Statistical language model is very important to speech recognition. To a system of special topic, domain dependent language model is much better than the general model. There are two problems in traditional method. (1) The corpus of special topic is not large enough as general corpus. (2) An article is always related to more than one topic, but these phenomena have not been considered during the process of model training. In this paper, the authors try to solve these two problems. They present a new method to organize the corpus——the method based on fuzzy training subset. And the training of domain dependent models is based on these fuzzy subsets. At the same time, self organized learning has been introduced in training process to improve the models' prediction ability. It can improve the performance of models evidently.