Speech emotion recognition based on DNN-decision tree SVM model

Abstract

Motivated by the development of DNN technology, a speech emotion recognition method based on DNN-decision tree SVM model is proposed. The proposed method can not only excavate the deep emotion information of the speech signal, but also extract more distinctive emotion features from the easily confused emotions. In this method, the decision tree SVM structure is firstly constructed by computing the confusion degree of emotion, and then different DNN are trained for diverse emotion groups to extract the bottleneck features that are used to train each SVM in the decision tree. Finally, speech emotion classification is realized based on this model. This model is assessed by using the Chinese Academy of Sciences Emotional Corpus. The experiment results show that the average emotion recognition rate based on the proposed method is 6.25% and 2.91% higher than traditional SVM and DNN-SVM classification method, respectively. It is proved that this method can effectively reduce the confusion between emotions, thus improving the speech emotion recognition rate.