Design and Implementation of a Phonological Analyzer for the Irula

Language: A Computational Approach to Endangered Language

Preservation

Vaishakh S. Kamath

, Safa Salim

, Jyothi Ratnam

, Arun Kumar K. P

and G. Veena

Department of Computer Science and Applications, Amrita School of Computing, Amritapuri, India

Amrita-CREATE, Amrita Vishwa Vidyapeetham, Amritapuri, India

Keywords:

Phonological Analyzer, Irula Language, Low-Resource Languages, Computational Linguistics, MFCC,

Machine Learning.

Abstract:

Irula is a South Dravidian language spoken by a small tribal community in India and is critically endangered

due to a decreasing number of speakers and the assimilation of its speakers into dominant regional languages

such as Tamil and Malayalam. This research focuses on creating a phonological analyzer that utilizes Mel-

Frequency Cepstral Coefﬁcients (MFCC) for feature extraction, aiming to tackle the computational limitations

faced by the Irula language. The study evaluates the performance of two machine learning models, Bi-LSTM

and SVM, in classifying audio data from Irula, Tamil, and Malayalam. Additionally, it explores the phonologi-

cal systems of these closely related languages, emphasizing articulatory features, variations in consonants, and

retroﬂex articulations. The ﬁndings indicate that the SVM model surpasses the Bi-LSTM model, achieving

a classiﬁcation accuracy of 90%. This project contributes to the preservation of endangered languages and

enhances the ﬁeld of low-resource language processing by creating a computational framework for analyzing

the phonological features of Irula. The results also open avenues for additional research in computational

linguistics and phonetics, particularly for languages that are underrepresented.

1 INTRODUCTION

The legacy and culture of a region is carried in its

languages which provides a sense of social identity.

However, many of the original Dravidian languages

that are native to India are in danger of becoming

extinct because of a lack of preservation and a drop

in speakers.The Irula community is an indigenous

tribal group belonging to the South Dravidian lan-

guage family, primarily residing in four southern In-

dian states: Andhra Pradesh, Karnataka, Kerala, and

Tamil Nadu. Their estimated population is around

200,000 individuals(The Criterion: An International

Journal in English, 2024). According to the 2011 cen-

sus, Irula, a South Dravidian tribal language, is an

endangered language, with 12,000 speakers in Kerala

currently (Kerala Institute for Research Training and

Development Studies of Scheduled Castes and Sched-

uled Tribes (KIRTADS), 2017). The speakers from a

small hamlet in Kerala’s Attapadi region are the sub-

ject of our attention. In the domains of computational

linguistics and natural language processing, Irula is

considered a zero-resource language. This study aims

to solve this problem by creating the ﬁrst phonolog-

ical analyzer speciﬁcally designed for the Irula lan-

guage. The goal is to determine whether the provided

audio data is from the sister Dravidian languages of

Tamil, Malayalam, or Irula.

Code mixing or code-switching refers to the jux-

taposition of linguistic units from two or more lan-

guages in a single conversation or sometimes even

a single utterance. It is the transfer of linguistic el-

ements or words from one language to another or

mixed together. Native speakers, mostly tribal peo-

ple, are attempting to switch to the main-stream lan-

guages of their area, resulting in code mixing. The

main external factors causing language endangerment

among the Irula people are cultural, educational, and

economic subjugation. Internal factors, such as the

community’s rejection of its own language and the

general decline of group identity, also contribute to

this issue. The Irula tribal communities link their tra-

ditional language and culture to their precarious so-

cial and economic status. They now feel that their

languages are useless and should not be preserved. In

S. Kamath, V., Salim, S., Ratnam, J., Kumar K. P, A. and Veena, G.

Design and Implementation of a Phonological Analyzer for the Irula Language: A Computational Approach to Endangered Language Preservation.

DOI: 10.5220/0013606300004664

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 2, pages 871-877

ISBN: 978-989-758-763-4

871

their pursuit to combat discrimination, achieve ﬁnan-

cial stability, and enhance social mobility for them-

selves and their children, they have relinquished their

languages and traditions. However, they are unaware

that they risk losing their identity if they are unable

to communicate in their native tongue. In such a sit-

uation, it became crucial to preserve and strengthen

the root language, or Irula, from both a cultural and a

computational standpoint. Irula presents a signiﬁcant

obstacle due to the absence of any digital corpus or

documentation. The objective of this project is to cre-

ate computational tools for Irula that will signiﬁcantly

inﬂuence the ﬁeld of low-resource language process-

ing and help preserve the language.

