We present a new approach for language modeling based on dynamic Bayesian networks. The philosophy behind this architecture is to learn from data the appropriate relations of dependency between the linguistic variables used in language modeling process. It is an original and coherent framework that processes words and classes in the same model. This approach leads to new data-driven language models capable of outperforming classical ones, sometimes with lower computational complexity. We present experiments on a small and medium corpora. The results show that this new technique is very promising and deserves further investigations.