Neural Networks for Language Independent Emotion Recognition in Speech

Abstract

This chapter introduces a neural network based approach for the identification of human affective state in speech signals. A group of potential features are first identified and extracted to represent the characteristics of different emotions. To reduce the dimensionality of the feature space, whilst increasing the discriminatory power of the features, a systematic feature selection approach which involves the application of sequential forward selection (SFS) with a general regression neural network (GRNN) in conjunction with a consistency-based selection method is presented. The selected parameters are employed as inputs to the a modular neural network, consisting of sub-networks, where each sub-network specializes in a particular emotion class. Comparing with the standard neural network, this modular architecture allows decomposition of a complex classification problem into small subtasks such that the network may be tuned based on the characteristics of individual emotion. The performance of the proposed system is evaluated for various subjects, speaking different languages. The results show that the system produces quite satisfactory emotion detection performance, yet demonstrates a significant increase in versatility through its propensity for language independence.