Speech Emotion Recognition of Tamil Language: An Implementation with Linear and Nonlinear Feature

Abstract

In the past decades, speech emotion recognition has made remarkable footprints in human–computer interaction and affective computing. Conventional techniques use acoustic, prosodic, lexical, and paralinguistic features and proved the importance and accuracy rate. Also, K-nearest neighbor, support vector machine, other standard classifiers are used to classify the different emotions. In this paper, video and audio signals are fetched from the Tamil speaking people, and the emotions of the people are recognized with good recognition rates. Both video and audio signals are preprocessed to remove the noise using the bandpass filter. Temporal feature extraction is used to improve the recognition rate. Remarkable comparison is made with nonlinear (Hurst parameter) and linear feature parameters (wavelet packet cepstral coefficient). To avoid false classification, operational, and computational cost, the dimensionality of the feature set is reduced with principal component analysis. The selected features are classified with KNN, GMM for five emotion classes, namely happy, sad, angry, fear, and disgust. The proposed mixed/hybrid model of KNN, HMM, and GMM is implemented to recognize the emotions, and the results are compared with traditional KNN and GMM. Also, our present database is compared with different standard databases like Tamil Speech Data-ASR, CMU-INDIC, and LDC-IL.