Speaker Independent Emotion Recognition Using HMMs Fusion System with Relative Features

Speaker independent emotion recognition is particularly difficult for the individual differences of acoustic character and culture background. So, relative features obtained by calculating the features change of emotion speech relative to natural speech are adopted to weaken the influence from the individual differences in the paper. Moreover, an improved ranked voting fusion system is proposed to combine the decisions from four hidden Markov model (HMM) classifiers which are based on different feature vectors respectively. The recognition results of the provided algorithm have been compared with the isolated HMMs with absolute features, by Berlin database of emotional speech, and the average recognition rate has reached 78.4% in speaker independent case.