Mixture Component Clustering for Efficient Speaker Verification

In speaker verification (SV) systems based on a support vector machine (SVM) using Gaussian mixture model (GMM) supervectors, a large portion of the test-stage computational load is the calculation of the a posteriori probabilities of the feature vectors for the given universal background model (UBM). Furthermore, the calculation of the sufficient statistics for the mean also contributes substantially to computational load. In this paper, we propose several methods to cluster the GMMUBM mixture components in order to reduce the computational load and speed up the verification. In the adaptation stage, we compare the feature vectors to the clusters and calculate the a posteriori probabilities and update the statistics exclusively for mixture components belonging to appropriate clusters. Our results, demomstrate that (on average) we can, reduce the number of a posteriori probability calculations by a factor up to 2.8× without loss in accuracy.