基于本征音因子分析的短时说话人识别

Eigenvoice Factor Analysis in Short Time Speaker Recognition

提出了一种基于本征音因子分析的文本无关的说话人识别方法。它解决了训练语音与测试语音均很短的情况下,传统的基于最大后验概率准则的混合高斯模型无法建立稳定的说话人模型问题。首先利用期望最大化算法在开发集上训练出说话人的本征音载荷矩阵,在说话人模型建模时通过将短时语音数据向本征音空间的降维映射来得到模型参数。实验结果表明,在NIST SRE 2006数据库中的10 s训练语音-10 s测试语音任务中,在传统的混合高斯模型的基线系统上,通过采用本征音因子分析的方法可以使系统等错误率降低18%。

A text-independent speaker verification method is proposed based on eigenvoiee fac- tor analysis algorithm. It focuses on the short-duration text-independent speaker verification. The Gaussian mixture model （GMM）-universal background model （UBM） based on maximum a posteriori（MAP） estimation cannot work when the training and test speech data are sparse. Firstly, the eigenvoice loading matrix is trained using the expectation maximuzation（EM） algo- rithm in the development corpus. Then, the speaker factor is calculated through the eigenvoiee space to obtain the speaker model. Experimental results show that the algorithm can improve the system performance. In the NIST speaker recognition evaluation （SRE） 2006 10 s-10 s corpus, the equal error rate （EER） of the proposed system can be reduced by 18% against the baseline GMM system.