Dynamic Bayesian network based speech recognition with pitch and energy as auxiliary variables

Abstra t. Pit h and energy are two fundamental features de-s ribing spee h, having importan e in human spee h re ognition. However, when in orporated as features in automati spee h re og-nition (ASR), they usually result in a signi ant degradation on re ognition performan e due to the noise inherent in estimating or modeling them. In this paper, we show experimentally how this an be orre ted by either onditioning the emission distributions upon these features or by marginalizing out these features in re og-nition. Sin e this is not obvious to do with standard hidden Markov models (HMMs), this work has been performed in the framework of dynami Bayesian networks (DBNs), resulting in more exibil-ity in dening the topology of the emission distributions and in spe ifying whether variables should be marginalized out.