Development of a Bengali Speech-based Emotion Analysis System

Detecting emotion from speech is one of the most demanding works in speech recognition and processing. This paper presents speech emotion recognition in the Bengali language by extracting features from speech samples and matching them with a feature database. In this work, five (5) classes of emotions are detected using Deep-Learning and Machine-Learning based algorithms on a data set in Bengali. The classes of emotion defined are Happiness, Sadness, Threat, Panic, and Neutral. The extracted features are Mel Frequency Cepstral Coefficient (MFCC), Chromagram, Mel- spectrogram, spectral roll-off, spectral centroid, and Teager Energy Operator (TEO) is applied to them to enhance the details of extracted spectral features, increasing the algorithm’s accuracy. Simulation results show that the devised method outperforms the existing speech emotion recognition method.