Call Center Customer Sentiment Analysis Using

ML and NLP
2023 14th International Conference on Intelligent Systems: Theories and Applications (SITA) | 979-8-3503-0821-1/23/$31.00 ©2023 IEEE | DOI: 10.1109/SITA60746.2023.10373715

1st Imad AATTOURI 2nd Hicham MOUNCIF 3rd Mohamed RIDA

Ain Chock Faculty of Science Polydisciplinary Faculty of Beni Mellal Ain Chock Faculty of Science
Hassan 2 University Sultan Moulay Slimane University Hassan 2 University
Casablanca, Morocco Beni Mellal, Morocco Casablanca, Morocco
aattouriimad@gmail.com h.mouncif@usms.ma mhd.rida@gmail.com

Abstract—In the contemporary digital era, call centers have nuances through the customer’s tone, choice of words, and
significantly incorporated automation through callbots, but they voice inflections. But a standard callbot can’t do this... at least,
often lack an essential aspect of customer service: empathy. This not without the right technological aid.
paper explores the integration of sentiment analysis into call
center operations to introduce an emotional dimension to callbot This is where sentiment analysis, combining ML techniques
interactions. Utilizing natural language processing and machine and NLP , comes into play. By equipping callbots with this
learning, the paper examines both text-based and signal-based capability, it becomes possible to detect a customer’s sentiment
sentiment analysis approaches. The proposed sentiment analysis and act accordingly, whether it’s to adjust the bot’s response
architecture encompasses data input and output interfaces, pre-
or to transfer the call to a human agent.
processing, sentiment analysis, and a decision manager module
that can, for example, escalate calls to a human agent based on In this article, we will explore how such integration could
sentiment analysis results. The preprocessing steps, both for text not only humanize automated interactions but also greatly en-
and signal analysis, are outlined in detail, along with a review of hance the customer experience in call centers. We will outline
relevant sentiment analysis algorithms. the techniques and methodologies we’ve adopted, focusing
Through comprehensive experiments, the study demonstrates
that the integration of sentiment analysis yields promising out-
on two main approaches: text-based analysis and signal-based
comes. In text-based analysis, SVM and LSTM models consis- vocal analysis.
tently performed well, achieving accuracy scores of 74% and
72%, respectively. For voice-based analysis, the MLP model II. L ITERATURE R EVIEW
exhibited the highest accuracy of 0.72 when using mel spec-
trogram features, while the RF model outperformed others This article is a continuation of our work on the develop-
with an accuracy of 0.78 using MFCC features. These results ment of an intelligent callbot [1], [2], [3].
showcase the potential of sentiment analysis in humanizing
callbot interactions, thereby enhancing customer satisfaction and Sentiment analysis is a firmly established subfield of NLP
service efficiency. that endeavors to ascertain the emotion or sentiment conveyed
Index Terms—Sentiment Analysis, Call Centers, Callbots, NLP within a textual entity [4]. It has found manifold applications,
, Machine Learning, Text-based Analysis, Signal-based Analysis, from social media analysis [5] to finance [6], and, of course,
Emotion Detection, SVM, Naive Bayes, Random Forest, LSTM, call centers.
MLP, Mel Frequency Cepstral Coefficients (MFCC), Customer
Satisfaction.
A. Text-based Sentiment Analysis in Call Centers
I. I NTRODUCTION Text-based sentiment analysis techniques have been exten-
In the era of digitalization and artificial intelligence, cus- sively studied and applied to various domains. Within the
tomer interaction has undergone a significant transformation. call center context, call transcriptions can be analyzed to
Nowadays, it is common to reach out to customer service gauge customer sentiment [7]. Common methods include those
and be greeted not by a human voice, but by a callbot. based on bag-of-words models, such as SVM [8] and logistic
These automated systems, powered by complex algorithms regression [9], as well as more advanced techniques employing
and sometimes even neural networks, aim to streamline and neural networks like LSTM [10] and BERT [11].
expedite the handling of customer requests, while optimizing
human resources. However, despite their technical efficiency, B. Signal-based Sentiment Analysis
these callbots can lack one essential thing: empathy. Beyond lexical content, voice tone, pacing, pauses, and
Feeling and emotion play a pivotal role in human com- other prosodic features can carry significant insights into a
munication. A displeased, frustrated, or confused customer speaker’s sentiment or emotional state [12]. Techniques like
requires a different approach than a satisfied or neutral one. Convolutional Neural Networks (CNN) [13] and Gaussian
In a traditional call center, a human agent can discern these Mixture Models (GMM) [14] have been employed to classify
979-8-3503-0821-1/23/$31.00 ©2023 IEEE emotions from voice signals.

