Continuous speech recognition using structural learning of dynamic Bayesian networks

We present a new continuous automatic speech recognition system where no a priori assumptions on the dependencies between the observed and the hidden speech processes are made. Rather, dependencies are learned form data using the Bayesian networks formalism. This approach guaranties to improve modelling fidelity as compared to HMMs. Furthermore, our approach is technically very attractive because all the computational effort is made in the training phase.

1 Introduction

First order Hidden Markov Models (HMM) are the most commonly used stochastic models in speech recognition. They are defined with a set of imposed conditional independence assumptions. Indeed, the observations are assumed to be governed by a hidden (unobserved) dynamic process. The associated independence assumptions state that the hidden process is first-order Markov, and each observation depends only on the current hidden variable. There is, however, a fundamental question regarding these dependency assumptions: are they realistic hypothesis for any kind of speech recognition task?