Factor Analysis and Majority Voting Based Speech Emotion Recogntion

There are some problems to be resolved for speech emotion recognition, such as the dimension of feature sets is usually too high and the redundancy among various features is relatively stronger. Considering these problems, the factor analysis and majority voting based speech emotion recognition was proposed. How to extract emotional factors from global statistical features and GMM super vectors was researched. Several classifiers were adopted to perform the majority voting based decision fusion. Experimental results demonstrated that factor scores are stronger correlated with emotional states and the majority voting based method effectively enhanced the cohesion of training data.