Detection of Acoustic Change-Points in Audio Streams and Signal Segmentation

Abstract

This contribution proposes an efficient method for the detection of relevant changes in continuous stream of sound. The detected change-points can then serve for the segmentation of long audio recordings into shorter and more or less homogenous sections. First, we discuss the task of a single change-point detection using the Bayes decision theory. We show that it leads to a quite simple and computationally efficient solution based on the Bayesian Information Criterion. Next, we extend this approach to formulate the algorithm for the detection of multiple change-points. Finally, the proposed algorithm is applied for the segmentation of broadcast news audio-streams into parts belonging to different speakers or different acoustic conditions. Such segmentation is necessary as the first step in the automatic speech-to-text transcription of TV or radio news.