Hidden Markov models for trajectory modeling

Current state-of-the-art statistical speech recognition systems use hidden Markov models (HMM) for modeling the speech signal. However, it is well known that HMM's do not exploit the time-dependence in the speech process, since they are limited by the assumption of conditional independence of observations given the state sequence. Alternative techniques, such as segment modeling approaches, can e ectively exploit time-dependencies in the acoustic signal by discarding the observation independence assumption. However, losing the basic HMM structure is often a high computational price to pay for improved acoustic models. In this paper, we introduce the parallel path HMM that exploits the time-dependence in speech via parametric trajectory models while maintaining the HMM framework. We present preliminary results on Switchboard, a large vocabulary conversational speech recognition task, demonstrating both improved modeling and potential for improved recognition performance.