区分性训练在声纹密码中的新应用

Novel Application of Discriminative Training in Vocal Password

在声纹密码任务中由于数据稀疏的问题难以实现区分性训练,本文以一种表征距离度量的特征矢量为基础提出新的声纹密码区分性系统框架,对正反例样本的新特征矢量实现了基于最小分类错误准则的区分性训练,将声纹密码从确认问题转化为二类分类问题。在自由说话风格的60人数据集上,声纹密码区分性系统与混合高斯模型-通用背景模型(Gaussian mixture model-universal background model,GMM-UBM)系统融合后等错误率为4.48%,相对GMM-UBM,动态时间规划(Dynamic time warping,DTW)基线系统性能分别提升了17.95%和59.68%。

Due to data sparsity, discriminative training has not been successfully applied to the system of vocal password up to now. Therefor, a novel vocal password framework based on a specific pre-processing strategy is proposed. The new feature is used to represent the distance measure and the problem caused by data sparsity can be solved to some extent. As a consequence, the vocal password is actually transferred from verification to binary classification and the discriminative training of two class models is sueeessfully accomplished on the minimum classification error criteria. After fusing the discriminative system with Gaussian mixture mod- el-universal background model（GMM-UBM） system, the equal error rate （EER） performance decreases to 4.48%, relatively 17.95% and 59.68% lower than the GMM-UBM and the dynamic time warping（DTW） system respectively on the corpus including 60 speakers. The experiment results show that the new application of discriminative training in the vocal password system is feasible and effective.