Effects of feature domain normalizations on text independent speaker verification using sorted adapted Gaussian mixture models

Abstract

In this paper we evaluate sorted Gaussian Mixture Model (GMM) system performance for Text Independent Speaker Verification under the feature domain normalization conditions. Sorted GMM is a speed-up algorithm proposed for GMM based systems. Cepstral Mean Subtraction (CMS) and Dynamic Range Normalization (DRN) are the normalization schemes studied for sorted GMM system purposes. Effectiveness of these normalizations has been proved in speaker recognition systems while their effectiveness on the speed-up of GMM based speaker verification is showed in this study. The baseline system is a universal background model-Gaussian mixture model (UBM-GMM) system and evaluations were performed on the NIST 2002 speaker recognition evaluation database with NIST SRE rules. It is shown that CMS and DRN normalizations enhance both the baseline system and sorted GMM system performances. In other words, the performance loss due to reducing the computational load is mitigated by applying CMS and DRN. © 2008 Springer-Verlag.