基于得分域多维特征分类器的声纹密码系统

A Vocal Password System Based on Multi-Dimension Feature Classifier in Score Domain

针对不同类型数据对目标发音人区分能力不同的现象,在传统系统基础上提出利用UBM模型对测试数据进行分类,使用分类后的似然比得分形成多维特征,在此基础上利用SVM分类器进行声纹密码确认.该方法把传统的似然比检验策略转换成多维特征空间上的二类分类问题.测试与注册数据同信道情况时,在4种手机数据集上,文中系统相对文本相关GMM-UBM声纹密码系统等错误率分别下降41.25%、33.33%、37.49%和26.03%,在交叉信道上系统性能也获得改善.

As providing the score of different types of data the same weight, the average likelihood ratio verification measure used in GMM-UBM vocal password system brings decline in the system performance. Based on the different distinguished capacity between data types, a method is proposed in the score domain which classifies the test data by UBM, combines the likelihood ratio score of each class to form new multi-dimension feature, and then implements speaker verification by SVM. By use of the proposed strategy, the traditional likelihood ratio test is converted into a two-class classification problem in the multi-dimension feature space. The equal error rate of the proposed system is relatively 41. 25%, 33.33%, 37.49% and 26.03% less than that of text-dependent GMM-UBM system in the co-channel experiments on four telephone corpuses respectively. The improvement of performance is also demonstrated through the cross-channel experiments.