We describe a method based on two independent and complementary predictions mechanisms, which is used to improve the recognition and the understanding performances of the SUNDIAL1 speech and dialogue system. We exploit information produced by the dialogue manager component of the SUNDIAL system which consists of intentional contents (list of dialogue acts) and proposition al contents (task types associated to a list of domain related semantic types). The static predictions mechanism corresponds to the determination of characteristic states of the dialogue related to the resolution of the task. The dialogue state reference is then transmitted to the recognition module, which activates a specific sub-lexicon and word-pair grammar. The dynamic predictions mechanism is based on two trials (semantic and dialogic), which are applied to the candidate strings produced by the parser. Tests show a significant definite improvement of the sentence understanding rate of the dialogue system with both kinds of mechanism.