Finite-State Language Modeling

Thursday, November 2, 2000 - 2:00pm - 3:00pm
Keller 3-180
Mehryar Mohri (AT&T Laboratories - Research)
Weighted automata and transducers are successfully used in many text and speech processing applications. They give a unifying framework for the construction and representation of the components of speech recognition, speech synthesis and spoken-dialogue systems. This representation provides opportunities for the application of general and well-studied algorithms. We give a brief introduction to the theory of weighted transducers and describe some of these algorithms.

The automata used in language modeling for speech synthesis, speech understanding and speech recognition represent regular languages. Recent work in the theory of weighted automata shows that the same weighted automata as those currently used in speech processing can be used to recognize efficiently non-trivial classes of context-free languages. We present some results of this work, and illustrate the efficient recognition of some well-known context-free languages.