采用支持向量回归机的说话者确认系统

Support Vector Regression Machine Adopted Speaker Verification System

提出一种基于支持向量回归机的说话者确认方法.该方法利用高斯混合模型中的均值向量连接构成一个超向量来模拟目标说话者的身份特性.以该超向量作为分类样本,利用支持向量回归机的方法进行分类,从而在一定程度上减轻了信道因素对系统识别精度的影响.该方法在NIST2006年说话者识别数据库上实验得到的识别等错误率比采用支持向量分类机方法有了相对12.8%的降低.

A speaker verification system based on support vector regression machine （SVR） is presented in this paper. In this paper, we model characteristics of the target speaker by integrating all the mean vectors of Gaussian mixture model （GMM） into a supervector. And then these supervectors are taken to as observations of the Support Vector Regression Machine for classification. This action makes the verification system more robust against outliers or noisy vectors and alleviates the variability of channel affects. Experiments show that the proposed SVR approach outperforms the Support Vector Classification Method at relative reduction of up to 12.8% in equal error ratio （EER） on the NIST 2006 speaker recognition corpus.