Proto-Dravidian refers to the linguistic recon-

struction of the common ancestor of the Dravidian

languages indigenous to the Indian subcontinent. This

language set has descendants in Irula, Tamil, and

Malayalam. Analyzing these three languages reveals

clear commonalities in syntax, semantics, and lexi-

con. The phonological system is where the degree of

divergence appears. These parallels and discrepancies

are examined in this study. It also emphasizes how the

phoneme inventory and sound patterns of these lan-

guages have changed over time.

Malayalam has a more signiﬁcant phonetic dis-

tinction for every consonant. Irula, which is more

like Tamil, has distinct phonological characteristics,

as evidenced by the change in where some consonants

are articulated. These changes emphasize how each

language’s phonetics—particularly its fricatives and

retroﬂexes—make these sister languages distinct. The

aim of this study is to assess and analyze the phono-

logical distinctions that exist. Irula shares phonetic

similarities with Tamil and Malayalam, but it also has

distinctive phonetic characteristics, such as consonan-

tal shift and phoneme simpliﬁcation. For instance, the

word ”singam” in Tamil means ”lion,” whereas the

word ”simham” in Malayalam indicates both phono-

logical and lexical diversity. Irula is closely con-

nected to languages like Malayalam, as seen by the

term ”shivaya,” which is pronounced similarly. This

word is pronounced ”sivaya” in Tamil, and ”chivaya,”

”shivaya,” or ”sivaya” in Irula. This research fo-

cuses on phonetic differences, as well as syntactic and

grammatical similarities.

This effort also aims to analyze the similarities

and differences between these languages. Despite

having a similar lexicon and grammatical organiza-

tion, their phonetic components differ greatly. Frica-

tives and retroﬂexes aid in the development of a thor-

ough phonological model that precisely differentiates

between these languages. Vowel articulation and pho-

netic markers offer a strong foundation for examining

these languages’ development. The necessity for in-

struments that can precisely record the phonetic vari-

ety among the proto-Dravidian languages and aid in

their preservation is discussed in this work.

Section II provides an overview of related re-

search in phonological analysis, emphasizing stud-

ies that focus on language preservation and linguis-

tic diversity, especially concerning low-resource lan-

guages. This section reviews prior work on acous-

tic feature extraction, phoneme segmentation, and

the creation of language models speciﬁcally designed

for endangered languages. Section III describes the

methodology of this study, including the dataset uti-

lized for phonological analysis, feature extraction

methods like MFCC, and the models used for pho-

netic comparison. Section IV presents the results

from the acoustic analysis, highlighting both the sim-

ilarities and differences in phonetic characteristics

among Malayalam, Tamil, and Irula while assessing

the effectiveness of various feature extraction tech-

niques. Section V assesses the phonological mod-

els based on their ability to identify language-speciﬁc

features, particularly regarding their role in preserv-

ing Irula. Finally, Section VI concludes with a discus-

sion on the implications of these ﬁndings for linguis-

tic research and outlines potential directions for future

work, such as incorporating machine learning tech-

niques for automatic phoneme segmentation and de-

veloping speech-to-text systems for low-resource lan-

guages.

2 RELATED WORKS

The majority of the existing computational linguis-

tics literature on voice analysis concentrates on lan-

guages with abundant resources or extensive docu-

mentation. However, research on resource-centered

and endangered languages has brought attention to the

need for computational tools designed speciﬁcally for

these languages. Speech segmentation and pronunci-

ation modeling are the subjects of two studies that are

particularly relevant to this endeavor.

