Experiment on fast scoring for GMM based speaker verification

Each speaker is modeled with a mixture of Gaussians in Gaussian mixture model (GMM) based speaker recognition. During verification, a match score is computed between the test feature vectors and the claimant speaker model. In order to make a fast verification, each feature vector might be scored only against the most likely mixtures instead of all mixture components of the model. The most likely mixtures might be selected during the universal background model (UBM) scoring. In this paper, we test this method using two separate text-dependent, Turkish speaker recognition databases. In our experiments, we observed that the number of the most likely mixtures can be reduced to a few mixtures without degradation in verification accuracy. This reduction significantly improves the verification speed.