Real life emotion classification using VOP and pitch based spectral features

In this work, vowel onset points (VOPs) and pitch based spectral features are used for speech emotion classification. VOP is an anchor point from which vowel begins in a CV unit (generally a syllable). These are estimated using energy values of linear prediction (LP) residual, short time spectrum and modulation spectrum. Identification of vowel, consonant and CV transition regions of a syllable is done based on VOP locations. Spectral features are separately extracted from vowel, consonant and CV transition regions. Emotion recognition models (ERMs) are developed separately using these spectral features. Gaussian mixture models (GMMs) are used as classifiers. Pitch synchronous analysis is done on speech signal and spectral features are extracted from each pitch period to develop emotion models. Simulated emotion speech corpus, IITKGP-SESC is used as a database. Results of IITKGP-SESC are compared with those of Berlin emotion speech database Emo-DB. Emotion recognition performance observed is about 92% and 89% for IITKGP-SESC and Emo-DB respectively. Later the approach has been extended to the recognition of real life emotions, using IITKGP-movie emotion speech corpus (IITKGP-MESC). About 82% of emotion recognition performance is observed in this case.