基于深度神经网络的说话人信道自适应方法  

Channel Adaptation Based on Deep Neural Networks for Speaker Verification

作　　者：龙艳花[1] 倪继锋[1] 叶宏[1] 

机构地区：[1]上海师范大学电气信息系,上海200234

出　　处：《四川大学学报（工程科学版）》2016年第2期151-155,共5页Journal of Sichuan University (Engineering Science Edition)

基　　金：上海市青年科技英才扬帆计划资助项目(14YF1409300)

摘　　要：针对说话人确认中的复杂信道环境干扰问题,提出一种基于深度神经网络的信道自适应方法。该方法首先在不同信道类型下训练得到音素信息相关的深度神经网络模型(deep neural networks,DNNs),将说话人语音的声学特征参数在这些DNNs上进行自适应,得到各信道类型下的深瓶颈特征(deep bottleneck feature,DBF)。然后将这些参数进行拼接并通过PCA降维,最后采用目前最有效的基于身份认证矢量(identity vector,i-vector)的建模技术对降维后的DBF进行建模,得到目标说话人模型和测试语音段的i-vector矢量用于最终说话人确认打分判决。在NIST SRE2010核心评测数据库上的实验结果表明,利用提出的方法能有效消除信道干扰对说话人确认的影响,在很大程度上提升了基于i-vector的说话人确认基线系统的性能。In order to handle the channel condition distortions between train and test speech in speaker verification,based on the deep neural networks,a channel adaptation approach was proposed. First,several phonetic deep neural networks（ DNNs） were trained on the speech datasets with different types of channel conditions. The acoustic features derived from speaker utterances were then adapted to obtain deep bottleneck features（ DBFs） using these DNNs. DBFs were concatenated and a feature dimension reduction was performed using PCA. Finally,these DBFs were modeled by the identity vector（ i-vector） modeling technique which is the most popular and efficient approach for speaker verification. The achieved i-vectors for target speaker and test utterances were then used to achieve the final verification scores. Results on the NIST SRE2010 coretest evaluation task demonstrated that compared to the i-vector baseline system,the proposed approach is effective to eliminate channel distortions for speaker verification,and achieves significant performance improvements.

关 键 词：信道自适应 深度神经网络 深瓶颈特征 i-vector 说话人确认 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]