Helmut Lucke
On the Applicability of
Bayesian Belief Networks to
Language Modeling in Speech
Recognition
Abstract:Baysian Belief Networks are a powerful tool for combining different knowledge sources with various degrees of uncertainty in a mathematical sound and computationally efficient way. Surprisingly they have not yet found their way into the speech processing field, despite the fact that it in this science multiple unreliable information sources exists. The present paper provides an introduction to the theory of Bayesian Networks. It also proposes several extensions to the classic theory as described by Pearl by describing mechanisms for dealing with statistical dependence among daughter nodes (usually assumed to be marginally independent) and by providing a learning algorithm based on the EM-algorithm with which the probabilities of link matrices can be learned from example data. Using these ideas a possible language model for speech recognition is constructed. It is evaluated over a text data base.