基于VQ-HMM的无教师说话人自适应方法

An Unsupervised Speaker Adaptation Method Based on VQ-HMM

提出了一种新的语音识别方法 ,该方法综合了VQ ,HMM和无教师说话人自适应算法的优点 .该方法首先在每个状态通过用矢量量化误差值取代传统HMM的输出概率值来建立VQ HMM ,同时采用无教师自适应矢量量化算法 ,来改变VQ HMM的各状态的码字 ,从而实现对未知说话人的码本适应 .本文通过非特定人汉语数码 (孤立和连续数码 )识别实验 ,把新的组合方法同基于CHMM的自适应和识别方法进行了比较 ,实验结果表明该方法鲁棒性好 ,所需计算量较少 ,自适应和识别效果远优于基于CHMM的方法 .

We propose a new speech recognition method by the integration of the VQ, HMM and an unsupervised speaker adaptation algorithm, it complies a VQ distortion measure at each state instead of a discrete output probability used by a discrete HMM, and uses an adaptive VQ algorithm to alter the codewords for speaker adaptation. In this paper, the new combined method is compared with CHMM by the task of speaker independent Chinese spoken digit (isolated/connected) recognition, and the experiments illustrate that this new method is simple and robust, and has good performance.