Comparative Analysis of Speech Features for Speech Emotion Recognition

Abstract: In this paper, comparative analysis of speech cepstral features is performed to recognise emotion. We identify two effective feature namely, Mel Frequency Cepstral Coefficient (MFCC) and Cochlear Filterbank cepstral coefficients extracted from speech signal. MFCC as a baseline approach is compared to the feature extracted from cochlear filterbank with zero crossing at the output of each channel. Extracted features are fed to Support Vector Machine (SVM) classifier. As shown in our results, cochlear feature provide highest recognition accuracy provided using linear kernel. It gives 89.67% classification accuracy for Berlin Emotional Speech Database. A study on noise robustness of above mentioned feature was also carried out. MFCC and cochlear feature have recognition accuracy of 81.9% and 86% respectively in clean testing conditions with RBF kernel function but when mismatch between training and testing set increases as in real time situations, recognition accuracy of MFCC feature is 11% while cochlear feature gives accuracy of 25%, which shows that cochlear feature is more robust to noise.

Keywords: Cochlea, cochlear filterbank, mel filterbank, mel scale, noise robustness, linear polynomial and RBF kernel, speech emotion recognition, Zero Crossing Peak Amplitude, SVM.