Relative Speech Emotion Recognition Based Artificial Neural Network

Abstract

Artificial neural network (ANN) models based on static features vector as well as normalized temporal features vector, were used to recognize emotion state from speech. Moreover, relative features obtained by computing the changes of acoustic features of emotional speech relative to those of neutral speech were adopted to weaken the influence from the individual difference. The methods to relativize static features and temporal features were introduced individually and experiments based Germany database and Mandarin database were implemented. The results show that the performance of relative features excels that of absolute features for emotion recognition as a whole. When speaker is independent, the hybrid of relative static features vector and relative temporal features normalized vector achieves the best results.