Component score weighting for GMM based text-independent speaker verification

GMM/UBM framework is wildly used in Automatic Speaker Verification (ASV), however, due to the insufficiency of the training data, both the hypothesized speaker and impostors are not well modeled, especially to some of the Gaussian component mixtures. Thus, the Gaussian mixtures in each GMM model have different discriminative capabilities, and the mismatch between testing and training data will also aggravate this situation. In this paper, we propose a novel approach, namely, Component Score Weighing (CSW), to reweight the Gaussian mixtures and highlight those which have high discriminative capability by post-processing the log-likelihood ratio (LLR). The original log-likelihood in GMM systems is assigned to each Gaussian component mixture, deriving two component score serials, which we called the dominant score serial and the residual score serial. A nonlinear score weighting function is then applied to reweigh those scores, respectively. Experiments on NIST 2006 SRE corpus show that, this approach achieves notable performance gains over our previous baseline system (about 12% relative improvement in minimum detection cost function (DCF) value).