A Hybrid Speech Emotion Perception Method of VQ-based Feature Processing and ANN Recognition

This paper constructs a VQ/ANN (vector quantization/artificial neural network) based speech emotion recognition system. The system first extracts the basic prosodic parameters and Mel-frequency cepstral coefficients (MFCC) frame by frame. Recent researches reveal that MFCC convey detailed emotional relevant information of syllable. However, the statistic measures of MFCC confuse the information at sentence level. Therefore, this paper proposes a VQ-based method different to statistic method to generate measures of MFCC. Then the combination of VQ-based MFCC measures and the statistic measures of prosodic parameters is used as input feature vector. The ANN is performed to process the combination features and the statistic measures of all extracted parameters respectively. The experiment results reveal that the combination features outperform the statistic measures. More detailed analysis indicates that the combination features could characterize the emotion space better than the statistic features. Besides, the rationality of VQ/ANN based framework is also demonstrated.