Content-based audio segmentation using support vector machines

Abstract

Audio exists at everywhere, but is often out-of-order. It is necessary to arrange them into regularized classes in order to use them more easily. It is also useful, especially in video content analysis, to segment an audio stream according to audio types. In this paper, we present our work in applying support vector machines (SVMs) in audio segmentation and classification. Five audio classes are considered: silence, music, background sound, pure speech, and non-pure speech which includes speech over music and speech over noise. A SVM learns optimal class boundaries from training data to best separate between two classes. A sound clip is segmented by classifying each sub-clip of one second into one of these five classes. Experiments on a database composed of clips of 14870 seconds in total length show that the average accuracy rate for the SVM method is much better than that of the traditional Euclidean distance based (nearest neighbor) method.