The ﬁrst study segmented voice samples using

a hybrid segmentation system that integrated signal

processing and machine learning methods(Prakash

et al., 2016). particularly Speech data is divided into

syllable-level segments using a hidden Markov model

(HMM), which is initialized using the global aver-

age variance. Along with examining the distribution

of acoustic qualities among the languages, the study

also examined the acoustic characteristics of sylla-

bles in six Indian languages. and determine the par-

allels and discrepancies The study’s ﬁndings demon-

INCOFT 2025 - International Conference on Futuristic Technology

872

strate how these models could enhance TTS systems

and speech technologies for low-resource and poly-

glot languages. However, the lack of variation in lan-

guage, dialect, and speakers makes research difﬁcult.

Having trouble understanding natural or loud speech

also adds to it.

For compatibility with Indian languages, a sec-

ond study created a pronunciation model that in-

cludes spelling and pronunciation techniques trans-

ferred from UTF-8 to ISCII(Singh, 2006). The model

employs Dynamic Time Warping (DTW) for string

alignment and incorporates telemetry to gauge pho-

netic similarity. Spell checking is one of the many

applications that the model supports. Normalization

of text and identiﬁcation of related terms. The re-

sults demonstrate how standardizing a phonemic or-

thographic framework for the Brahmi script has im-

proved natural language processing for Indian lan-

guages. However, issues like resource efﬁciency were

identiﬁed. limited pronunciation format presentation

and language generalization.

Hegarty et.al. (2019) phonetic comparison of-

fers crucial tools for comprehending the phonetic

diversity of languages. An online database and

tools for analyzing phonetic similarities and con-

trasts between languages are made available by this

project(Nallanthighal et al., 2021). It is a useful tool

in the ﬁeld of comparative linguistics. Dialectics and

sociolinguistics ”Phonetics Comparison” emphasizes

the difﬁculty of developing phonetics resources for

those with low proﬁciency despite its wide breadth.

This project’s technique, which includes the capac-

ity to search, ﬁlter, and examine phonetic data, es-

tablishes a standard for recording and examining the

phonological characteristics of underrepresented lan-

guages.

Tohru et al. represent the overview of evolution

of speech recognition, speech synthesis, and machine

translation.Key approaches include a minimum mean

square error (MMSE) estimator with a Gaussian mix-

ture model (GMM) and a particle ﬁlter to reduce noise

and interference(Shimizu et al., 2008). An MDLSSS

algorithm determines the number of parameters based

on training data size using the maximum description

length criterion. Despite well-trained models, fac-

tors like speaker variability and background noise can

still cause errors, which are mitigated by applying

generalized word posterior probability (GWPP) for

post-processing. The translation process uses TATR

and EM modules, with TATR leveraging translation

examples and employing probabilities and penalties

when necessary.The XIMERA speech synthesis en-

gine, trained on large corpora from Japanese, English,

and Chinese speakers, uses an HMM-based statisti-

cal prosody model to produce natural pitch patterns

without heavy signal processing. Evaluation across

the BTEC, MAD, and FED corpora ensures the sys-

tem’s robustness in various speaking styles. Results

indicate high accuracy (82-94 percentage for Chinese)

and BLEU scores (0.55-0.74) for Chinese-Japanese

and Chinese-English translations. The paper con-

cludes that integrating these advanced techniques sig-

niﬁcantly enhances the performance and consistency

of speech recognition, synthesis, and machine trans-

lation systems

Nallanthighal et al. investigates the connection

between speech’s respiratory effort and phoneme for-

mation. The results of the study on the respiratory pat-

terns linked to particular phonetic classes provide in-

sights into the physical components of speech produc-

tion, despite the controlled setting and small sample

size(Heggarty et al., 2019). These observations could

help with the phonological analysis of Irula by exam-

ining the differences in respiratory effort between the

language’s phonetic components. However, the ac-

curacy of lung volume changes—which is crucial for

practical applications—is limited by the absence of

calibration in respiratory data measurement.

While previous studies have examined various

methods for phonological analysis and language

preservation, there has been limited research compar-

ing machine learning models for phonetic analysis in

low-resource languages such as Irula. Furthermore,

the automation of phonological feature extraction us-

ing models like BiLSTM and SVM has not been ex-

tensively studied, particularly in the context of Indian

languages. This research seeks to ﬁll this gap by as-

sessing BiLSTM and SVM models for phoneme seg-

mentation and comparing the acoustic properties of

Malayalam, Tamil, and Irula. The aim is to enhance

the accuracy of phonological analysis tools for endan-

gered languages and evaluate the feasibility of these

models for automating phonetic feature extraction in

less-explored languages.

