In smart home based on speech recognition, the corpus used for training is not complete and the application scenario is complex. Besides, the false acceptance rate of natural language speech recognition is much higher than that of small vocabulary speech recognition. During the procedure of designing and trying to implement smart home system based on natural language speech recognition, the author makes an intensive study of the MAP, MLLR algorithm based on the role of HMM acoustic model parameters. This paper presents a comprehensive adaptive method, based on which the author completed the system by using open source tools CMU SPHIN. The experiment result shows that the presented new adaptive algorithm is feasible and effective, and makes the system performance better in different scenarios.