Authorized licensed use limited to: Indian Institute of Technology Palakkad. Downloaded on July 08,2025 at 14:07:23 UTC from IEEE Xplore. Restrictions apply.
C. Challenges in Emotion Detection for Call Centers steps involved in assessing and responding to the feelings of
Accurate emotion detection within call centers presents customers as they interact with a callbot.
unique challenges. In addition to background noise that may
hamper signal quality [15], individual variability in emotion
expression [16], and the need for real-time data processing
[17] complicate the endeavor.

III. M ATERIALS AND M ETHODS

A. Materials
In this article, we explore the integration of sentiment
analysis into call center operations to introduce an emotional
dimension into interactions with callbots. To capture this emo-
tional dimension from two complementary perspectives, we
utilize two distinct datasets: one based on voice recordings and
another based on text transcriptions. This methodology allows
for a more comprehensive comprehension of the sentiments
conveyed by customers during their interactions with callbots.
The use of two different datasets is justified for several
reasons. First, vocal and text-based interactions differ in how
emotions are conveyed. Emotions in the voice can be transmit-
ted through elements such as tone, intensity, and pace, whereas
in text, emotions rely on word choices, expressions, and phras-
ing. By combining voice recordings with text transcriptions,
Fig. 1. Sentiment Analysis Architecture for the Call Center.
we encompass a broader range of emotional cues.
Furthermore, the combination of the two datasets enhances
a) Input Interface: The first step in this architecture
the robustness and generalizability of our sentiment analysis
captures the client’s interaction. Whether it’s a vocal or textual
approach. ML models trained on multiple datasets better
communication, this interface is pivotal as it stands as the entry
capture variations and nuances in the sentiments expressed
point for all the data the system needs to analyze.
by customers.
1) Text-based Sentiment DataSet: For the text-based senti- b) Preprocessing: After capturing, the data undergoes a
ment analysis aspect of our research, the selection of an appro- preprocessing phase. In this step, any superfluous or irrelevant
priate dataset played a pivotal role. After careful consideration information is filtered out, and the data is formatted for
of various available options, we opted for the ”French Twitter optimized analysis. This phase ensures only pertinent data is
Sentiment Analysis” dataset [18], which aligns closely with the forwarded to the analysis stage.
objectives of our study. With 1.5 million tweets, the dataset c) Analysis: This is the heart of the architecture. Once
offers a substantial amount of textual content for analysis, preprocessed, the data is analyzed to discern the client’s
enabling us to obtain robust and comprehensive insights into sentiment. Employing advanced ML and NLP techniques, the
sentiment patterns. system can evaluate whether the client is satisfied, frustrated,
2) voice-based Sentiment DataSet: For our voice-based neutral, or exhibits any other relevant emotion.
sentiment analysis, we harnessed the ”Canadian French Emo- d) Decision Manager: Based on the results from the
tional (CaFE) Speech Dataset” [19], a notable resource in the analysis phase, the decision manager determines the next
field. This dataset features a diverse array of emotional speech course of action. If, for instance, the client seems discontented
recordings in the French language, comprising six distinct or frustrated, the manager might decide to transfer the call to
sentences uttered by twelve actors representing both genders. a human agent for a more personalized handling.
Covering six fundamental emotions and a neutral state, the e) Output Interface: Lastly, this final step implements
CaFE dataset consists of three parts. Initially, the dataset the determined decision. This could manifest as an automated
encompassed 864 audio files capturing various emotional response to the client, a closing message, or a transfer to a
expressions. human agent, depending on what the decision manager has
determined.
B. Methods 2) Preprocessing for Text-based Analysis: To ensure accu-
1) Sentiment Analysis Architecture Proposed for the Call rate sentiment analysis on text data, a series of preprocessing
Center: In order to facilitate a comprehensive analysis of steps (Figure 2) were applied to the extracted content from the
customer interactions within the call center, we have designed dataset.
a sentiment analysis architecture as depicted in Figure 1. a) Text Cleaning: The text data underwent thorough
This architecture provides an overview of the consecutive cleaning, involving the removal of punctuation, special char-

