Modeling And Algorithm Of Hidden Markov Models Based On State Duration

Hidden Markov models ( HMM) are statistic models. Each model consists of a Hidden Markov process and an observation process. There are three problems needed to be solved by using Hidden Markov models, which are training, decoding and recognizing. The answers for those three problems consist of the theory of Hidden Markov models. Parameter estimation is the core problem of the training process.Presently, in HMM applied in speech recognition, some error often generate in the process because the characters of the models are not quite accurate. Especially in speaker recognition, as the HMM presented are not very accurate, the true speaker maybe can not be recognized. Those recognition errors will bring great loss in many actual projects such as speech password guard. So, further accuracy of HMM become an important precondition for correct recognition.Aiming at the two problems in speech recognition, which are ,(1) The connection between state duration and observed character vector; (2) the dependent problem of frames in speech recognition. Two kinds of new models MDHMM(modified discrete HMM)and MCHMM(modified continuous HMM)are set up in this paper, in which state duration is considered in both transition probability and character probability and new model parameters are made, not only do the two new models make use of the important information "state duration", but also do speech recognition with observed character vector which output by state transition-arc added state duration. This considers the frame dependency effectively, and reduces the complexity in operation.A modified Viterbi algorithm is applied to recognition in this paper. The parameter re-estimation algorithms of MDHMM and MCHMM are also discussed, and relatively simulations of the two models are made, in theories and simulations which present the accuracy and superiority of those HMMs.