Analysis And Identification Of Emotion Specific Features For Speaker Independent Emotion Recognition System Using Gaussian Mixture Models (GMMs)

摘要

Emotion recognition from speech is yet challenging and important task. Emotion specific features for emotion classification using speech signals is addressed in this paper. Emotional speech is produced by human speech production mechanism and the emotion information in the speech signal is the combination of both excitation source (prosody features) and the vocal tract system (spectral features). Emotion recognition performance decreases by using either spectral or prosody features alone. We propose emotion specific features by combining both spectral and prosody features for Emotion Recognition. Experiments were conducted for different combinations of spectral and prosody features for classifying four emotions viz. angry, fear, happy and neutral using Gaussian Mixture Models (GMMs) It is identified that the combinations of Fundamental Frequency (FO), Pitchchroma(PC), Formants and Mel Frequency Cepstral Coefficients (MFCC gives better recognition performance. The database used for the study is IITKGP.