Third International Conference on Spoken Language Processing (ICSLP 94)
In the conventional hidden Markov model (HMM), the prob- ability of duration of a state decreases exponentially with time. It is not appropriate for representing the temporal structure of speech. To overcome this problem, the use of HMMs with duration models or time-dependent transition probabilities has been proposed . These models accomplish the task with a large increase in the computation complexity. In this paper we present the Viterbi best-first searching algorithm using duration-controlled HMMs. To set a heuristic score appropriately, how the constraint is imposed on HMMs used in the backward Viterbi is investigated. The new searching algorithm is evaluated on isolated word recognition experiments. Experimental results show that the conventional Viterbi search with duration control takes 280-290 times of computation cost necessary for that without duration control, while the new searching algorithm takes only 10~15% of the computation cost keeping the same recognition rate.
Bibliographic reference. Katoh, Masaharu / Kohda, Masaki (1994): "A study on viterbi best-first search for isolated word recognition using duration-controlled HMM", In ICSLP-1994, 263-266.