Lay summary
Regulus is an Open Source platform used for building linguistically motivated, grammar-based systems for speech translation and spoken dialogue, which has been developed at Geneva University, NASA Ames Research Center and elsewhere over the last five years. The basic idea of the system is to encode most of the domain-independent information needed to construct a speech recogniser inside a library of substantial, linguistically motivated domain-independent unification grammars. Efficient domain-specific recognisers are then constructed by first writing a domain-specific lexicon and training corpus, and then automatically deriving the recogniser using a set of corpus-driven transformations. The platform has been used successfully to build several large applications, most notably the FNRS-funded MedSLT medical speech translation at Geneva and the Clarissa procedure browser at NASA. The Regulus project has a good international reputation, with numerous publications, including a recent book from CSLI Press, and includes resource grammars for English, French, Spanish, Catalan, Japanese and Arabic.

The project described here proposes to extend and stabilise the Regulus platform so as to make it suitable for adoption as a general Swiss platform for spoken language system construction in academia and industry. The key goals are adding resource grammars for new languages including German and Italian; building tools to simplify the process of adding lexicon entries and tuning grammars to specific domains; extending the platform to add more support for construction of spoken dialogue applications, as opposed to speech translation systems; adding support to allow Regulus to be run on mobile platforms; and improving the quality of recognition by making it easier to incorporate domain and discourse constraints into the speech understanding process. All functionality will be made available through a well-documented Java-based GUI.