Research On Key Issues Of Mandarin Speech Emotion Recognition

How to realize natural Human Computer Interaction (HCI) is one of the most important research directions of computer science. Since the speech recognition is the key approach of the HCI, and the recognition of emotion in the speech is so important that to accomplish the HCI with natural mode. With the development of psychology, physiology, neuroscience and computer technology, the affective comuting, especially the speech emotion recognition had great progress both in theoretical research and particular application. There are a lot of methods about the emotion defination and category, the acoustic features related with emotion, and classification model. Develop serval emotion recognition system with different language. There are pilot studies with in the speech emotion recognition frameworks. Howere, with the demand of the high performance in speech emotion recognition, and the application's requirement, the existent technologies and methods could not meet the request. Especially the lack of research in mandarin speech emotion recognition, we should work hard even more, and supply the gap.Four challenges existed in speech emotion recognition: Build a mandarin emotional speech database which meets the requirement of the number, the quality, the management and diversity; Search for a set of acoustic feature vector which have strong relationship with the emotion state; Reduce the disturbance of result from the difference of speaker and text, and to shorten the distance of the intra-class; Decrease the dimension of the features using feature selection and dimensionality reduction to select the best subset of features which are most important in distinguishing emotions, and to improve the generalization ability of the classifier.Based the background of natural HCI applications, this paper research the speech emotion recognition techology, especially the characteristic of mandarin speech emotion recognition and some unsolved problem. The paper proposed a novel method for speech emotion recognition, the kernel of the method is consis of feature selection, relative features and emotion focus. The main content includes as follows:(1) Mandarin emotional speech database. Because the research of mandarin speech emotion recognition was at the beginning, there is a lack of the technology of mandarin emotional speech database. So the paper built a mandarin emotional speech database, and the corpus was collect from studio recording and movie clips. The total number of the database is 1376, which consist of 5emotions include anger, fear, happy,...