Authorized licensed use limited to: Indian Institute of Technology Palakkad. Downloaded on July 08,2025 at 14:07:23 UTC from IEEE Xplore. Restrictions apply.
 Fig. 3. Preprocessing for Voice-based Analysis

b) Audio Segmentation: Continuous audio streams from

Fig. 2. Preprocessing for Text-based Analysis the CaFE dataset were segmented into smaller, meaningful
units, such as sentences or utterances. This segmentation
process not only organized the data but also allowed for
acters, and numerical digits. This step aimed to reduce noise focused analysis of distinct emotional expressions.
and standardize the data. c) Audio Feature Extraction: Relevant acoustic features
b) Tokenization: Tokenization segmented the text into were extracted from each segmented audio unit. These features
individual words or tokens. This process provided the basis for included pitch, intensity, speech rate, and pauses, among
subsequent analysis, breaking down the text into manageable others. The extraction process transformed raw audio data into
units. quantifiable attributes for subsequent analysis.
c) Stop-word Removal: Common stop words, such as d) Normalization: To ensure consistency and compara-
articles, conjunctions, and prepositions, were eliminated from bility across various recordings, the extracted audio features
the text to focus on significant content and enhance the quality underwent normalization. This step standardized the range and
of analysis. scale of features, facilitating accurate and meaningful analysis.
d) Stemming/Lemmatization: Stemming or lemmatiza- e) Vector Representation: Normalized audio features
tion was applied to reduce words to their base forms. This step were then converted into numerical vectors, rendering them
aimed to normalize variations and facilitate more meaningful suitable for input to ML algorithms. This vector representation
analysis. formed the basis for our sentiment analysis techniques.
e) Named Entity Recognition: Named Entity Recogni- The preprocessing of voice data from the Canadian French
tion (NER) identified and categorized proper nouns, such as Emotional (CaFE) Speech Dataset lays a solid groundwork
names of people, places, and organizations. NER enriched the for our subsequent sentiment analysis, enabling us to delve
understanding of context and entities within the text. into the intricate emotional nuances present within call center
f) Vector Representation: Text data was transformed into interactions.
numerical vectors through techniques like CountVectorizer and 4) Algorithms:
TF-IDF. This conversion enabled ML algorithms to process a) SVM (Support Vector Machine): is a classification
and analyze the text. algorithm used in sentiment analysis to separate positive
g) Feature Engineering: Additional features were en- and negative comments based on textual features. It seeks
gineered to enhance analysis, including n-grams to capture the optimal hyperplane that maximizes the margin between
contextual information and phrase structures. classes, effectively classifying new examples based on their
The comprehensive preprocessing process lays the founda- proximity to this hyperplane[20].
tion for accurate sentiment analysis, enabling the extraction of b) Naive Bayes: is an algorithm that employs the princi-
insightful emotional expressions from the call center context. ples of Bayes’ theorem in order to forecast the category of a
3) Preprocessing for Voice-Based Analysis: In order to given sample. This is achieved by calculating the conditional
facilitate a comprehensive sentiment analysis on voice data, probability of each class, taking into account the input features.
a series of preprocessing steps were applied to the ”Canadian Despite its simplistic assumption of independence among
French Emotional (CaFE) Speech Dataset.” (Figure 3). features, NB can be effective for sentiment analysis tasks.[21]
a) Audio Data Cleaning: The initial preprocessing step c) Random Forest: is a collective algorithm that con-
involved thorough cleaning of the audio recordings within structs numerous decision trees in the process of training and
the CaFE dataset. Background noise, artifacts, and potential merges their forecasts to generate the ultimate outcome. In
distortions were meticulously removed to enhance the clarity the context of sentiment analysis, Random Forest has the
and fidelity of the voice data. capability to apprehend intricate associations among character-

