Abstract

In this talk, I will present an overview of Regulus, an Open Source platform that has been under development since 2001 at NASA Ames Research Center and Geneva University. Regulus contains tools and resources to support construction of grammar-based language models for medium-vocabulary spoken dialogue applications. Language models are compiled out of general linguistically motivated unification grammars using example-based methods driven by small corpora, typically of a few hundred examples. I will describe the compilation process, present results demonstrating that Regulus-derived language models are very competitive with statistical ones derived from the same small training sets, and show how the well-defined Regulus compilation process enables systematic investigation of the properties of grammar-based recognisers.

Regulus has already been used for several large projects, including NASA ’s Clarissa astronaut assistant (as far as we know, so far the only speech enabled system in space), and Geneva University’s MedSLT, a multilingual medical speech translator. I will close by giving a short demo of MedSLT.