crucial in giving a succinct description for several English
dialects, as it identified all lexical set variances.
Kessler's approach of grouping by analyzing phonetic transcription was another. The
approach, like the isoglosse method, was based on determining linguistic distance
between two places. Isoglosses were the usual way for determining dialect borders. This
was accomplished by drawing border lines between locations with a group dialect;
individuals who pronounce the same manner. The disadvantage of isoglosses is the
inconsistency of border lines caused by lexical feature change within the same area. The
baseline metric and the Levenshtien metric were the new distance metrics used to
aggregate phonetic measures. In the baseline measure, two locations on the map had
"0" distance for phone strings that were similarly comparable, and "1" distance otherwise.
The Levenshtien range between phonetic string was measured in the Levenshtien metric,
where the Levenshtien range is defined as the most inexpensive sequence of deletions,
insertions, and substitutions required to transform one phone string into another. The
phonetic and features string comparisons were utilised; in the former, each operation
turns one phone string into another phone string, when in the latter, each operation
changes the characteristic feature. The distance was calculated by averaging the
distances between several landmarks. The ultimate finding was that the phone string is
superior for automated dialect identification grouping. Furthermore, the term was defined
as an ADI unit that may be separated into string to get exact measurements.
The last DI approach is that supplied by Huckvale, who proposed ACCDIST, a new
measure for the quantitative evaluation of a speaker's accent similarities. The idea behind
ACCDIST is that the author was able to group together accents of the speakers by storing
the inter-segment distance measurements for each speaker in a table and then detecting
autocorrelation measurements between the data collected and the ones under test. The
ACCDIST speaker classification approach captures a speaker's pronunciation qualities
rather than voice parameters.
III. SPEECH DIALECT SYSTEM
In speech dialect recognition scheme, the recognition of the dialect or accent is done
through the training speech utterance of special dialect utilizing the single phone detector.
Speech recognition is the capability of the program to recognize the words or phrase in
spoken language and transform them machine readable format. In the given figure, the
speech input is analyzed for the selection of the template. After that the features are
extracted where the decoded procedure is done through the different models. The models

ISSN: 1671-5497
Vol: 42 Issue: 11-2023
DOI: 10.5281/zenodo.10203386
Nov 2023 | 466
are acoustic, linguistic or pronunciation. The final output value is achieved or
pronunciation. The final output value is achieved.
The main types of the speech dialect system are described as;
1. Acoustic Phonetic Approach: In this method, the different phonetic unit in vocal
speech and the speech is measured through the group of the acoustic features. The
features of the phonetic elements ate mainly changeable with speaker and closest
sound. It is unspoken in this method that regulates the leading the changeability are
simple and educated through equipment [10]. The initial stage of this method is
language spectral study. In the other stage, the conversion of the spectral dimensions
in to group of the feature is done during feature extraction that explained the acoustic
features of various phonetic units. The other stage is the division and marking phase
where the language signal is divided in to remote areas follow by linking single or
more markers to every divided area [11]. The final phase searches the suitable word
from phonetic mark series created by segmentation to marking. The input the speech
signal. After that, speech signal pre-processing is experienced. And then the
recognition of the phonetic unit is done by the acoustic modeling. The decoded data
is practiced.
2. Pattern Recognition Approach: This method is used for the extraction of the pattern
dependent on desired situations and to divide single class from other. It is based on
four phases as feature dimension, training and classification the pattern. The test
pattern is defined, a series of the dimensions is completed on inner signal [12]. The
related pattern is developed by acquiring more than pattern related language sounds.
It can be given in the speech pattern. It is analyzed on phrase, speech or word. During
the pattern classification phase of this method, a straight comparison is done among
the undesired test pattern. Input the speech signal and features are analyzed for the
classification of the pattern along with language, word lexicon and acoustic model.
The utterance is verified to receive the final value.
3. Template Method: In this technique, the undesired language is comparable opposite
to group of templates for searching desired match [13]. The group of the prototype
language pattern are recorded as location pattern demonstrating the vocabulary of
words of the applicant. The recognition are done by matching of undesired language
word along with every of the related template and choosing the class of suitable
matching pattern. This technique presents the high recognition rate. Input the speech
signal and the pre-processing and feature extraction is done. Moreover, the template
and pattern classification are done by testing. In addition, acoustic and language are
determined by to get the speech transcript through data classification.
4. Artificial Intelligence Method: It is mixture of the acoustic phonetic method and pattern
detection system. A number of the investigators established detection method
through acoustic phonetic data to establish the classification regulation for language
sounds. Despite the fact that the template dependent technique presents the small
relation about the social speech processing, but these methods are more efficient in
design of the language recognition methods. In contrast, the language and phonetic

