Speech Emotion Recognition Based On Filter-wrapper Model

In the study of speech emotion recognition,a large number of effective speech emotion features are proposed to improve the accuracy of speech emotion recognition.However,in the process of the feature extraction and feature fusion,the feature set usually contains noise or redundant data,which not only increase the cost of speech emotion calculation,but also affect the accuracy of emotion recognition.Therefore,it is necessary to do feature selection of speech emotion.According to the evaluation criterion of subset,feature selection mainly includes two categories: filter and wrapper.The filter model relies on a variety of fast evaluation criteria to pick up the subset,it has fast subset searching speed.But its classification of speech emotion is not high.The wrapper model takes the classification performance as the evaluation criterion directly,so the subset has higher emotion recognition.On the other hand,search process with wrapper is related to the classifier closely,the time complexity is higher than filter.Based on the analysis of the advantages and disadvantages of the filter and the wrapper model,this study proposed a hybrid search model with filter and wrapper for speech emotion feature selection.According to this model,the experiment combines information gain and the harmony search algorithm to select the subsets.Firstly,the information gain as the filter part,it can quickly remove a large number of noise and redundant data because of its quick search ability.The pre-filtered subset with filter will be entered into the wrapper for further selection.The wrapper is composed of harmony search and support vector machine.And the speech emotion recognition adopts 10-fold validation.In our experiments,four feature sets are extracted from two emotional corpus of EMODB and EESDB.And conducting the feature selection with method which proposed in our study on the extracted features.The results show that the feature selection method we proposed can greatly reduce the dimension of the features and have the high recognition accuracy as the original data.Compared with the single strategy model-wrapper,our model reduced about half running time.