一种基于说话者话路变化的主成分分析方法

A PCA Method Based on Speaker Session Variability

在文本无关的说话人确认中,训练与测试语音中信道环境的不匹配是一种说话者话路变化问题.这种不匹配会严重降低说话人确认系统的性能.为了有效解决该问题,本文提出一种基于说话者话路变化的主成分分析方法,将其应用在说话者确认中,我们将这种方法称为面向话路变化的主成分分析方法.这种方法能够与类内协方差归一化结合,进一步提高识别效果.在NIST2006年说话者识别数据库上进行实验,证明该方法不仅在系统识别等错误率上比基线系统有了24.2%的降低,而且在计算复杂度上相对于目前传统的方法也有很大的优势.

In the text-independent speaker verification systems, the mismatch and variability of the channel and environment between training and testing is a session variability problem. It can greatly degrade the speaker recognition performance. To deal with the problem more efficiently, a modified PCA method is proposed called session variation principal component analysis （SVPCA） which can integrate with within class covariance normalization （WCCN）. In the NIST 2006 verification task, the proposed method is compared with our previous baseline general linear discriminative sequence-support vector machine （GLDS-SVM） system. The experimental results show a relative reduction of up to 24.2% in error equal ratio （EER）. Moreover, the proposed method has advantages in computational and memory costs, compared with the state-of-art systems.