ISSN: 1671-5497
Vol: 42 Issue: 11-2023
DOI: 10.5281/zenodo.10203386
Nov 2023 | 467
survey of social language processing is done. On other hand, this method has one
sided achievement because of the complexity to quantify the experience data [14].
IV. INDIAN LANGUAGES
About languages and number’s (Indian constitution recognizes):
Language is normal method and it may not be essential to get direct boundaries that were
simple for the definition and identification. Language is related to communal, educational
factors instead of intrinsic real and realistic computation of language characteristics of
speech [15].
The Indian constitution determines the eighteen representative Indian speech or
language. These languages include various dialect or the changes of that speech. In
addition, the regional languages are considered by central government and worked as
official language. The state limits of the state generated was dependent on restrictions of
the major Indian language that is considered by Indian Constitution.
Dialect is the kind of the speech that is vocal in one region which varies from another kind
of the similar speech (words). It demonstrated the characteristics of the syntax and
terminology and aspect of accent. It is the kind of the language that is vocal in single
region which may change from other kind of similar speech. There is no universal criterion
for differentiating and distinction is subject of point. It is not easy to present direct
demonstration for speech and dialect [16].
The language when utilized by the people from various areas may be analyzed to check
the selection of the words [17]. If some standard forms of the words are spoken then the
variation in the spectral features of sound created. Every person creates a method of
speaking at the initial age. The method is based on the native language. The standard
form of the language is spoken through native language, and then the speaker acquires
the traits of the method. The mode may be affected by the human language or region
surrounding of the speaker. The speaking variations worked as accent. Various accents
provide the variation in the understanding of the utterance. During the acoustic training,
the effectiveness of the automated system recognition has been improved by dialect.
V. FEATURE EXTRACTION TECHNIQUES
Generally, the speech feature extraction is utilized for decreasing the size of the input
vector, whereas the management of the sensitive energy of the signal. Feature extraction
is the specific type of the database and extraction of the specified features are acquired.
The features received structure of the desired data about the speech [18]. The major
kinds of the feature extraction group are MFCC (Mel Frequency Cepstral coefficient),
Linear Discriminant analysis (LDA), (LPC) Linear Prediction Coefficient. However, MFCC
is the major method for feature extraction. Feature extraction is the major kind of the
speech recognition scheme. It is measured as the major part of the network.

