The Research Of Speech Emotion Recognition Based On The Fusion Features

Emotion recognition is an important part of emotion computing, it has raised much more concern in recent years and speech is one of the most significant means of communication, which carries a lot of emotional information, so the research about speech emotion recognition is very important. The technology of speech emotion recognition can help improve the human-computer interactive ability of the computer. This paper mainly studies the speech emotion recognition based on the fusion features and in this paper the recognition ability of common features are analyzed and we improve the recognition ability of the fusion features in two ways:fusion in decision level and feature level. We use the improved adaptive weighted fusion SVM-KNN to finish fusion in decision level and Deep Belief Network(DBN) to finish fusion in feature level. The main work and contributions of this paper are shown as follows:(1) This paper outlines the research background and significance of speech emotion recognition and make a summary of the research of the definition of emotion, emotional database, emotional features, feature dimension reduction methods and emotion classifiers.(2) An emotional database which contains five kinds of emotions including joy, anger, sadness, fear and neutral is established and the voice signals are test to ensure their effectiveness and then they are preprocessed. After that, we extracted 261 features from the signals. All of this is the basis of the research about speech emotion recognition.(3) Make comparison between common features and reduce the dimension of the feature set. In this paper we use J1 criterion based on Fisher criterion to compare features and use the scatter plot drew by features calculated by LDA and KNN algorithm to verify the conclusion. Using minimum redundancy maximum relevance algorithm to get the best subset for further study.(4) Studying a method to combine the result of the improved SVM-KNN classifiers in the decision level. We first improve the SVM-KNN classifier by using the combined kernel function and QPSO, then use the adaptive weighted fusion algorithm to fuse the result of these classifiers. By this way, this paper succeeded to raise the recognition rate.(5) Studying a method to finish fusion in feature level and designing two different models called DBN21 and DBN22. We first use the DBN21 model to combine the traditional features in the Chapter 3 and then compare it with the one not use the DBN21.The result proves that this method can raise the rate of emotion recognition rate. What's more,we extract new features of spectrogram using the STB/Itti model and improve it by DBN21 and then use the SVM-KNN to test the effectiveness of the features.To strengthen the ability of emotion recognition we combine the traditional features and the new features by DBN22 and also test the effectiveness of the fusion features.