Regroupement bayesien variationnel des locuteurs

In this paper we explore the use of Variational Bayesian (VB) learning in unsupervised speaker clustering. VB learning is a relatively new learning technique that has the capacity of doing at the same time parameter learning and model selection. We run experiments on the NIST 1996 HUB-4 evaluation test for speaker clustering. Two cases are considered : the speaker number is a priori known and it has to be estimated. We evaluate results in terms of average cluster purity and average speaker purity. VB shows a higher accuracy compared to the Maximum Likelihood solution.