论文部分内容阅读
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.
Aiming at the different distinguishing ability of target pronunciation between different types of data, the UBM model is proposed to classify the test data based on the traditional system, and the likelihood ratio score is used to form multi-dimensional features. Based on this, SVM classifier is used This method converts the traditional likelihood ratio test strategy into the second-class classification problem in multidimensional feature space. On the four mobile phone data sets, the system relative text correlation GMM -UBM vocal cipher system error rates decreased by 41.25%, 33.33%, 37.49% and 26.03%, respectively, and the cross-channel system performance also improved.