Sixth European Conference on Speech Communication and Technology

Most stateoftheart statistical speech recognition systems use hidden Markov models (HMM) for modeling the speech signal. However, limited by the assumption of conditional independence of observations given the state sequence, current HMM's poorly model the trajectory constraints in speech. In [1], we introduced the parallel path HMM, where each phonetic unit is represented by a parallel collection of HMM's that model the phone trajectory variability. The trajectory constraint is imposed by disallowing transitions across parallel paths. In this paper,we investigate improvements to two critical components ofthis new framework: (i) initializing the sets of trajectoriesper phone that will form the basis of the parallel collection of HMM's, and (ii) evaluating alternative parameter sharing strategies related to distributing the number of model parameters. Recognition results on Switchboard, a large vocabulary conversational speech recognition task, demonstrate 0.71.0% absolute performance improvements with the parallel path HMM in the Nbest rescoring paradigm.
Full Paper (PDF) GnuZipped Postscript
Bibliographic reference. Iyer, Rukmini / Kimball, Owen / Gish, Herbert (1999): "Modeling trajectories in the HMM framework", In EUROSPEECH'99, 479482.