The Research On Speech Emotion Recognition Based On Fuzzy Support Vector Machines

In the intelligent man-machine interactive system,voice emotion recognition is one of the research hotspots,and support vector machine(SVM)method is widely used in speech emotion recognition.However,it is a binary classification method.When the emotional confusion is large,it is not only necessary to construct the multivariate classifier,but also some non-separable regions in the classification process,which can affect the classification results.Many classification problems in real life is highly unbalanced characteristics,when the data set of positive and negative samples imbalance is larger,the traditional support vector machine for the minority class recognition effect is very poor,is sensitive to noise and outlier is also at the same time,influence the final classification result.So this article in view of the traditional fuzzy support vector machine(SVM)is sensitive and the support vector classification problem such as membership value is not precise enough,by putting forward a new design method of membership function to improve the algorithm principle,after the speech emotion library do recognition.The main research contents are as follows:(1)Introduced about prospects of statistical learning theory and the principle of SVM algorithm based on statistical learning theory,explain in detail the implementation process,raises the concept of fuzzy support vector machine(FSVM)at the same time,gives the fuzzy support vector machine model in fuzzy mathematics involved in the related definitions and theorems.(2)The characteristic parameters of single feature and feature fusion is studied to identify the impact of network,through the experiment to select optimal feature fusion method,and then studied based on the fuzzy support vector machine algorithm of speech emotion recognition method,through the combination of fuzzy theory of support vector,to improve learning and generalization ability of support vector machine method to solve the nonlinear and small sample and high dimensional pattern recognition problems,effectively reduce the effects of outliers and noise points,improve the classification accuracy of emotional speech.(3)By studying the existing algorithm of FSVM,in terms of principle on the improvement of designing new samples give weight method,this paper proposes a new class based in hyperplane distance measure membership function,in order to realize effective outstanding support vector and weakened the effect of noise,and give the accurate weight for each sample points,through the experiment for all kinds of algorithm performance comparison,verify the validity of the algorithm,and then to the applicability of the new membership function is discussed.(4)In view of the unbalanced emotion datasets,the traditional fuzzy support vector machine is sensitive and the support vector classification problem such as membership value is not precise enough,the introduction of theunbalanced adjustment factor,the comprehensive considering the similar sample status and the distribution of sample point under the influence of neighboring sample density,to the applicability of the fuzzy support vector machine algorithm is improved,and applied to the balance of speech emotion recognition.The results show that compared with the traditional fuzzy support vector machine,the algorithm has a TYUT2.0 emotion with a sample imbalance of3.89 The performance of the voice database was improved by 4.95%,and the classification of CASIA Chinese emotional corpus was improved by 10.57% for the unbalanced rate of 13.28.The greater the imbalance rate of the sample,the more obvious the classification effect.