We present here an oral dialogue system designed for air-traffic controller training with an air-traffic simulator. The system has been developed with two main objectives in mind: we try to make it both habitable and robust. The dialogue manager is designed to be able to cope with actual operational conditions and to allow users flexibility and freedom in particularly in the choice of their formulations and their pronunciation modes. Knowledge sources comprising the vocabulary, the phraseology, the universe of the task and the history of dialogue are represented in the form of hierarchized frames. The speech recognition system which is being used is AMADEUS developed at the LIMSI. It is a DTW-based continuous speech recognizer. The language model used is represented by binary rules with frequential weights obtained from automatic training on corpora. Weights are dynamically modified during the dialogue according to the universe state of the task, the dialogue history and even to the user's message. Those weights make some messages allowed or forbidden.