Authorized licensed use limited to: Indian Institute of Technology Palakkad. Downloaded on July 08,2025 at 14:07:23 UTC from IEEE Xplore. Restrictions apply.
istics and sentiments, consequently augmenting the precision intensity of each matrix element indicates the magnitude of the
of predictions.[22] corresponding frequency component. The Mel Spectrogram
d) Long Short-Term Memory: LSTM is an Recurrent serves as a fundamental feature for various audio-related tasks,
Neural Network(RNN) created to solve the vanishing gra- such as analyzing the sentiment in call center interactions.[27]
dient issue in sequence modeling. It can capture long-term b) MFCCs (Mel Frequency Cepstral Coefficients):
dependencies in sequential data, making it useful for sentiment MFCCs represent a succinct portrayal of an audio signal’s
analysis tasks involving text sequences..[23] spectral characteristics. Commonly employed in speech and
e) Multi-Layer Perceptron: a category of artificial neural audio processing, they capture the unique timbral and pho-
networks is recognized for its numerous layers of intercon- netic attributes of a sound. The MFCC extraction process
nected nodes. It encompasses an initial layer for input, one involves multiple steps, encompassing computation of the Mel
or more concealed layers, and a concluding output layer. The Spectrogram, followed by the application of Discrete Cosine
training of MLPs is achieved through the utilization of activa- Transform (DCT) to decorrelate the Mel frequency elements.
tion functions and backpropagation, rendering them adaptable The resultant coefficients outline the audio signal’s spectral
for a multitude of tasks, such as sentiment analysis.[24] envelope shape, effectively compressing information while
5) Text Vector Representation Techniques: Leveraging ap- retaining pertinent attributes. MFCCs find extensive use in
propriate text vector representation techniques is crucial for automatic speech recognition and emotion recognition tasks,
effective sentiment analysis and other NLP tasks. In this rendering them a valuable tool for extracting distinguishing
section, we delve into two widely used techniques: TF-IDF features from call center voice recordings.[27]
and CountVectorizer. Both the Mel Spectrogram and MFCCs play a pivotal role
a) TF-IDF (Term Frequency-Inverse Document Fre- in metamorphosing raw audio data into meaningful features
quency): TF-IDF is a fundamental text vectorization method that encapsulate the auditory traits of speech. These features
that evaluates the importance of words in a document relative empower sentiment analysis models to decode emotional ex-
to their frequency across a corpus. The technique combines pressions and sentiments conveyed through voice interactions
Term Frequency (TF), which measures the occurrence of in the call center milieu.
a word within a specific document, and Inverse Document 7) Comparison Metrics: To assess the performance of the
Frequency (IDF), which gauges the rarity of the word across sentiment analysis models applied to call center interactions,
the entire collection. [25] several key metrics are employed:
b) CountVectorizer: This method is a straightforward
a) Recall: Recall measures the model’s ability to cor-
and efficient approach for transforming textual data into
rectly identify actual positive instances, calculated as:
numerical vectors. Its functioning involves the creation of
a matrix, where each row represents a document and each
column represents a distinct word from the complete collection True Positives
Recall =
of texts. The entries in the matrix indicate the frequency at True Positives + False Negatives
which each word appears in its respective document.[26]
Both TF-IDF and CountVectorizer are valuable tools for b) F1-Score: is the harmonic mean of precision and
transforming raw text into structured numerical data, enabling recall, offers a balanced assessment of a model’s performance.
ML algorithms to effectively analyze and classify text docu- This metric takes into account both false positives and false
ments for sentiment analysis and various other applications. negatives and is particularly valuable when working with im-
6) Audio Feature Extraction: Efficient audio feature extrac- balanced datasets. It is computed using the following formula.
tion is crucial for converting raw audio signals into meaningful Precision × Recall
numerical representations that can be utilized for sentiment F1-score = 2 ×
Precision + Recall
analysis and other audio processing tasks. In this section, we
delve into two prominent techniques: the Mel Spectrogram and c) Precision: The precision metric determines the cor-
MFCCs. rectness of positive predictions by calculating the ratio of accu-
a) Mel Spectrogram: The Mel Spectrogram is a widely rately predicted positive instances to all instances predicted as
utilized technique in the field of audio analysis, which provides positive, which can be computed using the provided formula.
a visual representation of the frequency characteristics of an
audio signal over a period of time. Its effectiveness in captur- True Positives
ing variations in pitch and timbre makes it extremely valuable Precision =
True Positives + False Positives
for detecting subtle emotional nuances in speech. The Mel
Spectrogram divides the audio signal into short overlapping d) Accuracy: The accuracy of a model’s predictions is
intervals, calculates the magnitude of the Fourier Transform determined by the proportion of correctly classified instances
for each interval, and then maps the resulting spectrum onto out of the total number of instances.
the Mel scale—a frequency scale that is perceptually signifi-
cant. This process generates a two-dimensional matrix where True Positives + True Negatives
one axis represents time, another represents frequency, and the Accuracy =
Total Instances

Authorized licensed use limited to: Indian Institute of Technology Palakkad. Downloaded on July 08,2025 at 14:07:23 UTC from IEEE Xplore. Restrictions apply.
 B. voice-based Analysis

Fig. 4. Performance of Text Analysis Algorithms with CountVectorizer.

