Abstract : The aim of this thesis is to build a sentence regognition system based on an existing word regognition system. Two research axes are considered: the sentence segmentation int words as well as the integration of linguistic knowledge to take into account the context of the sentences. We studied several types of statistic language models by comparing their respective impact on the recognition system performances. We also tried to find the best strategy to introduce them efficiently into the whole recognition system. One of the originality of this study is the integration of a representation of the different sentence hypotheses in the form of a confusion network; which is then used to detect and correct the remaining regognition errors. Using the aforementioned techniques allowed us to considerably reduce the number of recognition errors among the words of the sentences.