Abstract: The existing computer-aided system uses the algorithm of HMM based log posterior probability to judge the tester's pronunciation, but the confusion between HMM models is big. In order to improve the validity and reliability of the system, the author reconstructs the recognition network in algorithm based on the introduction of linguistic knowledge of Putonghua pronunciation, and optimizes the probability spaces in algorithm. Experimental results indicate that the improved recognition networks can not only significantly reduce the system's operation time, but also effectively reduce the probability space impact on scoring, and improve the system of evaluating performance.