Automatic segmentation of broadcast news audio using self similarity matrix

Audio broadcast news is composed of music, commercials, news from correspondents and recorded statements; apart from the newsreader who reads out the news during the news program. Automatic segmentation of news audio is essential, for automatic time alignment of audio and the available text transcription to build speech corpora. In this paper we address this particular problem of segmentation of broadcast news audio, in order to extract the audio segments corresponding to the newsreader. We observe that the existing systems and techniques proposed in literature tend to face difficulties when used on our task of automatic segmentation and extraction of the newsreader segments; and thus produce sub- optimal results. Hence, we propose a technique which identifies the actual acoustic change points using an acoustic Self Similarity Matrix. We discuss the proposed algorithm, which uses a two pass technique, and present experimental results on broadcast news audio from All India Radio.