Probabilistic Modeling With Bayesian Networks For Asr

This paper describes the application of Bayesian networks to automatic speech recognition. Bayesian networks enable the construction of probabilistic models in which an arbitrary set of variables can be associated with each speech frame in order to explicitly model factors such as speaking rate or articulator positions. Once the basic inference machinery is in place, a wide variety of models can be expressed and tested. We present algorithms for inference with Bayesian networks, and provide experimental results on the PhoneBook database. These results indicate that performance improves when the observations are conditioned on an auxiliary variable modeling acoustic/articulatory context. The use of multivalued and multiple context variables further helps. 1. INTRODUCTION Hidden Markov models are the most widely used method for doing automatic speech recognition, and are based on a very simple set of concepts and assumptions: the models are expressed in terms of phonetic states and aco..