语音识别中统计与规则结合的语言模型

LANGUAGE MODEL FOR SPEECH RECOGNITION APPLICATIONS

在分析语音识别系统中，基于规则方法和统计方法的语言模型，提出了一种对规则进行量化的合成语言模型．该模型既避免了规则方法无法适应大规模真实文本处理的缺点，同时也提高了统计模型处理远距离约束关系和语言递归现象的能力．合成语言模型使涵盖６万词条的非特定人孤立词的语音识别系统的准确率比单独使用词的ＴＲＩＧＲＡＭ模型提高了４．９％（男声）和３．５％（女声）．

In this paper a hybrid language model integrating rule base grammar and Markov language model for speech recognition applications is described.This hybrid language model not only avoids the disadvantage of rule base grammar in processing very large real text but also has a good performance in processing Chinese language recursive nature and long distance constrained relations,which has been applied to large vocabulary isolated work speech recognition.The male voice recognition accuracy is improved from 81 7% with Trigram language model to 86 6%,the female recognition accuracy from 87 7% to 91 2%.