A scanning window scheme based on SVM training error rate for unsupervised audio segmentation

Audio segmentation has applications in a variety of contexts, such as automatic broadcast news transcription, audio information retrieval, and as a pre-processing step in automatic speech recognition (ASR). The Support vector machine (SVM), as a binary classifier, is commonly used for supervised audio signal segmentation and classification. In this study, inspired by the idea of scanning window, we present and evaluate an unsupervised audio segmentation approach based on the SVM training error rate. The approach is unsupervised in the sense that it does not require prior knowledge of audio classes. Experimental results indicate that the segmentation technique outperforms traditional Bayesian information criterion (BIC), generalized likelihood ratio (GLR), and Gaussian mixture models (GMM) methods, particularly in detecting audio landmarks of short duration.