采用M-矢量和支持向量机的说话人确认系统

Speaker verification system based on M-vector and support vector machine

将UBM子空间中的说话人MLLR自适应得到的M-矢量应用于SVM中,提出了一种新的说话人确认系统.该系统有效地将扰动属性映射算法整合到SVM核函数中,实现在核空间中直接对M-矢量进行信道补偿,从而提高系统对信道干扰的鲁棒性能.实验结果表明:相比传统基于音素类的MLLR-SVM和基于I-矢量的I-vector-SVM基线系统,在不需要大量有文本内容标注的语音数据、复杂度和运算量都很高的自动语音识别系统、因子空间统计量的估计的情况下,本系统可获得与最好的基线系统几乎相当的性能,同时还表现出很强的互补特性.在NIST SRE2008说话人评测数据库上测试结果表明:提出系统的性能与基于I-矢量的说话人确认系统的性能接近,并表现出很强的互补性,融合后的等错误率相对下降了13.3%.

A new speaker verification system based on the M-vectors and support vector machine (SVM ) was proposed in this paper .The M-vectors were derived from multiple maximum likelihood linear regression (MLLR) speaker transformations which were calculated for a given speech data with respect to each subspace of the universal background model (UBM ) .Furthermore ,a nuisance attrib-ute projection was introduced into the SVM kernel space to project the M-vectors into a speaker-de-pendent space ,to alleviate the channel and session variability during training and testing .Compared with the traditional phone-class based MLLR-SVM and I-vector -SVM systems ,experimental results show that the proposed system can achieve almost the same good performance as the best baseline sys-tem without large factor statistical calculations and any automatic speech recognition system w hich needs large labeled training data ,complexity and computations .In the NIST SRE2008 evaluation task ,the proposed system can achieve almost the same performances as the state-of-the-art I-vector based system .Large complementary information has been demonstrated in a relative 13.3% EER re-duction after system fusion .