Towards a more efficient SVM supervector speaker verification system using Gaussian reduction and a tree-structured hash

Speaker verification (SV) systems that employ maximum a posteriori (MAP) adaptation of a Gaussian mixture model (GMM) universal background model (UBM) incur a significant teststage computational load in the calculation of a posteriori probabilities and sufficient statistics. We propose a multi-layered hash system employing a tree-structured GMM which uses Runnalls’ GMM reduction technique. The proposed method is applied only to the test stage and does not require any modifications to the training stage or previously-trained speaker models. With the tree-structured hash system we are able to achieve a factor of 8× reduction in test-stage computation with no degradation in accuracy. Furthermore, we can achieve computational reductions greater than 21× with less than 7.5% relative degradation in accuracy.