Impostor Selection for SVM Models Training in Speaker Verification
Jiping Xiong · Jisuanji gongcheng · 2009
In text-independent Support Vector Machine(SVM) speaker verification,impostor selection for SVM training directly determines its efficiency and performance.This paper proposes two Gaussian Mixture Model(GMM)-based methods for impostor selection.By GMM likelihoods,the most similar impostors to the target speaker are selected for SVM training,which makes the target speaker models more discriminative.Experiments on text-independent SVM speaker verification in NIST'04 1side-1side data show significant improvement.