话者识别的信道补偿

Model Session Variability in Speaker Verification

在文本无关的说话人识别中,训练与测试语音中信道环境的差异是影响其性能最重要的因素.近年来,利用因子分析对信道建模成为说话人识别领域的重要方法,大大降低了说话人确认的错误率,但运算复杂度限制了实时的应用.本文介绍了一种简化的因子分析方法:首先在混合高斯模型的模型域训练信道空间,然后在特征域进行信道补偿,得到的新特征可用于各种系统.在NIST2006的数据库上,利用本文的方法相对基线系统在等错误率上有31%的降低.

In the text-independent speaker recognition systems, the variability of the channel and environment is the most important factors affecting the performance. More recently, factor analysis has been proposed to model session variability and has provided impressive reductions in verification error rates, but the computation load prevents its application in real-time cases. In this paper, we introduce a simplified factor analysis technology： At first the channel sub-space is trained in the Gaussian Mix Model domain, then the session variability is compensated in feature domain. The transformed feature can be used by any other systems. In the NIST 2006 SIRE corpus, the equal error rate（EER） of the proposed system can reduce by 31% against the baseline GMM system.