The study we present here is a basic structure for a continuous speech recognition system. Our research concentrated on the general analysis algorithm and communication between the different linguistic levels. These modules and the algoritm are driven by a supervisor which processes information provied by the modules as theorical constraints, to reduce the branching ratio. These constraints are marked (phonetic, lexical, or syntactic) but have a meaning only for the module which produces them. The separation of the information sources and the processing operations enables the system to adapt rapidly to new applications.