Improved Speech Emotion Recognition in Bengali Language using Deep Learning
Syem Aziz, Nokimul Hasan Arif, Sakif Ahbab, Sabbir Ahmed, Tasnim Ahmed, Md. Hasanul Kabir
Abstract
Speech Emotion Recognition (SER) is an evolving field at the intersection of artificial intelligence and signal processing. Despite there being notable advancements leveraging sophisticated Machine Learning (ML) techniques, the focus has largely been on widely spoken languages like English, resulting in limited exploration of under-resourced languages like Bengali. In this connection, this work proposes a Convolutional Neural Network (CNN) based solution for SER in the Bengali language utilizing Mel Frequency Cepstral Coefficients (MFCC) features and data augmentation techniques. The proposed model achieved an impressive accuracy of 90% and 78%, respectively, on the SUBESCO and BanglaSER datasets. Several models have been devised to attain promising results on these datasets individually; however, they fail to exhibit consistent performance across both datasets. In contrast, the proposed approach demonstrated commendable performance when applied to both datasets, showcasing its robustness in the field of emotion recognition from Bengali speech signals.