Enhanced Speaker Verification Using GMM-Supervector Based Modified Adaptive GMM Training

摘要

In this paper, an enhanced speaker verification is proposed by exploring a novel modified adaptive Gaussian mixture model (GMM) training. Based weight factor of observation called the observation reliability; we propose to apply a modified Expectation maximization (EM) algorithm, combined with a modified Maximum a posteriori (MAP) estimation to train the modified adaptive GMM model. Using this proposed model, we generate GMM-supervectors which are combined with SVM for verification task. We evaluate performance of speaker verification system based the proposed approaches on utterances from Korean movie database ("You came from the stars"). Experimental results demonstrate that our proposed approaches can outperform the standard GMM-UBM and GMM-supervector approaches in noise conditions.