BAYESIAN ANALYSIS OF SIMILARITY MATRICES FOR SPEAKER DIARIZATION

Inspired by recent success of speaker clustering in Total Variabil-ity space we propose a new probabilistic model for speaker diariza-tion based on Bayesian modeling of pairwise similarity scores. The recordings are represented by symmetric similarity matrices of like-lihood ratio scores from probabilistic linear discriminant analysis (PLDA) trained on short-term i-vectors. We employ Bayesian ap-proach to address the problem of unknown number of speakers in conversation. Diarization error rates on the NIST 2008 SRE tele-phone data indicate comparable performance with state-of-the-art eigenvoice-based diarization. But unlike the eigenvoice approach, our method finds the number of speakers automatically, making the proposed model more viable for practical applications. Index Terms — Speaker diarization, variational Bayesian infer-ence, similarity matrix 1