In this article, we show how the performances of a dictation machine can be improved following the integration of a semantic analyser, which currently is dedicated to the specific French prepositional phrases of the form [NP de NP]. Besides, the dictation machine works on the basis of a bi- and tri-class model taking essentially into account the syntactic phenomena occuring on the corpus sentences and yielding - because of the uncertainty bound to speech recognition - a set of syntactically correct sentences, among which some cannot be associated with any meaning.The semantic inference engine thus provides a selection among the different occuring 'NdeN' sentences (the number of occurences of which is high in French), in order to set aside those which do not bear any meaning and therefore to lower the number of possible solutions. This is to make a tool such as the dictation machine more efficient and thus more user-friendly.