Speech Emotion Recognition Based on Multi-Output GMM and SVM

For the poor ability of discrimination in the case of recognizing emotion by using GMM model, an algorithm based on multi-output GMM and SVM is proposed, which combines the advantages of both GMM and SVM. The multidimensional output of GMM for one test speech are regarded as feature of emotion for SVM. This method takes advantage of the statistical properties of characterization of GMM and the strong discrimination ability of SVM. Experimental results on emotional speech databases demonstrate that the proposed method achieves significant improvements about 2% to 4% than standard GMM on speech emotion recognition.