Human speech emotion recognition via feature selection and analyzing

Abstract

Abstract Speech emotion recognition is one of the important research topics in the field of multimedia processing and human-machine interface. To obtain the most influential features of the speech data for emotion recognition, in this paper, 64 statistical features of the speech signal including short-term energy, pitch, frame, format, and spectrum energy were extracted with speech emotion database. Mean Impact Value (MIV) and the improved Correlation-based Feature Selection (CFS) were employed to select the most influential feature set. BP neural network (BPNN) was used to identify the accuracy. The proposed MIV-CFS method selected the features related to speech emotion, with less recognition error, the recognition accuracy all higher than 88%, and the highest recognition accuracy is 91.61%.