Building a dictation machine involves the management of a large amount of linguistic knowledge. We are dealing here with the explicit integration of a phonological module in the automatic dictation machine, MAUD, of which a first version has already been implemented. MAUD is made up of three main modules: an acoustic-phonetic decoder, a lexical module and a syntactic-semantic module. Some of the problems at the lexical level can be addressed with the help of a phonological module. The one we have developed comprises a set of phonological rules taken from generative phonology in order to transform phonetic forms into phonological ones. The most important problem is to formalize the rules in order to use them in speech recognition. The implementation of these rules leans on IRIT work based on two original notions: multi-pronunciation groups (mpg's) and contextual phonological groups (cpg's). Experimental results illustrate the impact of phonological knowledge in the overall recognition process of a dictation machine.

Keywords: Dictation Machine, Continuous Speech Recognition, Multi-agents Architecture, Phonological Knowledge