This paper deals with a language modeling approach that is dedicated to an on-line handwriting recognition system. Three main goals are set: i) performances, ii) versatility, and iii) resources. To achieve these goals we propose a statistical word n-class approach, which uses a learning stage to cluster words in classes and defines an estimation of the probability distribution of sequences of classes. Very large corpora from three different languages (English, French and Italian) have been used to train and test the language models. The efficiency of these models are evaluated not only from a linguistic point of view, using perplexity measurements, but also combined inside the recognition system on real ink signals corresponding to written sentences. Using a tri-class model allows a word error rate reduction ranging from to 50 to 60% according to the language.