最大后验估计和加权近邻回归结合的说话人自适应方法  

Speaker adaptation with MAP estimation and weighted neighbor regression

作　　者：何磊[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

出　　处：《清华大学学报（自然科学版）》2001年第1期60-63,共4页Journal of Tsinghua University(Science and Technology)

摘　　要：提出了一种最大后验 (m aximum a posteriori,MAP)估计和加权近邻回归 (weighted neighbors regression,WNR)相结合的说话人自适应方法。在 MAP自适应中 ,只有自适应数据对应的模型参数可以得到调整。针对这一缺点 ,提出一种基于变换的模型插值 /平滑方法 - WNR,利用模型近邻信息和 MAP自适应结果 ,建立距离加权的回归模型 ,对没有自适应数据的模型完成模型调整。实验证明 ,该方法可以有效地提高 MAP自适应的速度。在自适应数据为 10句时 ,音节误识率降低近 15 % ;而在自适应数据为 2 5 0句时 ,误识率降低 5 0 %以上。此外 ,证明了向量域平滑 (vectorfield sm oothing,VFS)是This paper describes a novel speaker adaptation framework that combines the maximum a posteriori (MAP) estimation and wighted neighbor regression (WNR) methods. A great deal of adaptation data is required in MAP adaptation because only the parameters of those models with adaptation data can be updated. To alleviate this disadvantage, a technique called WNR is presented in which the parameter relationships between the speaker independent models and the speaker adaptation models are trained by applying distance weighted regression to a set of neighbor model parameters with and without MAP adaptation. The Chinese syllable recognition error is reduced nearly 15 percent with 10 adaptation utterances and more than 50 percent with 250 utterances. In addition, vector field smoothing (VFS) can be proved to be a degenerate case of WNR.

关 键 词：说话人自适应 最大后验 向量域平滑 语音识别 最大后验估计 加权近邻回归 

分 类 号：TN912.3[电子电信—通信与信息系统]