Probabilistic Modeling with Bayesian Networks for Automatic Speech Recognition

Bayesian networks are an extremely general prob-abilistic modeling framework, and are increasingly being applied to complex real-world problems. In this paper, we describe the use of a Bayesian net-work system in large vocabulary isolated word recognition. We brie y review the algorithms and network structures used, and present results show-ing that signicant improvements in word error rate result from modeling acoustic and articu-latory context with a multivalued context vari-able. The network parameters are highly corre-lated with simply dened acoustic characteristics, and utterance clustering results in a partitioning according to both speaker gender and the presence of liquid consonants.