基于多微商核函数的SVM话者确认  

Multiple Derivative Kernel for SVM Based Speaker Verification

作　　者：许敏强[1] 戴蓓蒨[1] 刘青松[1] 许东星[1] 

机构地区：[1]中国科学技术大学电子科学与技术系,合肥230027

出　　处：《数据采集与处理》2011年第5期508-514,共7页Journal of Data Acquisition and Processing

摘　　要：给出了一种基于多微商核函数(MDK)的结合高斯混合模型(GMM)和支持向量机(SVM)的方法,并应用于SVM文本无关话者确认。从GMM话者语音特征概率分布出发,用多阶微商描述GMM概率分布,将GMM和SVM结合的问题转化为用多阶微商建立SVM话者模型的问题。首先对说话人语音进行基于因子分析的参数域失配补偿,用GMM描述失配补偿后的话者语音特征的概率分布;然后对GMM求多阶微商;最后构建多微商核函数,建立多SVM话者模型。在NIST’01 2min-1min话者确认数据库上的实验表明,基于多微商核函数的SVM话者确认系统性能优于基于失配补偿的GMM系统,也比基于失配补偿的Fisher核函数SVM话者系统和基于失配补偿的Kullback-Leibler(KL)距离SVM话者系统有较大的提高。A multiple derivative kernel （MDK） based method is proposed, combining Gaussian mixture model （GMM） and support vector machine （SVM）, and it is applied to text-independent speaker verification. In order to combine GMM and SVM, MDK computes multiple derivatives from speaker feature distribution, which is modeled by GMM. Then, the multiple derivatives are taken as the input of SVM. The framework of the multiple derivative kernel based SVM method （MDK-SVM） for speaker verification is as follows. Firstly, features are abstracted from utterances and are compensated using factor analysis method in the feature domain. Secondly, these features are used for training GMM distribution. Thirdly, multiple derivative kernel is computed from the GMM distribution, and used as the input of the SVMs for speaker modeling. Finally, the performance of MDK-SVM is evaluated on the NIST SRE 01 2min-1min dataset. The proposed MDK-SVM system gives reduction in equal error rate （EER） and minimum detection cost function （MinDCF） compared with factor analysis Gaussian mixture model （FAGMM） system, Fisher kernel SVM system and Kullback-Leibler divergence based SVM system.

关 键 词：话者确认 支持向量机 多微商核函数 特征概率分布 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]