【摘要】 本文研究了在汉语语音识别中如何根据识别任务的主题相关性自动调整语言模型 ,即语言模型的主题自适应问题。提出了利用梯度投影法在最大似然估计准则下将不同主题的语言模型进行线性插值的方法。实验表明 ,该方法可以有效地提高系统的识别率和稳健性 ,特别是对于主题明确的识别任务改善尤为明显。同时 ,为了解决新系统识别速度较慢的问题 ,本文在音字转换过程中采取了多路搜索策略 ,在与基线系统识别速度相当的情况下识别率仍获得了明显改善。更多还原

【Abstract】 In this paper the problem of adapting language model automatically according to the topic-dependence of recognition task,that is language mdel topic-adaptation,is studied.A method is proposed to implement the linear interpolation of several topic language models based on the rule of maximum likelihood estimation by using Gradient Project (GP) algorithm.This method shows an effective improvement in terms of word right rate and robustness in the experiments,especially for the recognition task with definite topic.At the same time,the strategy of multi-pass is adopted in the processing of pinyin-to-character conversion in order to solve the problem that the rrecognition of the new system is slow.In comparison to the baseline system,the word right rate is increasing apparently while the efficiency is as much as that is in baseline system.更多还原