This paper describes some aspects of the linguistic expert system in continuous speech recognition, which makes part of the DIRA (Integrated Dialogue and Automatic Recognition) project, being developed in our laboratory. This system consists in a rule compiler and an analyzer. The compiler accepts context free or context sensitive rules and transform rules. It produces a transition network (ATN) which contains the nodes corresponding to the syntactic, lexical or phonetic categories and the transition arcs, whose traversal is conditioned by rules and predicates. These conditions arise both from semantic attributes for syntactic semantic analysis, and from phonologic rules describing the phonetic lexical network. In fact, it is possible to produce one multi-layer network in which interactions between syntactic-semantic level and lexical level are strong. For example, one chooses the input level of the grammar (axiom) and the description level of the terminal vocabulary. One can define also phonologic rules constrained by syntax or syntactic constraints from lexical informations, etc...