Fig. 6. Performance of voice Analysis Algorithms with melspectrogram.

IV. R ESULTS AND D ISCUSSION
A. Text-based Analysis Diving into the realm of voice-based analysis using the
melspectrogram, the MLP model emerged as the frontrunner
When utilizing the CountVectorizer for feature extraction, with an accuracy score of 0.72. An intriguing facet was the
the models demonstrated comparable performances. Specif- NB model’s impeccable recall of 1.00. While this might seem
ically, both the SVM and LSTM algorithms displayed the impressive at first glance, it warrants careful scrutiny as it
highest accuracy, achieving a score of 0.74. The recall metric, might insinuate the model’s propensity to overfit or possess a
which measures the model’s ability to correctly identify all bias towards the positive class. Precision values, which were
relevant instances, showed the LSTM slightly ahead with a mainly sandwiched between 0.60 and 0.67, didn’t showcase
score of 0.75. When evaluating precision, which represents the any standout performer. However, in terms of F1 scores, the
accuracy of the positive predictions, all the models gravitated MLP and LSTM shined once more.
around the 0.70-0.72 range. The F1 scores, which balance pre-
cision and recall, further accentuated the superior performance
of the SVM and LSTM models.

Fig. 7. Performance of voice Analysis Algorithms with mfccs.

When the voice-based features were represented using

Fig. 5. Performance of Text Analysis Algorithms with TfidfVectorizer. MFCCs, a significant shift in model performance was evident.
The RF algorithm notably outpaced its competitors with an
For the TfidfVectorizer, the performances slightly dwin- impressive accuracy of 0.78. Mirroring the previous set’s trend,
dled in comparison to the CountVectorizer results. The SVM the MLP boasted a flawless recall score of 1.00. However,
achieved the pinnacle of accuracy with a score of 0.73. as mentioned earlier, such perfect scores require cautious
An interesting observation arose in the recall department, interpretation. The RF maintained its supremacy in precision
with the MLP model showcasing the highest value at 0.74. with a score of 0.77, and similarly, the F1 scores portrayed it
Precision scores remained consistent with the previous set, as the best performer.
floating within the 0.69-0.72 spectrum. Finally, the F1 scores In conclusion, for text-based analysis, while there were
reemphasized the balanced and steady performances of all slight discrepancies between the two vectorization techniques,
models. SVM and LSTM consistently emerged as top contenders. On

