基于因子分析信道失配补偿的SVM话者确认方法

SVM Speaker Verification Method of Mismatch Compensation Based on Factor Analysis

针对信道失配和统计模型区分性不足而导致话者确认性能下降问题,文中提出一种将因子分析信道失配补偿与支持向量机模型相结合的文本无关话者确认方法.在SVM话者模型前端采用高斯混合模型-背景模型(GMM-UBM)方法对语音特征参数进行聚类和升维,并利用因子分析(FA)方法,对聚类获得的超矢量进行信道补偿后作为基于SVM话者确认的输入特征,从而有效解决SVM用于文本无关话者确认的大样本、升维问题,以及信道失配对性能影响问题.在NIST06数据库上实验结果表明,文中方法比未做失配补偿的GMM-UBM系统、GMM-SVM系统在等误识率上有50%以上的改善,比做了FA失配补偿的GMM-UBM系统也有15.8%的改善.

The poor performance of speaker verification system results from the channel mismatch and the lack of distinction between statistical models. A text-independent speaker verification method is proposed which combines the channel compensation based on factor analysis and the discriminative support vector machine （SVM） model. Gaussian mixture model （GMM） is used to make the speech parameter clustered and ascended, then the channel information of GMM mean super-vectors is wiped off by using factor analysis. The parameters, which are used as inputting parameters, are employed for the construction of SVM speaker verification system. The proposed method solves the problems of large samples, dimension raising and channel mismatch effectively when SVM is used for the text-independent speaker verification. Experimental results on NIST 06 male speaker corpus show that the proposed method improves system performance. Compared with the baseline system Gaussian mixture model-universal background model （GMM-UBM） , GMM-SVM without channel compensation, the system improves the equal error rate （EER） more than 50%. Compared with the system factor analysis （FA）-GMM-UBM which uses channel compensation based on factor analysis without discriminative models, it also gets the improvement of EER by 15.8%.