Research On The Segmentation And Classification For Broadcast Audio

Nowadays the large vocabulary speech recognition have played a high recognition rate in the quiet environment. However in order to put speech recognition into a more broad application,there is a strong need of system robustness and calculation speedBroadcast audio as a ordinary audio which is complicated and it is different from the audio in the laboratory. For example,broadcast audio contains a variety of audio elements,such as vioce, music, long silent segement ,niose ,etc. How to extract audio structure and content is the basis of deeper process of audio information ,information retrieval and improvement of system robustness . Now the research on the segmentation and classification for broadcast audio has become one of the most hottest topic.The main topic of this paper is the discrimination of speech and music. A speech music discrimination system based on support vector machines has been built by the specific features. In addition ,we study pitch which is a common feature of audio ,and use it to distinguish between speech and music. We also do some experiment to test the system.Besides, this article summarizes the main method of audio segmentation and build a Speaker Change Detection system using the sequential metric-based segmentation method via BIC.The study work of this thesis provides the preparation for the integrated audio segmentation system and and pushes forward further studies.