传统NAP,投影矩阵的训练需要繁杂的参数调整和大量的标注语料,投影后不能彻底去除干扰信息且会造成一定的信息损失。为此,该文提出一种DWNAP算法,首先通过统计各语种训练语音协方差矩阵的特征值离散度,对干扰源进行量化估计,利用规整的估计值作为各语种的区分性权重参与投影矩阵的训练。汉日英三种语言的测试结果表明,相对于传统NAP提出的DWNAP有效地提高了系统识别性能,EER相对降低了7.51%。
Abstract
The conventional NAPs training method of projection matrix requires laborious parameter tuning process over the training corpus with information labels. It cannot remove all unwanted information and result in loss of desirable information. To tackle these problems, a discriminating weighted nuisance attribute projection (DWNAP) is proposed in this paper. DWNAP quantitatively estimates the source of nuisance based on the normalized scatter of the given languages eigenvalues for discriminating weighting in training of projection matrix. Experiments on Chinese, Japanese and English show the advantage of the proposed DWNAP, with a relative reduction in the equal error rate (EER) for about 7.51% compared with the traditional NAP.
Key wordslanguage recognition; mismatch compensation; discriminating weighted nuisance attribute projection(DWNAP); nuisance attribute projection(NAP)
关键词
语种识别 /
失配补偿 /
区分加权干扰属性投影 /
干扰属性投影
{{custom_keyword}} /
Key words
language recognition /
mismatch compensation /
discriminating weighted nuisance attribute projection(DWNAP) /
nuisance attribute projection(NAP)
{{custom_keyword}} /
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}
参考文献
[1] V. Wan, W. M. Campbell. Support vector machines for speaker verification and identification.[C]//Procedings of Neural Netw. Signal Proces. 2000:775-784.
[2] W. M. Campell, E. Singer, P. A. Torres-Carrasquillo, et al. Language recognition with support vector machines[C]//Proceedings of IEEE Odyssey: The Speaker and Language Recognition Workshop Toledo, Spain, 2004: 41-44.
[3] W. Campbell, D. Sturim, P. Torres-Carrasquillo, et al. A Comparison of Subspace Feature-Domain Methods for Language Recognition[C]//Proceedings Interspeech, 2008: 309-312.
[4] A. Solomonoff, W. M. Campbell. Advances in channel compensation for SVM speaker recognition[C]//Proceedings of ICASSP, 2005:629-632.
[5] A. Solomonoff, W. M. Campbell. Channel compensation for SVM speaker recognition[C]//Proceedings of Odyssey, Speaker Lang. Recognition Workshop, 2004:57-62.
[6] Alex Solomonoff, W. M. CamPbell, Ian Boardrnan. Advances in channel compensation for svm speaker recognition[C]//Procedings of ICASSP. 2005(l):629-632.
[7] R. Vogt, S. Kajarekar, S. Sdridharan. Discriminant NAP for SVM speaker recognition[C]//Proceedings of Odyssey, 2008.
[8] 雷文辉. 基于支持向量机的语种识别研究[D].中国科学技术大学硕士学位论文.2009,5.
[9] W. M. CamPbell, D. E. Sturim,D. A. Reynolds, et al. SVM based speaker verifieation using a gmm supervector kemel and nap variability[C]//Proceedings of ICASSP. 2006(1): 97-100.
[10] Torres-carrasquillo P A, Singer E, Campbell W, et al. The MITLL NIST LRE 2007 Language Recognition System[C]//Proceedings of Interspeech. 2008: 719-722.
[11] Matejka P. Phonotatic and Acoustic Language Recognition[D]. Brno: Brno University of Technology, 2008.
[12] 姜洪臣, 郑榕, 张树武, 等. 基于SDC特征和GMM-UBM模型的自动语种识别[J]. 中文信息学报, 2007,21(1):49-53.
[13] L. Lamel, L. Labiner, A. Rosenberg, et al. An Improved Endpoint Detector for Isolated Word Recognition[C]//Proceedings of IEEE Transactions on Acoustics,Speech,and Signal Processing,Aug 1981,29(4): 777-785.
[14] Douglas A.Reynolds,Thomas F.Quatieri, Roberr B.Dunn.Speaker verification using adapted Gaussian mixture models[J].Digital Signal Processing,Jan.2000,10: 19-41.
[15] C. Lin. LIBSVM: a library for support vector machines [EB/OL.]2010-12-14. http://www.csic.ntu.tw/cjlin/libsvm/index.html. 2010.
[16] PedroJ. Moreno, Purdy P. Ho, Nuno Vasconcelos. A Kullback-Leibler divergence based kernel for SVM classification in multimedia applications[J]. in Adv. in Neural Inf.Procedings Systems 16,S.Thrun,L.Saul, and B.Schlkopf,Eds. MIT Press,Cambridge,MA,2004.
{{custom_fnGroup.title_cn}}
脚注
{{custom_fn.content}}
基金
国家863计划重点资助项目(2011AA010603)
{{custom_fund}}
PDF(1601 KB)
