Speech Emotion Recognition Using Stationary Wavelet Transform and Timbral Texture Features

Nowadays, researchers are paying more attention to recognize the human emotions from various modalities such as facial features, physiological or biological signals and speech signals. In the last decade, a number of research works have been carried out on emotion recognition using speech signals. In this work, emotion recognition system was developed using the features extracted from the emotional speech signals (ES) and its glottal waveforms (GW). Inverse filtering and linear predictive analysis were used to derive GWs from the speech signals. ES and GW were decomposed into five levels using stationary wavelet transform (SWT) and timbral texture features were extracted from the decomposed SWT coefficients. A total of 288 features were extracted from ES and GW respectively. Two-stage feature reduction was proposed to reduce the number of features and also to enrich the discriminatory power of the extracted features. The raw and enhanced features were used as input for extreme learning machine (ELM) and k-nearest neighbor (kNN) based classifiers. Several experiments were carried out and the results shows that timbral texture features derived from the decomposed stationary wavelet coefficients can be used as possible cues for emotion differentiation.