First International Conference on Spoken Language Processing (ICSLP 90)
This paper describes an effort to extend the theory of hidden Markov models (HMM). It is rather revealing to find out that the scaling factors are a conditional probability of observing the current symbol given all the past observed symbols. And the single symbol emission probability is independent of time if the initial state distribution is the limit distribution of the corresponding Markov chain. To make the real-time implementation of reestimation possible, we derive the forward evaluation of the backward probabilities. Binomial distribution is introduced to model the state transitions of HMM. This increases not only the flexibility of the model but also the modelling power of HMM. Simulation results are presented.
Bibliographic reference. Chang, Lu / Bayoumi, M. M. (1990): "New results on theory of hidden Markov models", In ICSLP-1990, 53-56.