基于最大似然线性回归矩阵的说话人识别算法研究

Research on MLLR Based Speaker Recognition Algorithm

研究了将自适应领域的最大似然线性回归(Maximum likelihood linear regression,MLLR)变换矩阵作为特征进行文本无关的说话人识别算法.本文引入了基于统一背景模型的MLLRSV-SVM说话人识别算法,并在此基础上进行高层音素聚类以进一步提高识别性能.在采用多种信道补偿技术后,在NISTSRE2006年1训练语段-1测试语段同信道和跨信道数据库上,基于MLLR特征的系统与其他最好的系统性能接近并有很强的互补性,经过简单线性融合可以极大提高识别性能.

This paper uses the maximum likelihood linear regression （MLLR） as feature for text-independent speaker recognition algorithm. We introduce a universal background model （UBM） based MLLRSV-SVM algorithm first, and then extend the algorithm to multi-class for improvement. After channel compensation, in terms of the NIST 2006 SRE lconv4w-lconv4w/mic corpus, the MLLR based system is comparable with and complementary of the state of the art systems. The performance is greatly improved by simply linear fusion.