In view of the shortage of the traditional HMM model with a short state duration, and the low accuracy of speech recognition, long training time and high training error in the case of large computation, a HMM model based on a long state duration of speech was proposed. First, the diagonal elements of the state transition matrix are all 0, the self-transition arc is removed, and a Gaussian distribution describing the duration with a parameterized function is added. Then, each frame is calculated according to the degree of correlation between frames, and the specified transition probability of each arc and the most primitive numerical probability of the visible symbol sequence output are repeatedly calculated by the re-evaluation formula until convergence, and the operation is stopped. The ratio of the difference between its probability output and its previous probability output and the probability output value is greater than the initial value set by the HMM model. Compared with the traditional HMM model experiment, the HMM model based on the duration state can reduce the number of training times and shorten the training time to a certain extent, improve the accuracy of speech recognition, and basically complete the function of the speech recognition system.