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.
出处
《电子学报》
EI
CAS
CSCD
北大核心
2000年第11期55-58,共4页
Acta Electronica Sinica
关键词
说话人自适应
最大后验估计
最近邻线性回归
speaker adaptation
maximum a posteriori
MAP
vector field smoothing VFS.
