基于目标驱动的多层MLLR自适应算法

Multi-Layer Structure MLLR Adaptation Algorithm Based on Target-Driven

本文在对语音识别中基于自适应回归树的极大似然线性变换 (MLLR)模型自适应算法深刻分析的基础上 ,提出了一种基于目标驱动的多层MLLR自适应 (TMLLR)算法。这种算法基于目标驱动的原则 ,引入反馈机制 ,根据目标函数似然概率的增加来动态决定MLLR变换的变换类 ,大大提高了系统的识别率。并且由于这种算法的特殊多层结构 ,减少了许多中间的冗余计算 ,算法在具有较高的自适应精度的同时还具有较快的自适应速度。在有监督自适应实验中 ,经过此算法自适应后的系统识别率比基于自适应回归树的MLLR算法自适应后系统的误识率降低了 10 % ,自适应速度也比基于自适应回归树的MLLR算法快近一倍。

In this paper, a new algorithm called Target Driven based multiple layer maximum likelihood linear regression (TMLLR) is proposed for model adaptation in speech recognition. The algorithm can be regarded as the improvement of maximum likelihood linear regression (MLLR) using the generation of regression class trees for model adaptation. Different from conventional MLLR, the regression classes of TMLLR are generated dynamically based on increment of target function and a multi layer feedback mechanism. Because of the special multi layer structure of TMLLR, some redundant computing cost can be reduced, which caused much faster adaptation speed. The target driven strategy is aimed at increasing the likelihood probability, which is same to measure of speech recognition, so a higher recognition accuracy of the system can be achieved. In comparison with the conventional MLLR using the generation of regression class tree, TMLLR achieved a further word error rate reduction by 10% and had only about half computational time consuming in supervised adaptation experiments.