【作者】 苏韬； 汪俊杰； 孙甲松； 王作英；

【机构】 清华大学电子工程系；

【摘要】 本文阐述了在语音识别中如何通过对自适应数据的概率分布进行最大似然估计,在线的将几个不同主题的语言模型进行线性插值以实现主题的自适应.本文提出了一种利用梯度投影法来解决最大似然参数估计问题的语言模型自适应方法,在线更新各主题模型的插值系数,使得语言模型与当前的识别任务更加匹配.实验结果表明,这种语言模型自适应方法使相对误识率平均下降了6.4%,同时在混淆度方面也获得了一定的改善.更多还原

【Abstract】 In this paper we propose an algori thm to implement the linear interpolation of several topic language models on-line by using maximum likelihood estimation for adapted data in order to achieve topic adaptation. To overcome the mismatch between language model and recognition task, we update each topic model’ s coefficient on-line by applying the GP algorithm to solve the problem of maximum likelihood estimation. Experiments show that 6. 4% average reduction of relative word error rate can be gained over the baseline system by using this language model adaptation method. At the same time, this method also shows improvement in terms of perplexity.更多还原