3 METHODOLOGY

MFCCs are used (Han et al., 2006)(Paulraj et al.,

2009)in this study’s methodology to extract phonetic

features from audio recordings of considered lan-

guages Malayalam, Tamil and Irula.Both deep learn-

ing model and a machine learning model were im-

plemented to classify the phoneme analysis for Irula,

Malayalam, and Tamil. The procedure can be divided

into several key steps, including dataset preparation

,feature extraction, train the model, and evaluation.

Design and Implementation of a Phonological Analyzer for the Irula Language: A Computational Approach to Endangered Language

Preservation

873

3.1 Dataset Preparation

We gathered voice samples from one male and one

female speaker of each language for our study, result-

ing in a total of six participants. For dataset devel-

opment, we collected one hundred comparable words

from Irula, Malayalam, and Tamil. These words share

similarities in phonology, morphology, and seman-

tics. The recordings are made by native speakers of

the respective languages and saved in WAV format.

3.2 Feature Extraction

The participants’ voice signals were(Paulraj et al.,

2009) captured. at a sample rate of 16 kHz. The

audio samples were processed using MFCC, a pop-

ular method for extracting phonetic characteristics in

audio sources . The audio samples were examined

using MFCC(Han et al., 2006), a commonly em-

ployed method for extracting phonetic features from

(Paulraj et al., 2009) audio sources. The primary ob-

jective of MFCC is to mimic human auditory per-

ception. For each audio ﬁle, thirteen MFCC coefﬁ-

cients—an established standard in speech processing

research—were extracted. Consequently, each audio

frame generated a total of thirteen MFCC features, ef-

fectively capturing essential phonetic traits.

Figure 1: An illustration of the MFCC block

The components of MFCC(Roul and Rath, 2023)

feature extraction algorithm:

3.2.1 Frameblocking

The continuous voice stream is segmented into frames

of N samples, with neighboring frames overlap-

ping by M samples (where M is less than N). The

ﬁrst frame contains the original N samples. (mfc,

2023)(Paulraj et al., 2009). This pattern continues as

the second frame overlaps the ﬁrst by M samples and

begins M samples after it. This process continues un-

til each spoken phrase is captured in either a single

frame or multiple frames(mfc, 2023)(Paulraj et al.,

2009) . Typically, N and M are set to 256 and 100,

respectively(mfc, 2023)(Paulraj et al., 2009) .

3.2.2 Windowing

Windowing reduces spectral distortion and signal dis-

continuities at the start and end of each frame by

smoothly tapering the signal to zero(Paulraj et al.,

2009). When N denotes the number of samples in

each frame(Paulraj et al., 2009), the resulting signal

is obtained through this windowing technique(mfc,

2023)(Paulraj et al., 2009). The Hamming window,is

usually utilized.

3.2.3 Fast Fourier Transform

After applying a Hamming window to reduce spectral

leakage, the Fast Fourier Transform (FFT) is used to

convert the signal from the time domain to the fre-

quency domain. The FFT is an efﬁcient algorithm

designed to compute the Discrete Fourier Transform

(DFT) for a set of N samples, {x

}. In contrast

to the DFT, which has a computational complexity

of O(N

), the FFT optimizes this to O(N logN) by

taking advantage of the symmetry and periodicity of

complex exponentials.

3.2.4 Mel-Frequency Wrapping

The frequency characteristics of spoken sounds do not

conform to a conventional linear scale(Paulraj et al.,

2009). As a result, a specialized scale known as the

”Mel” scale is utilized to relate perceived pitch to

an actual frequency f measured in Hertz (Hz)(Paulraj

et al., 2009). On this scale, frequencies below 1000

Hz are spaced linearly, while those above 1000 Hz

shift to a logarithmic spacing. Typically, the number

of coefﬁcients obtained from the Mel-frequency spec-

trum, represented as K, is commonly set at 20(Paulraj

et al., 2009).

