IMPROVED SPEAKER SEGMENTATION AND SEGMENTS CLUSTERING USING THE BAYESIAN INFORMATION CRITERION

Abstract

Detection of speaker, channel and environment changes in a continuous audio stream is important in various applications (e.g., broadcast news, meetings/teleconferences etc.). Standard schemes for segmentation use a classifier and hence do not generalize to unseen speaker/channel/environments. Recently S.Chen introduced new segmentation and clustering algorithms, using the so-called BIC. This paper presents more accurate and more efficient variants of the BIC scheme for segmentation and clustering. Specifically, the new algorithms improve the speed and accuracy of segmentation and clustering and allow for a real-time implementation of simultaneous transcription, segmentation and speaker tracking.