In this paper, a method for analytic handwritten word recognition based on causal Markov random fields is described. The words models are hmms where each state corresponds to a letter; each letter is modelled by a nshp-hmm (Markov field). Global models are build dynamically, and used for recognition and learning with the baum-welch algorithm. Learning of letter and word models is made using the parameters reestimated on the generated global models. No segmentation is necessary: the system determines itself the best limits between the letters during learning. First experiments on a real base of french check amount words give encouraging results of 83.4% for recognition.