ISSN: 1671-5497
Vol: 42 Issue: 11-2023
DOI: 10.5281/zenodo.10203386
Nov 2023 | 468
The main speech signal is forwarded to the per emphasis stage. In this stage, the
decreasing of the maximum frequency region of the acoustic was muted in the social
audio development method [19]. The speech after the pre-emphasis sounds deeper with
less volume rate. The speech signal is forwarded to maximum arrangement filter. In the
frame blocking, an audio digital signal is segmented in to the amount of the frames and
few portions of the frames overlapped to each other. Every frame may be estimated
without the loss of the data. In the given table 1, types of feature extraction techniques,
characteristics, advantages and disadvantages are given.
Table 1: Advantages and Disadvantages of Feature Extraction Techniques.
Feature Extraction
Method
Characteristics Advantages Disadvantages
Linear Predictive
Coding (LPC)
Regression based
speech Feature
Accurate and Robust Method.
Encode speech at less bit rate
Does not differentiate
words with the same
vowels.
MEL Frequency
Cestrum (MFCC)
Used in speech
Processing
Provide high accuracy.
MFCC presents the major
features of the phone of speech.
The performance is
influenced by the
amount of the filters.
Probable Linear
Discriminate
Analysis (PLDA)
Used the variables
based on HMM.
It is flexible acoustic method that
used variable amount of the
related frames.
Class discrimination
may be inaccurate.
Relative spectra
(RASTA) filtering
It is used in the
noisy speech
signals.
It acquires the frequency with
less modulation related to
speech.
An error occurs during
the filtering.
Feature selection is the stage when an automatic or the manual selection of the features
that took place in prediction of the variable or the desired outcome. The unrelated features
reduce the accuracy models. Feature selection is the necessary stage in the
establishment of the network for the speech recognition.
Presently, the interference among the features created from single audio resource is
rarely measured, that creates the redundant features and improves the calculation price
[20]. The selection of the features used for the automated speech recognition influenced
the performance rate. The features are extracted through the input data. After that, the
feature set re arranged with the optimum feature subset. In addition, the speech features
are classified as word, phrase, language or phenomena.
Advantages:
1. Less redundant information with minimum opportunity for the selection based on the
noise and it decreases over fitting.
2. Improved accuracy with minimum misleading information.
3. Reduced training time where the minimum data trained at faster rate.
4. Removing the unrelated and redundant features, the performance of the prediction
model can be improved.
5. The improvement through the elimination of the effect of dimensionality.
6. Improved the general performance rate.

ISSN: 1671-5497
Vol: 42 Issue: 11-2023
DOI: 10.5281/zenodo.10203386
Nov 2023 | 469
7. Improved the learning process.
VI. DEEP LEARNING MODELS
Deep Learning is the method as the subgroup of the machine learning that that is the
internal subgroup of the artificial intelligence. Over the past few decades, deep learning
is the technique of the learning information demonstrations. It may be supervised,
unsupervised or semi supervised. It presents the group of the algorithms and methods
that learns the features and jobs directly from the information [21]. The information may
be structured unstructured that includes the pictures, textual data or the sound. DL is
related as the point-to-point learning for the direct demonstration of the information. In
addition, the deep learning methods worked in absence of the human intervention and it
supports an accurate outcome. DL is mostly utilized in regions such as computerized
vision, natural language processing, pattern and object detection. Generally, the deep
learning methods processed the data in same way to the human brain that may acquire
to multiple job by the people. It is mainly used in image pre-processing, normal language
processing and speech recognition.
The deep learning architecture followed the approach as,
 Build network containing the input and hidden layer along with the required hops.
 Trained the system.
 Addition of the other hidden layer at the higher level of the learned system for creating
a new system.
 Retrain the system.
 Addition of maximum layers and retrains the system.
Advantages of DL:
1. Higher level features are learned.
2. Identify direct relations.
3. High measurement information.
4. Unlabelled information
Some of the methods of the deep learning models are described as [22];
1. Auto-Encoder: It is an artificial neural network that is helpful in different coding
patterns. The normal type of the auto encoder is same as the multiple perceptron
comprising an input layer or hidden layer. The major variation among the multiple
layer perceptron, FFNN and auto-encoder is the amount of the hops at the external
layers. The external layer comprises the similar kinds of the hops as internal layer.
2. Deep Belief Network: It is the output to the issue of management of the non-convex
selective values and local minima through specific multiplayer perceptron. It is the
kind of the deep learning with multiple layers of the latest variables with the
interconnection among the layers. This network may be displayed as the Boltzmann

