Variance-Based Gaussian Kernel Fuzzy Vector Quantization for Emotion Recognition with Short Speech

Automatic recognition of emotion is becoming an important part in the design of process for affect-sensitive human-machine interaction (HMI) systems. This work proposes variance-based Gaussian kernel fuzzy vector quantization (VGKFVQ) method for speech emotion recognition. By non-linear kernel mapping, it mapped the data into the high-dimensional feature space, and made the dissimilarity among different emotions enlarged. VGKFVQ used the clustering centers to form the codebooks, and employed the minimum overall average fuzzy weighted vector quantization error (FWVQE) rule to classify emotions: happiness, anger, neutral and sadness. VGKFVQ used membership to present the ambiguous of an unknown emotion instead of a single hard label compared with non-fuzzy method such as Support Vector Machine (SVM) algorithm and sample variance replaced the dispersion parameter in the Gaussian kernel to realise adaptively adjustment of the parameter. Experimental results show that the recognition rate of this method is higher than SVM method with short speech as well as Fuzzy C-means Clustering Vector Quantization (FVQ) method.