论文部分内容阅读
本文提出一种改进的基于模型差别度量的说话人聚类(Speaker Clustering)方法,并将该说话人聚类算法结合最大似然线性回归算法(Maximum Likelihood Linear Regression,MLLR)构成整体的说话人自适应框架。将该方法应用于以音素为识别基元的汉语连续语音识别系统中,可能够提高系统的识别率,较好的满足快速性和渐进性。实验结果表明,该方法能够在仅有一句自适应数据的情况下,使系统字正识率由40.43%提高到50.86%.
In this paper, an improved speaker clustering method based on model difference metrics is proposed, and the speaker clustering algorithm is combined with Maximum Likelihood Linear Regression (MLLR) to form the overall speaker self Adapt to the framework. Applying this method to Chinese continuous speech recognition system using phonemes as the recognition primitives may improve the recognition rate of the system, and better meet the requirements of fastness and gradualness. Experimental results show that the proposed method can increase the awareness of system words from 40.43% to 50.86% with only one adaptive data.