ISSN: 1671-5497
Vol: 42 Issue: 11-2023
DOI: 10.5281/zenodo.10203386
Nov 2023 | 470
equipment when every hidden layer of the sub system works as observable internal
layer for the adjoining layer of the system.
3. Convolution Neural Network: It is the other alternative of the FFMP (feed forward
multilayer perceptron). It is kind of the FFNN in which single neuron are arranged for
complete overlapping areas. CNN works as successive model of data and linking in
the system. The boundaries of the initial layer are identified and create the templates
for the edge recognition. After that, succeeding layer tried to link in easy design and
templates of the diverse locations. The last layer matched as input picture with
complete templates and last possibility worked as addition of the weighted. Hence,
the deep CNN method provides high exact predictions.
4. Recurrent Neural Network: It worked as the permanent amount of the input value,
creates as static dimension vectors as output value along already defined stages.
This network permits for operating the series of the vectors in input and output. The
interconnection among the units creates as directed cycle in RNN. In this network,
the input and output values are not automated but interrelated. In addition, RNN
connects the normal metrics at each layer. RNN trained in traditional neural system
through back propagation technique.
5. Re-enforcement Learning to NN: This network is the type of the hybridization of the
dynamic programming and supervised learning. The distinctive element of the model
is background, agent, measures, guidelines and cost values. The agents work as the
regulator of the network; guidelines recognize the operations, and output value
determines the selective of the re-enforcement learning issue. This learning provides
the performance with maximum output for the desired stage.
VII. CONCLUSION AND FUTURE SCOPE
The main issue of dialect recognition system has been viewed as challenging than that
of language classification or recognition due to the maximum similarity among dialects of
the similar language. While, dialects may differ in any dimensions of the linguistic
spectrum such as syntactic, lexical, morphological, phonological differences, these
changes are likely to be more indirect across dialects than those across languages such
as Hindi, Punjabi and English etc.
DI mainly helps in ASR system since speakers with various dialects pronounce few words
like differently, consistently changing certain phones, and even morphemes. This is proof,
like example with the Hindi Language, which has various variants the formal written SL
(Standard Language) of the education, media and culture and information spoken dialects
that are the preferred technique of communication in a daily life. Although, there are
commercially available ASR (Automatic Speech Recognition) systems for recognizing
few standard dialects with minimum error rates, these recognizers failure, when a native
Hindi speaker speaks in her/his regional dialect.
The deep learning is very efficient technique used for this purpose as in this technique
the higher-level features are learned, it can identify direct relations, high measurement
information and can deal with unlabelled information.

ISSN: 1671-5497
Vol: 42 Issue: 11-2023
DOI: 10.5281/zenodo.10203386
Nov 2023 | 471
References
1) Chittaragi, N. B., Limaye, A., Chandana, N. T., Annappa, B., and Koolagudi, S. G.,”Automatic text-
independent kannada dialect identification system”, In Information Systems Design and Intelligent
Applications (pp. 79-87). Springer, Singapore, 2019.
2) Dharmale, G., Thakare, V. M., &Patil, D. D., “Implementation of Efficient Speech Recognition System
on Mobile Device for Hindi and English”, International Journal of Advanced Computer Science and
Applications (IJACSA), Vol. 10, No. 2, 2019.
3) Graves, A., Mohamed, A. R. and Hinton, G., “Speech recognition with deep recurrent neural networks,”
In 2013 IEEE international conference on acoustics, speech and signal processing (pp. 6645-6649).
IEEE, (2013).
4) Abdel-Hamid, O., Mohamed, A. R., Jiang, H., Deng, L., Penn, G. And Yu, D. (2014). Convolutional
neural networks for speech recognition. IEEE/ACM Transactions on audio, speech, and language
processing, vol22, no. 10, pp. 1533-1545, 2014.
5) Pulz, G., Blanchard, H. E., Lewis, S. H., Zhang, L, "System and method for recognizing speech with
dialect grammars." U.S. Patent 9,361,880, issued June 7, 2016.
6) Biadsy, F.,”Automatic dialect and accent recognition and its application to speech
recognition”, (Doctoral dissertation, Columbia University), 2011.
7) Biadsy, F., Hirschberg, J. And Habash, N. "Spoken Arabic dialect identification using
phonotacticmodeling." In Proceedings of the eacl 2009 workshop on computational approaches to
semitic languages, pp. 53-61. Association for Computational Linguistics, 2009.
8) Hermansky, H., Ellis, D. P and Sharma, S. Tandem connectionist feature extraction for conventional
HMM systems." In 2000 IEEE International Conference on Acoustics, Speech, and Signal Processing.
Proceedings (Cat. No. 00CH37100), vol. 3, pp. 1635-1638. IEEE, 2000.
9) Lamel, L., Gauvain, J. L and Adda, G. (2002), “Unsupervised acoustic model training." In 2002 IEEE
International Conference on Acoustics, Speech, and Signal Processing, vol. 1, pp. I-877.IEEE, 2002.
10) Singer, E., Torres-Carrasquillo, P. A., Gleason, T. P., Campbell, W. M. and Reynolds, D. A. "Acoustic,
phonetic, and discriminative approaches to automatic language identification." In Eighth European
conference on speech communication and technology. 2003.
11) Truong, K "Automatic pronunciation error detection in Dutch as a second language: an acoustic-
phonetic approach." (2004).
12) Desai, N., Dhameliya, K. and Desai, V. "Feature extraction and classification techniques for speech
recognition: A review." International Journal of Emerging Technology and Advanced Engineering 3,
no. 12 (2013): 367-371.
13) Chou, W. "Discriminant-function-based minimum recognition error rate pattern-recognition approach
to speech recognition." Proceedings of the IEEE 88, no. 8 (2000): 1201-1223.
14) Alhawiti, K. M. " World Academy of Science, Engineering and Technology, International Journal of
Computer, Electrical, Automation, Control and Information Engineering 9, no. 6 (2015): 1351-1354.
15) Cooper, H., Holt, B. and Bowden, R. "Sign language recognition." In Visual Analysis of Humans, pp.
539-562.Springer, London, 2011.
16) Ali, A., Dehak, N., Cardinal, P., Khurana, S., Yella, S. H., Glass, J and Renals, S. Automatic dialect
detection in arabic broadcast speech." arXiv preprint arXiv:1509.06928 (2015).
17) Cavell, S. Conditions handsome and unhandsome: the constitution of Emersonian perfectionism: The
Carus Lectures, 1988. University of Chicago Press, 2018.

