SVM-based News Audio Classification

In the age of digital information, with the great improvement of multimedia processing, rapid increase in speed and capacity of computers and widespread internet, more and more digital multimedia information including image, video, and audio has became one part of our daily life. It presents new challenges for managing and searching large digital multimedia information collections. Therefore, content-based video and audio processing and retrieval have been the focus of the researches in multimedia application, information retrieval and data management.The existing systems of content-based video retrieval place more emphasis on the structure and content of the video than audio, while the audio information always provides assistant information for content changes. So the audio information, except for content based in audio retrieval, is an important part in video retrieval systems. Raw audio data is non-semantic and non-structured binary data stream, which lacks semantic content description and structured organization. How to extract structured information and semantic content from raw audio data and choose the optimal classification method are key points in the research.This paper provides a SVM-based news audio classification system. It contains four subsystems, including speech/non-speech/silence SVM, pure speech/non-pure speech SVM, male/female SVM and music/noise SVM, which can classify the audio into male speech, female speech, non-pure speech, noise, music and silence. Audio features are firstly analyzed in frame level and clip level, including MFCC, SF, SR, FE, SED, SC, HZCRR, LSTER features and so on, then employs sequential forward selection to select the most appropriate features. Also SVM is chosen as the classifier in this system. The experimental results show that the performance is as good as expected.