Abstract : This paper investigates the integration of different language models into an on-line sentence recognition system. The impact of n-gram and n-class (based on statistically and on morpho-syntactically classes) models, built on the Brown corpus, is compared in terms of word recognition rate. Furthermore, their integration in different steps of the recognition process (during it or to rescore the Nbest list of proposed sentences) is considered, thus showing better performances when used the sooner. Combinations of these models are also studied, in addition to the integration in the aforementioned recognition steps. All experiments are carried out on sentences from the Brown corpus which were written by several writers.
