Study of prosodic feature extraction for multidialectal Odia speech emotion recognition

Abstract

In this paper a speaker-independent and text-dependent speech emotion recognition system has been presented for the Cuttacki, Sambalpuri and Berhampuri dialects of the Odia language. A dialect is any distinguishable variety of a language spoken by a group of people. Emotions provide naturalness to speech. Here prosodic features are extracted from speech and used for classification of emotions. Prosodic features are represented by pitch, energy, duration, and formant. In order to evaluate the system performance for prosodic features the Orthogonal Forward Selection(OFS) algorithm is used for significant feature selection, and the Gaussian Mixture Model(GMM) and Support Vector Machine(SVM) for classification. The analysis of results, after significant features were found using the OFS algorithm, shows that SVM is a better classification algorithm compared to GMM. The study also shows distinctions between emotions of males and females after feature extractions.