最大后验概率自适应方法在口令识别中的应用

Maximum A Posteriori adaptation in password recognition

自适应技术是提高非特定人语音识别系统识别性能的有效手段,其中应用最广泛的两种自适应方法是基于最大后验概率的自适应方法和基于最大似然线性回归的自适应方法,分析了它们各自的特点并将最大后验概率的自适应方法应用到基于隐马尔可夫模型的口令识别系统中,实验结果表明,该方法能够在每个词自适应一次的情况下,使系统的识别率由40%提高到90%以上,并在此基础上实现了一个实用的中等词汇量的口令识别系统。

Adaptation technologies are the effective means of improving the performance of the speaker independent speech recog- nition system. Two of the most widely used adaptation methods are Maximum Likelihood Linear Regression（MLLR） and Maxi- mum A Posteriori （MAP）. This paper analyzes their features and through experiments on MAP probability in the parameter and the data selection, how to improve the effectiveness of the speaker adaptation is discussed. Experimental results show that the method is an efficient adaptation approach in this system. The system accuracy rate is increased from 40% to more than 90% by on- ly a few data. On this basis, it realizes a practical medium vocabulary password recognition system.