利用i-vectors构建区分性话者模型的话者确认

Discriminative Speaker Models Based on i-vectors for Speaker Verification

对于电话手机语音的文本无关话者确认,运用联合因子分析构建话者信息子空间与信道信息子空间来进行失配信道补偿取得了较好的效果.然而研究表明,信道信息子空间仍然包含了可以用来区分话者的信息.因此,本文运用一种既包含话者信息又包含信道信息的全变量信息子空间来提取i-vectors低维特征矢量,再运用类内协方差规整进行失配信道补偿,最后用补偿后的i-vectors特征矢量构建支持向量机话者模型.在NIST08数据库上实验表明,本文所构建系统的性能在等误识率和最小检测代价函数上有相对近70%的提高.

Joint Factor Analysis provides an effective means for text independent speaker verification system. It is a powerful technique for compensating the variability caused by different channels and speakers. However, studies show that, the channel information sub- space also contains information that can be used to distinguish between speakers. In this study, we propose a new speaker representa- tion called i-vectors which is a low-dimensional vector. Firstly , it is extracted from a total variability space which models both the speaker and channel variability. Then, within this total variability space, Within-Class Covariance Normalization, a common used channel compensation method, is performed to reduce the channel variability. Finally, the compensative i-vectors are used to train discriminative models based on Support Vector Machines. Experiments on NIST08 SRE database show that the proposed strategy can improve the system performance as much as 70% both in EER and MinDCF over the baseline system.