3.2.5 Cepstrum

The logarithmic time is again transformed from the

Mel spectrum, resulting in the MFCCs(Han et al.,

2006). For the speciﬁed frame analysis, the cepstral

representation of the speech spectrum provides an ef-

fective depiction of the signal’s local spectral char-

acteristics(Paulraj et al., 2009). The Discrete Cosine

Transform (DCT) can be applied to convert the real-

number Mel spectrum coefﬁcients into the(Paulraj

et al., 2009) temporal domain.

Using the previously described method, a set

of MFCCs is computed for each 30 ms speech

frame(Paulraj et al., 2009), incorporating overlap.

The DCT coefﬁcients for each speech input are ex-

tracted and used as the feature set for the neural net-

work(Paulraj et al., 2009) model.

INCOFT 2025 - International Conference on Futuristic Technology

874

3.3 Model Architecture and Training

3.3.1 Bi-LSTM

A speciﬁc type of recurrent neural network (RNN),

known as bi-LSTM (bidirectional long short-term

memory)(Subramanian et al., 2023) , is capable of

processing sequential data in both frontward and

backward directions(Nama, 2018). By combining

bidirectional processing with the strengths of LSTM,

this model can effectively integrate both the previous

and next context of the input sequence(Madamanchi

et al., 2024). This bidirectional approach signiﬁcantly

reduces the risk of losing important information that

may be inﬂuenced by factors present in the opposite

direction.

To evaluate the performance of a simpliﬁed BiL-

STM model, the Keras framework was employed for

building and training the architecture. The input data

was enhanced with Gaussian noise by adding random

values with a mean of 0 and a standard deviation of

0.1 to the training dataset. This augmentation aimed

to introduce variability and improve the model’s ro-

bustness.

The model architecture was constructed using a

sequential layout. The initial layer featured a Bi-

LSTM with 128 units, conﬁgured to avoid returning

intermediate sequences, thus ensuring a compact rep-

resentation of the input data. To avoid overﬁtting, a

Dropout layer with a probability of 0.5 was added to

deactivate nodes during training.

Subsequently, a dense layer with 32 units and a

hyperbolic tangent activation function was incorpo-

rated. This choice differed from the commonly used

ReLU activation function, allowing for an assessment

of how an alternative activation mechanism inﬂuences

the model’s learning process. The output layer con-

sisted of a dense layer with a softmax activation func-

tion, generating probabilities for each class corre-

sponding to the number of target categories in the one-

hot encoded labels. The model was compiled using

the Adam optimizer, with categorical cross-entropy

as the loss function and accuracy as the evaluation

metric. To minimize computational demands, training

was performed on a smaller subset of data, consisting

of only 100 samples. The model underwent training

for 10 epochs with a batch size of 32. These adjust-

ments, including the reduced dataset size and training

duration, were implemented to investigate the model’s

performance under resource-constrained conditions.

3.3.2 Support Vector Machine

Support Vector Machine (SVM) is a powerful super-

vised learning method that is mainly used for classi-

ﬁcation purposes.(Nair et al., 2019). It works by ﬁg-

uring out the best hyperplane—or decision boundary

in higher dimensions—to divide various data point

classes(Nair et al., 2019)(Ratnam et al., 2021). The

primary goal of Support Vector Machines (SVM) is

to maximize the distance between the nearest data

points from different classes, known as support vec-

tors. (Nair et al., 2019). This strategy improves

the model’s resilience and its ability to generalize to

new data. (Nair et al., 2019)(Ratnam et al., 2021).

MFCCs are frequently used to capture audio char-

acteristics, leading to a feature space that is typi-

cally high-dimensional, with 13 coefﬁcients for each

time frame. Given that SVMs perform well in high-

dimensional settings, they are particularly suitable for

tasks such as sound or vowel classiﬁcation. We im-

plemented SVM using the Scikit-learn library , apply-

ing a linear kernel function that is effective for high-

dimensional feature spaces like the MFCCs utilized in

this study. SVM is recognized for its efﬁcacy in sound

and vowel categorization tasks(Nair et al., 2019)(Rat-

nam et al., 2021).

4 EVALUATION AND RESULT

4.1 Comparison between Irula,

Malayalam and Tamil

Table 1: Comparison between Irula and its Sister Lan-

guages.

