Sample iterative likelihood maximization for speaker verification systems

Paper i proceeding, 2010

Gaussian Mixture Models (GMMs) have been the dominant technique used for modeling in speaker recognition systems. Traditionally, the GMMs are trained using the Expectation Maximization (EM) algorithm and a large set of training samples. However, the convergence of the EM algorithm to a global maximum is conditioned on proper parameter initialization, a large enough training sample set, and several iterations over this training set. In this work, a Sample Iterative Likelihood Maximization (SILM) algorithm based on a stochastic descent gradient method is proposed. Simulation results showed that our algorithm can attain high loglikelihoods with fewer iterations in comparison to the EMalgorithm. A maximum of eight times faster convergence rate can be achieved in comparison with the EM algorithm.