Speech emotion recognition based on wavelet packet coefficient model

Abstract

Conventional features have achieved good performance in speech emotion recognition. However, these features are based on short-time analysis without considering the non-stationary properties. In this paper we focus on wavelet packet techniques, which can provide an improved signal representation with a tradeoff between time and frequency resolution. We propose a wavelet packet coefficient model in speech emotion recognition. The wavelet packet coefficients at five decomposition levels are analyzed and used as input features to Support Vector Machine (SVM) classifiers. The performances of these features are evaluated for seven emotional states in two languages, German and Chinese. Results demonstrate that these wavelet packet coefficients features show improvement in emotion recognition performance compared with conventional Mel-Frequency Cepstral Coefficients (MFCC) features.