Abstract: A new Markov model,i.e.,asynchronous HMM(hidden Markov model) is proposed.By adding a new hidden time-stamp variable Ck,the presented model can be applied to the situations in which some frames are lost in the process of speech recognition to estimate the state sequence corresponding with the actual observation value,and can model the irregularly or incompletely sampled data.The forwards/backwards algorithm as well as EM(Expectation-Maximization) training algorithm which are suitable for asynchronous HMM(AHMM) are particularly introduced,and the calculation of transition matrix is optimized.At last,traditional HMM and AHMM models are used in experiment and simulation to recognize the same speech data of which the frame blocks are randomly extracted,and the results show that the recognition error rate of AHMM is lower than that of traditional HMM in condition of extracting the same amount of frame block.