Variational Bayesian speaker change detection

In this paper we study the use of Variational Bayesian (VB) methods for speaker change detection and we compare results with the classical BIC solution. VB methods are approximated learning algorithms for fully bayesian inference that cannot be achieved in an exact form. They embed in the objective function (also known as free energy) a term that penalizes more complex models. Experiments are run on the Hub4 1996 evaluation data set and show that the VB outperforms the BIC of almost 7%. Anyway as long as the decision must be taken on a limited amount of data the VB based method must be tuned as the BIC based method in order to produce reasonable results.