Author(s): Bimal Kumar Kalita and Pran Hari Talukdar
This paper works on the techniques for automated categorization of spoken sounds based on the emotional condition of the speaker. The information employed for the study comes from a corpus of selecte d natural dialogs recorded in different emotions deployed by Speech Works. In this study, Gaussian class-conditional likelihood distribution with respect to linear discriminant classification (LDC) and K-nearest neighborhood (K-NN) schemes are used to classify utterances into basic emotion states, positive and negative. The utterance level statistics of the fundamental frequency and the energy of the speech signal are used by the classifiers. The promising first selection and forward feature selection are used for feature selection to improve classification performance. The dimensionality of the features reduced to maximize classification accuracy. Improvements achieved up to 87%. Keywords- Bodo, LDC, K-NN, Gaussian.