An HMM for phonetic transcription is presented. The inter-state transitions are bounded around phone boundaries, which are estimated from the observation sequence by statistical phone boundary detectors. The detection is done using the ratio of two probabilities, a probability that the observation sequence in a window has a phone boundary and a probability that the observation sequence in a window does not have a phone boundary. Finding an optimal state sequence is done by a simple Viterbi algorithm with 2 variables(time and state). In phonetic transcription experiments the presented HMM achieved the best accuracy against HMMs with explicit modeling of state durations. The great improvement in total performance was due to reduction of insertion errors.

Keywords: Phonetic Transcription, Hidden Markov Model, Bounded Transition, Statistical Phone Boundary Detection, Ergodic HMM