Weight 1 Weight 2 Weight 3 Language Phoneme

-576.296 42.311 -4.741 Malayalam [a]

-442.539 51.198 7.521 Irula [a]

-463.308 73.609 -2.683 Tamil [a]

-563.650 24.217 17.235 Malayalam [i]

-384.451 49.462 -1.058 Irula [i]

-429.236 60.565 21.706 Tamil [i]

-609.626 68.138 30.401 Malayalam [u]

-431.740 46.375 2.256 Irula [u]

-458.865 85.762 20.168 Tamil [u]

-573.000 30.000 7.720 Malayalam [e]

-367.637 53.044 -8.653 Irula [e]

-4.732 57.919 13.994 Tamil [e]

-538.000 62.500 25.600 Malayalam [o]

-386.394 63.102 3.869 Irula [o]

-418.050 89.870 18.211 Tamil [o]

4.2 Evaluation

In this task, we assess both models using evaluation

metrics such as precision, recall, and F1-score(Tang

et al., 2020). Speciﬁcally, in this multi-class classi-

ﬁcation task, these metrics help us in evaluating the

model’s performance(Tang et al., 2020). Precision is

a statistical metric employed to assess the effective-

Design and Implementation of a Phonological Analyzer for the Irula Language: A Computational Approach to Endangered Language

Preservation

875

ness of a classiﬁcation model, especially in identify-

ing true positive predictions(Tang et al., 2020). Recall

is a performance metric that measures a model’s ca-

pability to accurately identify all pertinent instances

within a dataset(Tang et al., 2020). The F1 Score is

a statistical metric that integrates two essential evalu-

ation measures: precision and recall. It reﬂects their

harmonic mean, offering a single value that balances

both the accuracy (precision) and comprehensiveness

(recall) of a model’s predictions(Tang et al., 2020).

4.3 Result

The performance of the phonological analyzer was

evaluated using two machine learning models: Bi-

LSTM and SVM. The models were assessed based

on their classiﬁcation reports and confusion matrices,

which provide insights into the predictive accuracy

and error patterns.

The classiﬁcation performance of the models is

summarized in Tables 2 and 3, which provide de-

tailed metrics, including precision, recall, F1-score,

and support for each class.

Table 2: Classiﬁcation report for Bi-LSTM.

Class Precision Recall F1-Score Support

Irula 0.85 0.92 0.89 38

Malayalam 0.86 0.86 0.86 37

Tamil 0.90 0.69 0.78 13

Table 3: Classiﬁcation report for SVM.

Class Precision Recall F1-Score Support

Irula 0.95 0.97 0.96 38

Malayalam 0.89 0.86 0.88 37

Tamil 0.77 0.77 0.77 13

To further evaluate the performance of both mod-

els, a Confusion Matrix was employed. These matri-

ces are presented in Figures 2 and 3 for the Bi-LSTM

and SVM models, respectively.

Figure 2: Confusion Matrix for Bi-LSTM

Figure 3: Confusion Matrix for SVM

5 CONCLUSION

This study presents an innovative approach to analyz-

ing and classifying speech data from ﬁve individuals

into three distinct classes based on articulatory fea-

tures such as tongue position, tension, and lip place-

ment (front, middle, and rear). MFCCs for feature

extraction, the research assessed the performance of

Bi-LSTM and SVM. The empirical analysis reveals

that the SVM model outperformed the Bi-LSTM in

classiﬁcation accuracy, highlighting its potential as a

robust method for phonological analysis.

This research not only showcases computational

advancements but also has signiﬁcant implications

for the preservation of Irula, an endangered language

with a rich linguistic heritage. By combining compu-

tational methodologies with linguistic analysis, it es-

tablishes a foundational framework for studying Irula

and offers a scalable approach for other low-resource

languages. By emphasizing the phonological differ-

ences and common features among Irula, Tamil, and

Malayalam, this work contributes to the advancement

of computational phonology and linguistics.

Furthermore, it highlights the urgent need to

develop tools for preserving and analyzing endan-

gered languages in the digital age. The framework

presented here is anticipated to support future re-

search in phonological modeling, language preserva-

tion, and computational applications for low-resource

languages, ensuring that their cultural and academic

legacies endure.

REFERENCES

(2023). Speaker recognition using mfcc. Slideshare. Avail-

able at: https://www.slideshare.net/slideshow/speaker

-recognition-using-mffc/29354934\#9.

