A SVM-Based Audio Event Detection System

This paper proposes a SVM-based method to deal with the problem of detecting audio events(cheering and applause) by audio analysis. In our framework, a sliding window is first used to pre-segment the audio stream into short segments by moving from start to the end. Second, various kinds of audio features are extracted to represent different audio sounds in each segment. Third, SVM(super vector machine) is used as the classifier to detect audio events. Finally, smoothing rules are used to eliminate the false alarms caused by background noise. By integrating all the techniques, an average F value of 79.71% is achieved in the audio detection task evaluated on nearly 8 hour TV programs. In this study, we discuss the complementarity of various kinds of audio features for the audio event detection task. We also compare the result with the GMM-based audio event detection system.