Recognition Of Emotion In Speech Using Spectral Patterns

Abstract

Recent developments in man-machine interaction have intensified the need for recognizing human’s emotion from speech. In this study we proposed using Spectral Pattern (SP) and Harmonic Energy (HE) features for the automatic recognition of human affective information from speech. These features were extracted from the spectrogram of the speech signal using image processing techniques. A filter and wrapper feature selection scheme was used to avoid the curse of dimensionality. Here, a hierarchical classifier is employed to classify speech signals according to their emotional states. This classifier is optimized by the Fisher Discriminant Ratio (FDR) to classify the most separable classes at the upper nodes, which can reduce the classification error. Moreover, a tandem classifier is employed to increase the recognition rate of highly confused emotions pairs. Our experimental results have demonstrated the potential and promise of SPs and HEs for emotion recognition. The proposed method was tested on the male and female speakers separately and the overall recognition rate of 86.9% is obtained for classifying seven emotion categories in the Berlin database.