SPEECH AND

This paper describes the theory and implementation of Bayesian networks in the context of automatic speech recognition. Bayesian networks provide a succinct and expressive graphical language for factoring joint probability distributions, and we begin by presenting the structures that are appropriate for doing speech recognition training and decoding. This approach is notable because it expresses all the details of a speech recognition system in a uniform way using only the concepts of random variables and conditional probabilities. A powerful set of computational routines complements the representational utility of Bayesian networks, and the second part of this paper describes these algorithms in detail. We present a novel view of inference in general networks – where inference is done via a change-of-variables that renders the network tree-structured and amenable to a very simple form of inference. We present the technique in terms of straightforward dynamic programming recursions analogous to HMM a–b computation, and then extend it to handle deterministic constraints amongst variables in an extremely efficient manner. The paper concludes with a sequence of experimental results that show the range of effects that can be modeled, and that significant reductions in error-rate can be expected from intelligently factored state representations