Comparative study of machine learning and deep learning models for bengali speech sentiment analysis

Abstract

Sentiment analysis, a key area within Natural Language Processing (NLP), has garnered significant research attention across various languages. With the rise of Artificial Intelligence (AI), the use of speech in multiple sectors is expanding, highlighting the importance of Speech based Sentiment Analysis. Understanding emotions is crucial for many applications, as they form a fundamental component of human communication. Predicting human emotion from speech is a promising research topic with substantial potential for advancements in Human-Computer Interaction (HCI). Despite Bengali being the seventh most spoken language globally, research on Bengali speech sentiment analysis remains limited. In this study, we conducted a comparative analysis of sentiment analysis for the Bengali language using both machine learning and deep learning models. Specifically, we evaluated Random Forest Classifier (RFC), Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP) algorithms alongside Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) models. We utilized the SUBESCO and BanglaSER datasets to train our models. Our focus was on recognizing five emotional states: joy, sadness, anger, surprise and neutral. Initially, we employed Mel-Frequency Cepstral Coefficients (MFCC) feature extraction in Chapter 1, and later incorporated Mel-Frequency Cepstral Coefficients (MFCC) and Chroma Shift feature extraction methods in Chapter 2 to offer a comprehensive comparison between the models. Our experimental evaluations demonstrated the effectiveness of these algorithms in accurately predicting emotions in Bengali speech samples. The results indicated that the accuracy rates ranged mostly around 85% to 95%, with Multi-Layer Perceptron achieving the highest. The Majority Voting model, which combined the results from all models to enhance prediction accuracy, also performed effectively. The insights gained from this study provide valuable inputs for future explorations and improvements in the domain of Bengali speech sentiment analysis.