Non-spectral features-based robust speaker independent emotion recognition from speech signal

Abstract

To ensure the better and effective human-machine interaction, affective computing is playing a vital role in the current scenario. Since speech and glottal signals convey the information about the emotional state of the speaker in addition to the linguistic data, it is essential to recognise the speaker's emotions and respond to it expressively. This paper mainly discusses the effectiveness of non-spectral features and modelling techniques to develop a robust multi-speaker independent speaker's emotion/stress recognition system. Since our EMO-DB database is small it has become a challenging task to improve the performance of the system. The proposed non-spectral features and modelling techniques provides 81% and 100% as weighted accuracy recall for the stress recognition system concerning the classification done on individual models and emotion-specific group models. Weighted accuracy recall is found to be 100% for the classification done on emotion-specific group models by considering the utterances from the SAVEE database.