Han, W., Chan, C.-F., Choy, C.-S., and Pun, K.-P. (2006).

An efﬁcient mfcc extraction method in speech recog-

INCOFT 2025 - International Conference on Futuristic Technology

876

nition. In 2006 IEEE International Symposium on Cir-

cuits and Systems (ISCAS), pages 4 pp.–.

Heggarty, P., Shimelman, A., Abete, G., Anderson, C.,

Sadowsky, S., Paschen, L., Maguire, W., Jocz, L.,

Aninao, M., W

agerle, L., Appelganz, D., Silva, A.,

Lawyer, L., Camara Cabral, A. S., Walworth, M.,

Michalsky, J., Koile, E., Runge, J., and Bibiko, H.-J.

(2019). Sound comparisons: A new resource for ex-

ploring phonetic diversity across language families of

the world. INTERNATIONAL CONGRESS OF PHO-

NETIC SCIENCES 2019.

Kerala Institute for Research Training and Development

Studies of Scheduled Castes and Scheduled Tribes

(KIRTADS) (2017). Irular — KIRTADS. Accessed:

2024-12-18.

Madamanchi, S., Nedungadi, P., Sai, M., Reddy, R., and

Ratnam, J. (2024). Automated speech correction as-

sistive technology for malayalam articulation errors.

Nair, A. J., Rasheed, R., Maheeshma, K., Aiswarya, L., and

Kavitha, K. R. (2019). An ensemble-based feature

selection and classiﬁcation of gene expression using

support vector machine, k-nearest neighbor, decision

tree. In 2019 International Conference on Communi-

cation and Electronics Systems (ICCES), pages 1618–

1623.

Nallanthighal, V. S., H

arm

a, A., Strik, H., and Doss, M. M.

(2021). Phoneme based respiratory analysis of read

speech. In 2021 29th European Signal Processing

Conference (EUSIPCO), pages 191–195.

Nama, A. (2018). Understanding bidirectional lstm for se-

quential data processing. https://medium.com/@anis

hnama20/understanding-bidirectional-lstm-for-seque

ntial-data-processing. Accessed: 2024-11-25.

Paulraj, M. P., Yaacob, S. B., Nazri, A., and Kumar, S.

(2009). Classiﬁcation of vowel sounds using mfcc

and feed forward neural network. In 2009 5th Inter-

national Colloquium on Signal Processing & Its Ap-

plications, pages 59–62.

Prakash, A., Prakash, J. J., and Murthy, H. A. (2016).

Acoustic analysis of syllables across indian lan-

guages. In Interspeech 2016, pages 327–331.

Ratnam, J., Kp, S., T.K, B., Priya, M., and B., P. (2021).

Hybrid Machine Translation System for the Transla-

tion of Simple English Prepositions and Periphrastic

Causative Constructions from English to Hindi, pages

247–263.

Roul, A. K. and Rath, B. N. (2023). Text independent

speaker recognition using fuzzy based neural net-

work. In 2023 1st International Conference on Cir-

cuits, Power and Intelligent Systems (CCPIS), pages

1–5.

Shimizu, T., Ashikari, Y., Sumita, E., Zhang, J., and Naka-

mura, S. (2008). Nict/atr chinese-japanese-english

speech-to-speech translation system. Tsinghua Sci-

ence and Technology, 13(4):540–544.

Singh, A. K. (2006). A computational phonetic model for

indian language scripts. In Semanticscholar.

Subramanian, M., S, L. S., and V R, R. (2023). Deep learn-

ing approaches for melody generation: An evaluation

using lstm, bilstm and gru models. In 2023 14th Inter-

national Conference on Computing Communication

and Networking Technologies (ICCCNT), pages 1–6.

Tang, T., Tang, X., and Yuan, T. (2020). Fine-tuning bert

for multi-label sentiment analysis in unbalanced code-

switching text. IEEE Access, 8:193248–193256.

The Criterion: An International Journal in English (2024).

The Irula Language and Literature. Accessed: 2024-

12-18.

Design and Implementation of a Phonological Analyzer for the Irula Language: A Computational Approach to Endangered Language

Preservation

877