Fully Bayesian speaker clustering based on hierarchically structured utterance-oriented Dirichlet process mixture model

摘要

We have proposed a novel speaker clustering method based on a hierarchically structured utterance-oriented Dirichlet process mixture model. In the proposed method, the number of speakers can be determined from the given data using a nonparametric Bayesian manner and intra-speaker variability is successfully handled by multi-scale mixture modeling. Experimental result showed that the proposed method is computationally-efficient and effective in speaker clustering. The proposed method significantly improve the accuracy of speaker clustering systems as compared with the conventional method, particularly for the case in which the number of utterances varied from speaker to speaker.