摘要： 在大词汇连续语音识别系统中，语言模型权值和插入代价等语音解码参数对系统的识别率有较大的影响，而在实际应用中常通过实验手动调整其值寻求最佳识别结果。为此，提出一种利用二元文法进行词图重估的方法，自动优化语音解码参数。在重估的参数空间搜索过程中采用线性搜索与模拟退火搜索相结合的方法，使优化参数具有全局最优和对初值稳定性强的优点。实验结果表明，相比凭经验设置的参数，该方法估计出的参数值能大幅降低识别词错误率，与经典的N-best优化相比，其优化速度有较大提升。

关键词: 词图重估, 语言模型权值, 解码, 插入代价

Abstract: In Large Vocabulary Continuous Speech Recognition(LVCSR) system, speech decoding parameters——Language Model(LM) weight and insertion cost can greatly affects the recognition performance. But in practice, they are usually hand-tuned through experiment to obtain best recognition performance. This paper proposes the rescoring based method that uses bi-gram LM to optimize the parameters automatically, meanwhile the method of combine line search and Simulated Annealing(SA) search in parameters search space of rescoring, which is globally optimal and is insensitive to initial value of parameter. Experimental results show that the method can dramatically reduce the word error compared with empirical parameter setting method, and gains much faster optimization speed than classical N-best optimization.

Key words: word graph rescoring, Language Model(LM) weight, decoding, insertion cost

中图分类号: 

• TP391