Speech emotion recognition is an important branch in artificial intelligence. It has been widely used in the field of human-computer interaction and machine intelligence. How to use the effective feature selection algorithm to select the acoustic feature subset with strongly discriminant ability for enhancing the accuracy of speech emotion recognition plays an important role. In the speech emotion recognition, different acoustic features often have different discriminant ability for different emotion recognition. Some features even play a decisive role for certain emotion recognition, but can’t provide useful information for differentiating other emotion classes. So, selecting a common acoustic feature subset for all class leads to the result that feature subset is not optimal for individual emotion label.In order to enhance the accuracy of speech emotion recognition, this thesis proposes a novel class-dependent feature selection model to improve the performance of speech emotion recognition through transforming multi-classification into many binary classifications. But unlike class-independent feature selection, class-dependent feature selection can select each subset owning discriminant ability for each speech emotion category. Then, under the model we adopt the markov blanket technique to select discriminative features for each emotion class, and use the support vector machine as classifier to build the training model and prediction model of emotion recognition. At last, to solve the vote conflict problem in multiple classifier fusion, this thesis proposes that the class output of the support vector machine is transformed to the output of probability confidence, and combine with the output of class and the output of probability to predict the emotion class corresponding to each test sample.Extensive experimental results on a publicly available dataset show that in comparison with information gain, principal component analysis and class-independent feature selectors, the proposed method significantly reduces the feature dimensionality of original acoustic feature space and improves the accuracy of speech emotion recognition. In addition, the results of the Pearson correlation coefficient analysis show that class-dependent feature selection methods can remove these features not related to target class and reduce the redundancy between selected feature columns... |