Bayesian Approach to Text-independent Speaker Verification

In this paper, we propose a novel approach to speaker verification. One of the problems in conventional speaker verificaion techniques based on the likelihood ratio test (LRT) is that the detection performance varies widely for each hypothesized speaker when the decision threshold is held fixed. In order to cope with the problem, we incorporate the distribution of the log likelihood ratio (LLR) into the decision rule. The LLR distribution conditioned on each speaker hypothesis is approximated by a Gaussian probability density funciton (PDF) where the mean and variance are estimated from a set of training data. Experiments on text-independent speaker verification show the superioriy of the proposed approach compared to the conventional LRTbased method.