It is a great contribution that introduces hidden Markov model(HMM)in speech recognition.But the assumption of the traditional HMM that state transition probabilities a

ij

(#em/em#

j

)are independent of self-trainsiton probabilty a

ij

leads to disharmomy to the model.In fact

duration distribution probability and transition probability are dependent and one can be determined by another. From the viewpoint of duration distribution probabilty

this paper shows that the aussumption of independence of transition probability is unreasonable.And a new re-estimation algorithm of transition probability is deduced.The new algorithm performs better than the Baum-Welch algorithm.The average error reduction is about 5%.