Performance of the Vocal Source Related Features from the Linear Prediction Residual Signal in Speech Emotion Recognition

Abstract

Researchers concerned with Speech Emotion Recognition have proposed various useful features associated with their performance analysis related to emotions. However, a majority of the studies rely on acoustic features, characterized by vocal tract responses. The usefulness of vocal source related features has not been extensively explored, even though they are expected to convey useful emotion-related information. In this research, we study the significance of vocal source related features in Speech Emotion Recognition and assess the comparative performance of vocal source related features and vocal tract related features in emotion identification. The vocal source related features are extracted from the Linear Prediction residuals. The study shows that the vocal source related features contain emotion discriminant information and integrating them with vocal tract related features leads to performance improvement in emotion recognition rate.