Full-Text Search:
Home|Journal Papers|About CNKI|User Service|FAQ|Contact Us|中文
《Journal of Data Acquisition and Processing》 2011-05
Add to Favorite Get Latest Update

Multiple Derivative Kernel for SVM Based Speaker Verification

Xu Minqiang,Dai Beiqian,Liu Qingsong,Xu Dongxing(Department of Electronic Science and Technology,University of Science and Technology of China,Hefei,230027,China)  
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 MDKSVM 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.
【CateGory Index】: TN912.34
Download(CAJ format) Download(PDF format)
CAJViewer7.0 supports all the CNKI file formats; AdobeReader only supports the PDF format.
©2006 Tsinghua Tongfang Knowledge Network Technology Co., Ltd.(Beijing)(TTKN) All rights reserved