Speech Emotion Regognition Based On Mylti-Layer Classification

Nowadays, the development of artificial intelligence is highly developed, to make the machine really intelligent, it is too important to let the machine understand the emotional of human beings. In this paper, the study of speech emotion recognition is based on the Berlin speech emotion database. This article proposes a SVM-based multi-layer classifier, the proposed multi-layer SVM classifier improves the average accuracy by 5.42%. Then the author use PCA to optimize feature, and use the selected features for speech emotion recognition, the average recognition rate than the traditional SVM method promote 5.24%. We proposes a multi-layer SVM emotion classification algorithm combined with PCA at the end, compare to the traditional SVM method, the average recognition rate of this method increased 7.85%.This article provides a new voice activity detected method, we also discuss several emotional parameters including Mel-Frequency Cepstral Coefficients, pitch, formant, Delta, short-time zero-crossing rate short-time energy and so on. We use the traditional SVM method for speech emotion recognition obtained an average recognition rate of 58.69%. The article draws out the concept base on confusion matrix, then proposes a multi-layer SVM classifier. The first layer separates the emotions which can be easily distinguished, whereas the second one separates the emotions which cannot. Emotions are separated layer by layer, the proposed multi-layer SVM classifier improves the average accuracy by 5.42%, thus proves the effectiveness of the multi-layer strategy.However, the algorithm has high requirements on the running speed and storage space because of the umpteen features of the speech. The article combines the PCA and multi-layer SVM method to obtain the average recognition rate more than 5.24% over the traditional SVM method. At the last, we combine the multi-layer SVM with PCA, compare to the traditional SVM method, and obtained an average recognition rate of 66.54%, the average recognition rate of this method increased 7.85%. Thus proves the correctness of the proposed method.