A Two-Stage Multi-Feature Integration Approach to Unsupervised Speaker Change Detection in Real-Time News Broadcasting

This paper presents a two-stage multi-feature integration approach for unsupervised speaker change detection in real-time news broadcasting. We integrate MFCC and LSP features (i.e. a perceptual feature plus a articulatory feature) in the metric-based potential speaker change detection stage to collect speaker boundary candidates as many as possible. We adopt a weighted Bayesian information criterion (BIC) to integrate boundary decisions from MFCC and LSP features in the speaker boundary confirmation stage. This multi-feature integration strategy makes use of the complementarity between perceptual features and articulatory features to achieve a performance gain. Speaker change detection experiments show that the multi- feature integration approach significantly outperforms the individual features with relative improvements of 26% over the LSP-only approach and 6% over the MFCC-only approach.