AN HMM STATE DURATION CONTROL ALGORITHM APPLIED TO LARGE-VOCABULARY SPONTANEOUS SPEECH RECOGNITION

Although Hidden Markov Modeling (HMM) is widely acid successfully used in many speech recognition applications, duration control for HMMs is still an important issue in improving recognition accuracy since a HM M places no constraints on duration. For compensating this defect, som e duration control algorithms that employ precise duration models have been proposed. However, they suffer from greatly increased computatio nal complexity. This paper proposes a new state duration control algor ithm for limiting both the maximum and the minimum state durations. Th e algorithm is for the HMM trellis likelihood calculation, not for the Viterbi calculation. The amount of computation required by this algor ithm is only order one (O(1)) for the maximum state duration n; that i s, the computation amount is independent of the maximum state duration while many conventional duration control algorithm require computatio n in the amount of order n or order n(2). Thus, the algorithm can dras tically reduce the computation needed for duration control. The algori thm uses the property that the trellis likelihood calculation is a sum mation of many path likelihoods. At each frame, the path likelihood th at exceeds the maximum likelihood is subtracted, and the path likeliho od that satisfies the minimum likelihood is added to the forward proba bility. By iterating this procedure, the algorithm calculates the trel lis likelihood efficiently. The algorithm was evaluated using a large- vocabulary speaker-independent spontaneous speech recognition system f or telephone directory assistance. The average reduction in error rate for sentence understanding was about 7% when using context-independen t HMMs, and 3% when using context-dependent HMMs. We could confirm the improvement by using the proposed state duration control algorithm ev en though the maximum and the minimum state durations were not optimiz ed for the task (speaker-independent duration settings obtained from a different task were used).