Speaker verification based on improved updates to the SVM
Ren-Hua Wang · Journal of Tsinghua University(Science and Technology) · 2008
In support vector machine(SVM) based text-independent speaker verification system,the traditional average supervector does not give enough information to discriminate the voices.Improved discrimination was obtained using weighted and mean updates rather than the universal background model(UBM) as inputs to the SVM during training and testing.The algorithm gives better performance than the standard Gaussian mixture model(GMM) mean supervector and the GMM-UBM system with the 2006 NIST 1conv4w-1conv4w SRE corpus,with the EER reduced by 22% compared to the baseline GMM-UBM system.The results show that the weighting parameter can be optimized to improve the SVM-based speaker verification.In addition the supervector has more discrimination information when the effect of the UBM is not taken into consideration.