A Two-level Method for Unsupervised Speaker-based Audio Segmentation

In this paper, we propose a two-level segmentation method that detects speaker changes in a continuous audio stream effectively. In our approach, we divide the change detection process into two levels: region level that detects the potential change regions containing candidate speaker change points, and boundary level that searches and refines the true change points. At the region level, we employ the modified Generalized Likelihood Ratio (MGLR) metric to search for the potential change regions in continuous local windows. At the boundary level, we perform T2 and Bayesian Information Criterion (BIC) algorithm to detect segment boundaries within the potential windows. The experimental results on the 1997 Broadcast News Hub4-NE mandarin corpus show the efficiency of the proposed scheme.