Exploring speech features for classifying emotions along valence dimension

Abstract

Naturalness of human speech is mainly because of the embedded emotions. Today's speech systems lack the component of emotion processing within them. In this work, classification of emotions from the speech data is attempted. Here we have made an effort to search, emotion specific information from spectral features. Mel frequency cepstral coefficients are used as speech features. Telugu simulated emotion speech corpus (IITKGP-SESC) is used as a data source. The database contains 8 emotions. The experiments are conducted for studying the influence of speaker, gender and language related information on emotion classification. Gaussian mixture models are use to capture the emotion specific information by modeling the distribution. An average emotion detection rate of around 65% and 80% are achieved for gender independent and dependent cases respectively. © 2009 Springer-Verlag Berlin Heidelberg.