ISSN: 1671-5497
Vol: 42 Issue: 11-2023
DOI: 10.5281/zenodo.10203386
Nov 2023 | 472
18) Narang, S and Gupta, M. D. "Speech feature extraction techniques: a review." International Journal of
Computer Science and Mobile Computing 4, no. 3 (2015): 107-114.
19) Chiţu, A. G., Rothkrantz, L. J., Wiggers, P and Wojdel, J. C. "Comparison between different feature
extraction techniques for audio-visual speech recognition." Journal on Multimodal User Interfaces 1,
no. 1 (2007): 7-20.
20) Liu, Z. T., Wu, M., Cao, W. H., Mao, J. W., Xu, J. P., & Tan, G. Z. "Speech emotion recognition based
on feature selection and extreme learning machine decision tree." Neurocomputing 273 (2018): 271-
280.
21) Yu, D and Deng, L. "Deep learning and its applications to signal and information processing
[exploratory dsp]." IEEE Signal Processing Magazine 28, no. 1 (2010): 145-154.
22) Ruder, S. "An overview of multi-task learning in deep neural networks." arXiv preprint
arXiv:1706.05098 (2017).
23) Sainath, T. N., Weiss, R. J., Wilson, K. W., Li, B., Narayanan, A., Variani, E., and Misra, A.
"Multichannel signal processing with deep neural networks for automatic speech
recognition." IEEE/ACM Transactions on Audio, Speech, and Language Processing 25, no. 5 (2017):
965-979.

DIALECTAL VARIABILITY IN SPOKEN LANGUAGE: A COMPREHENSIVE SURVEY OF MODERN TECHNIQUES FOR LANGUAGE DIALECT IDENTIFICATION IN SPEECH SIGNALS

More Related Content

Similar to DIALECTAL VARIABILITY IN SPOKEN LANGUAGE: A COMPREHENSIVE SURVEY OF MODERN TECHNIQUES FOR LANGUAGE DIALECT IDENTIFICATION IN SPEECH SIGNALS (20)

More from indexPub (20)

Recently uploaded (20)

DIALECTAL VARIABILITY IN SPOKEN LANGUAGE: A COMPREHENSIVE SURVEY OF MODERN TECHNIQUES FOR LANGUAGE DIALECT IDENTIFICATION IN SPEECH SIGNALS