We first recall the importance of the prediction of hypotheses in man-machine command dialogue systems in order to improve the efficiency and the reliability of the recognition and to discuss the different possible sources of knowledge used to guide such predictions. We then show the interest of chiefly basing this process on pragmatics (which is defined here as the context of an application at a given moment) rather than on semantic information only. Next, we propose to explain the linguistic models (syntax, semantic, lexicon, reference and context) which have been chosen for this study. Finally, we present, in terms of general architecture, the data structures and the process implemented as well as an establishment of these proposals in the frame of the COMPPA system for which we give a first evaluation of its results. || Après un rappel sur l'importance des processus de prédiction d'hypothèses dans les systèmes de dialogue oraux homme-machine de commande pour augmenter l'efficacité et la fiabilité de la reconnaissance, nous discutons des diverses sources de connaissances