Authorized licensed use limited to: Indian Institute of Technology Palakkad. Downloaded on July 08,2025 at 14:07:23 UTC from IEEE Xplore. Restrictions apply.
the contrary, in voice-based analysis, feature representation [7] L. Zhao et al., “Call center customer complaint pre-
played a pivotal role in discerning the best models: MLP and diction with topic modeling and sentiment analysis,” in
LSTM for mel spectrogram and RF for MFCCs. However, International Conference on Web Information Systems
it’s paramount to consider factors such as interpretability, Engineering, 2012.
efficiency, and real-world application needs before finalizing [8] C. Cortes and V. Vapnik, “Support-vector networks,”
any model for deployment. Moreover, while standard metrics Machine learning, vol. 20, no. 3, pp. 273–297, 1995.
provide a robust understanding of model performance, an in- [9] D. Jurafsky and J. H. Martin, Speech and language
depth examination, especially regarding edge cases, remains processing. 2019.
indispensable. [10] S. Hochreiter and J. Schmidhuber, “Long short-term
memory,” Neural computation, vol. 9, no. 8, pp. 1735–
V. C ONCLUSION
1780, 1997.
The integration of sentiment analysis in call centers offers [11] J. Devlin et al., “Bert: Pre-training of deep bidirec-
a promising opportunity to enhance the customer experience tional transformers for language understanding,” arXiv
by introducing an emotional dimension into automated interac- preprint arXiv:1810.04805, 2018.
tions. Through the use of ML and NLP techniques, it becomes [12] B. Schuller et al., “A tutorial on paralinguistics in
possible to detect customer sentiments from text and voice speech and language processing,” Proceedings of the
signals, paving the way for tailored responses and intelligent IEEE, 2011.
call routing. The experiments have shown promising results, [13] Z. Zhang et al., “Pattern recognition in speech and
but there are challenges to overcome, including adaptation to language processing,” Electrical Engineering & Au-
different accents and languages, as well as ongoing improve- tomation, 2017.
ment of sentiment analysis models. The future of automated [14] D. A. Reynolds and R. C. Rose, “Robust text-
call centers could be much more than mechanical interaction. independent speaker identification using gaussian mix-
With the addition of artificial empathy, these systems could ture speaker models,” IEEE transactions on speech and
become true partners in delivering exceptional customer ser- audio processing, 1995.
vice. [15] U. Zölzer, Ed., DAFX: Digital Audio Effects. John
In future endeavors, we plan to advance our research by Wiley & Sons, 2011.
focusing on the seamless integration of the sentiment analysis [16] M. El Ayadi, M. S. Kamel, and F. Karray, “Survey
module within the callbot architecture. This integration aims to on speech emotion recognition: Features, classification
provide real-time monitoring and analysis of customer interac- schemes, and databases,” Pattern Recognition, vol. 44,
tions, offering valuable insights into the callbot’s performance no. 3, pp. 572–587, 2011.
and the emotional dynamics of the conversations. [17] E. Ribeiro et al., “Real-time speech emotion recognition
R EFERENCES using gpu and its applicability in neurology,” Health
Information Science and Systems, vol. 3, no. S1, 2015.
[1] I. Aattouri, M. Rida, and H. Mouncif, “A comparative [18] “French twitter sentiment analysis dataset.” (), [Online].
study of learning algorithms on a call flow entering of Available: https://www.kaggle.com/datasets/hbaflast/
a call center,” 2021. DOI: 10.1007/978-3-030-73103- french-twitter-sentiment-analysis.
8 36. [19] “Canadian french emotional (cafe) speech dataset.” (),
[2] I. Aattouri, M. Rida, and H. Mouncif, “Creation of a [Online]. Available: https://zenodo.org/record/1478765.
callbot module for automatic processing of a customer [20] C. Cortes and V. Vapnik, “Support-vector networks,”
service calls,” 2021. DOI: 10.1007/978-3-030-76508- Machine Learning, vol. 20, no. 3, pp. 273–297, 1995.
8 30. [21] A. Y. Ng and M. I. Jordan, “The optimality of naive
[3] I. Aattouri, H. Mouncif, and M. Rida, “Modeling of bayes,” Advances in neural information processing sys-
an artificial intelligence based enterprise callbot with tems, vol. 14, pp. 849–856, 2001.
natural language processing and machine learning al- [22] L. Breiman, “Random forests,” Machine learning,
gorithms,” IAES International Journal of Artificial In- vol. 45, no. 1, pp. 5–32, 2001.
telligence (IJ-AI), vol. 12, pp. 943–955, 2 2023. DOI: [23] S. Hochreiter and J. Schmidhuber, “Long short-term
10.11591/ijai.v12.i2.pp943-955. memory,” Neural computation, vol. 9, no. 8, pp. 1735–
[4] B. Liu, Sentiment analysis and opinion mining (Synthe- 1780, 1997.
sis lectures on human language technologies 1). 2012, [24] D. E. Rumelhart, G. E. Hinton, and R. J. Williams,
vol. 5, pp. 1–167. “Learning representations by back-propagating errors,”
[5] A. Pak and P. Paroubek, “Twitter as a corpus for Nature, vol. 323, no. 6088, pp. 533–536, 1986.
sentiment analysis and opinion mining,” in LREC, 2010. [25] K. S. Jones and S. Walker, “A study of automatic text
[6] B. Wuthrich et al., “Daily stock market forecast from classification,” Journal of Documentation, vol. 53, no. 1,
textual web data,” in IEEE International Conference on pp. 1–32, 1997.
Systems, Man, and Cybernetics, 1998.

Authorized licensed use limited to: Indian Institute of Technology Palakkad. Downloaded on July 08,2025 at 14:07:23 UTC from IEEE Xplore. Restrictions apply.
[26] C. D. Manning and H. Schütze, Foundations of Sta-
tistical Natural Language Processing. The MIT Press,
1999.
[27] J. P. Bello and L. R. Rabiner, “Mel frequency cepstral
coefficients for music modeling,” in 2010 IEEE Inter-
national Conference on Acoustics, Speech and Signal
Processing (ICASSP), IEEE, 2010, pp. 5542–5545.

Authorized licensed use limited to: Indian Institute of Technology Palakkad. Downloaded on July 08,2025 at 14:07:23 UTC from IEEE Xplore. Restrictions apply.