This article analyses the behavior of various hybrid architectures based on a multi-state neuro-markovian scheme (MS-TDNN HMM) applied to online handwriting word recognition systems. We have considered different cost functions, including maximal mutual information criteria with discriminant training and maximum likelihood estimation, to train the systems globally at the word level and also we varied the number of states from one up to three model basic hidden markov models at the letter levekl. We report experimental results for non constrained, write independent, word recognition obtained on the IRONOFF database.