最大后验估计和最近邻线性回归结合的说话人自适应方法  

A Novel speaker Adaptation Method based on Map and NNLR

作　　者：何磊[1] 武健[1] 方棣棠[1] 吴文虎[1] 

机构地区：[1]清华大学计算机科学与技术系,智能技术与系统国家重点实验语音技术中心,北京100084

出　　处：《电子学报》2000年第11期55-58,共4页Acta Electronica Sinica

摘　　要：本文提出一种新的说话人自适应方法 :最大后验 (MAP)估计与最近邻线性回归 (NNLR)结合的自适应 ,利用模型近邻信息和MAP自适应结果 ,建立线性回归模型 ,对没有自适应数据的模型完成模型调整 .实验证明 ,NNLR要优于另一种用于MAP自适应框架的模型插值方法 :向量域平滑 (VFS) .This paper describes a novel speaker adaptation method that combines maximum a posteriori(MAP)estimation and nearest neighbor linear regression(NNLR).In this scheme,the relationships between speaker independent models and speaker adaptation models are trained by applying the linear regression to neighbor parameters with and without MAP adaptation.Experiments show that the less adaptation data are repuired in MAP/NNLR adaptation with convergence to SD model held.In addition,experiments prove that NNLR is more efficient than vector field smoothing,(VFS)which is another model interpolation technique used in MAP adaptation frame work.

关 键 词：说话人自适应 最大后验估计 最近邻线性回归 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]