A neural network approach for human emotion recognition in speech

In this paper, we present a language-independent emotion recognition system for the identification of human affective state in the speech signal. A corpus of emotional speech from various subjects, speaking different languages is collected for developing and testing the feasibility of the system. The potential prosodic features are first identified and extracted from the speech data. Then we introduce 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. The selected features are employed as the input to a Modular Neural Network (MNN) to realize the classification of emotions. The proposed system gives quite satisfactory emotion detection performance, yet demonstrates a significant increase in versatility through its propensity for language independence.