A comparative study on Bengali speech sentiment analysis based on audio data

Abstract

Sentiment analysis is one of the most researched areas for every language. Due to the rise of AI, the use of speech in every sector is rapidly growing so is the importance of Speech Sentiment Analysis. Despite being the seventh most spoken language in the world, Bengali speech sentiment analysis studies are not much enriched. This study compared the Bengali speech sentiment analysis using machine learning and CNN, LSTM, and Bi-LSTM models. We have used the SUBESCO and BanglaSER datasets for training our models where the KNN model outperformed other models with an accuracy of 90%. Later, we evaluated the performance of the models with our custom-made test dataset. Experimental results show that AdaBoost and Bi-LSTM model performed best with 45% accuracy. Moreover, to understand the feature effect on the output, we used the interpretable SHAP model in the ML model outcomes as they provide the best results allowing us to have an explainable advantage to determine the results.