This paper reports on experiments performed on the Italian language in order to assess the efficiency of probabilistic language models with reference to a task of large-dictionary speech recognition. Two different types of models, an M-gram and an Mg-gram one, have been investigated for comparison purposes. The quality of the models trained on a corpus of 3.5 million words was measured in terms of perplexity and of the improvement achieved by integrating the language model in real speech recognition systems. Judging from this empirical measurement, the two language models exhibit equivalent preformance for Italian, although perplexity measurements would suggest otherwise.