Speech Emotion Recognition Based On Support Vector Machine

With the rapid development of science and technology,artificial intelligence gradually appears in people's perspectives,and the dialogue between computers and people becomes a reality.At the same time,people also hope that computers can understand people's emotions.Therefore,People also try to find ways to use computer technology to identify human emotions,that is,speech emotion recognition technology.At present,the application of speech emotion recognition technology has become more and more obvious,which is mainly used in medical field,service industry and education.With the development of artificial intelligence,more and more studies on speech emotion recognition show its theoretical value and application value in all aspects of human life.Firstly,the thesis studies the typical classification of speech emotion and the characteristic parameters of several affective types,and analyzes the extraction method and process of the characteristic parameters such as pitch frequency law,amplitude energy,formant,Mel cepstral coefficient.Secondly,A method of selecting characteristic parameters based on principal component analysis is proposed,which improves the recognition rate of primary features.Then the paper studies the processing methods of conventional speech signals,including the digital processing of speech signals,pre-emphasis,framing,windowing and endpoint detection preprocessing.Then,the paper studies the SVM model with high recognition rate for small sample classification.Then the Fisher criterion is used to adjust the interval between classes and the maximum entropy principle criterion is used to effectively achieve the uniform distribution within the class to improve the slow training of traditional SVM kernel parameters.Finally,the paper chooses five kinds of emotions,happy,angry,scared,sad and calm from the Berlin voice database.They are tested and verified by support vector machines,Fisher criterion optimization kernel parameters,maximum entropy principle to optimize class kernel parameters,and the Fisher criterion and maximum entropy principle fusion scheme.The test shows that the proposed method can obviously improve the training speed of nuclear parameters without affecting the recognition rate of several speech emotion signals.