Speech Emotion Recognition Research Based On Feature Selection

Recently, with the increase of the interest in human computer interaction (HCI), emotion recognition is becoming a hot research. Speech emotion recognition is to analyze and detect the special emotion state from given emotion speech emotion utterance and then to ascertain the subject's specific inborn emotion, which can achieve smarter and more natural interaction between human beings and computers. The study of speech emotion recognition has found its values in economy and society.In this thesis, we firstly discuss the background and then analyze the main exiting speech emotion feature extraction algorithms, feature selection algorithms and emotion recognition algorithms. After analyzing the methods currently used by others, we present the feature selection method based on contribution analysis algorithm of neural network, the confusion degree computation method of emotion category and speech emotion recognition algorithm based on selective feature and SVM decision tree. The main work is described as bellows:(1) Feature selection method based on contribution analysis algorithm of neural network is presented. There are many emotion features, including redundant features and irrespective features, so we present feature selection method based on contribution analysis algorithm of neural network in order to reduce the effect of redundant information, which also decrease the workload for calculation and increase the recognition rate. In this method, 101 features are extracted from the voice of quality, energy, Fundamental Frequency, Formant Frequency, MFCCs and Mel Frequency Energy Dynamic Coefficient, and then the features extracted are selected by contribution analysis algorithm of neutral network. Furthermore, the effectiveness of the selected features is analyzed by cluster analysis to demonstrate the effectiveness of the feature selection method.(2) Confusion degree computation method of emotion category based on incorrect recognition rate is presented. In emotion categories, some emotion states are similar and difficult to be differentiated, whereas some others are different and easier to be differentiated. In this method, the incorrect recognition rate between emotions is got by multi-SVMs, based on which, the confusion degree between the emotions and the confusion degree between groups are calculated. By calculating confusion degree between emotions and confusion degree between groups, the easily confused emotion states are grouped, and the minutely difference between emotion categories can be surveyed further, which can increase the recognition rate. (3) Speech emotion recognition algorithm based on selective features and SVM decision tree is presented. Since the contribution to different speech emotion states differs from speech emotion features, during the training and recognition for each emotion state, better contribution features for each emotion state are used. Meanwhile, according to confusion degree between the emotions, the easily confused emotions are classified as a group, which set up the intermediate node of SVM decision tree. Then each emotion group is classified by using better contribution features. In addition, for each emotion utterance, only the features in the set, which is the union of better contribution features for each emotion state, need to be extracted, which reduces the calculation and improves the recognition speed.(4) By combining Matlab with VC++, a prototype system of speech emotion recognition based on selective features and SVM decision tree is implemented, which demonstrates the effectiveness of the algorithms mentioned above.