Feature selection for emotion recognition of mandarin speech

In order to improve the accuracy of speech emotion recognition and reduce the feature dimension,four feature selection methods,i.e.promising first selection(PFS),sequential forward selection(SFS),sequential backward selection(SBS) and stepwise discriminant analysis(SDA) were employed to select acoustic features.After feature selection for the emotional mandarin database with speakertext-independent group and speaker-dependent group,two classifiers: linear discriminant analysis(LDA) and support vector machine(SVM) were used to compare the four feature selection methods according to the emotion recognition performance.Experimental results show that pitch,log energy,speed and the 1st formant are the most important factors for the feature selection.Meanwhile the discrimination of the other features changes with different speaker.SDA is the best among the four feature selection methods in both LDA and SVM,and the best feature number is 9 to 12.The results also show that the feature selection efficiency changes with training sample number,and the feature selection method is more effective in the condition of small training sample number.