Hidden Markov models with duration-dependent state transition probabilities (speech recognition)

摘要

A new method is proposed for incorporation of duration knowledge in the form of duration-dependent state transition probabilities in a left-right hidden Markov model. Duration-dependent transition probabilities are derived from integration of histograms of the state durations. The model re-estimation process becomes one of obtaining a new segmentation from which a new set of state and observation probabilities are derived.