【作者】 王晶莹； 王作英； 欧智坚；

【Author】 Wang Jingying, Wang Zuoying, Ou Zhijian Speech Recognition Lab, Electronics Engineering Department of Tsinghua University,100084

【机构】 清华大学电子工程系语音识别实验室；

【摘要】 本文提出了一种改进的最大似然线性回归说话人自适应方法一线性预测最大似然线性回归说话人自适应方法。由于LPHMM相比于经典的HMM充分利用了语音信号的时间相关性信息,能更充分地描述语音特性,因此将MLLR应用在LPHMM 中能得到更好的说话人自适应效果。自适应过程可以分为两个阶段:第一阶段,对特征去相关;第二阶段,对均值自适应。作者通过大词汇量连续语音识别实验果证明:这种做法可以明显降低语音识别系统的误识率。更多还原

【Abstract】 In this paper, author proposed using MLLR (Maximum Likelihood Linear Regression) in LPHMM (Linear Prediction Hidden Markov Model). Since LPHMM made fully use of temporal dependence information and it could better describe characteristics of speech signal than classical HMM, the use of MLLR in LPHMM can achieve better adaptation result. The adaptation process can be divided into two stages. In the first stage, the feature is decorrelated. In the second stage, mean is adapted. Experimental results show that system performance can be improved by doing so.更多还原