A sticky HDP-HMM with application to speaker diarization

We consider the problem of speaker diarization, the problem segmenting an audio recording of a meeting into temporal corresponding to individual speakers. The problem is particularly difficult by the fact that we are not to assume knowledge of the number of people in the meeting. To address this problem, we take a nonparametric approach to speaker diarization that on the hierarchical Dirichlet process hidden Markov model(HDP-HMM) of Teh et al. [J. Amer. Statist. Assoc.101 (2006) 1566–1581]. Although the basic HDP-HMM tends over-segment the audio data—creating redundant states and switching among them—we describe an augmented HDP-HMM provides effective control over the switching rate. We also that this augmentation makes it possible to treat emission nonparametrically. To scale the resulting to realistic diarization problems, we develop a algorithm that employs a truncated approximation of the process to jointly resample the full state sequence, improving mixing rates. Working with a benchmark NIST set, we show that our Bayesian nonparametric architecture state-of-the-art speaker diarization results.