This thesis deals with the study, the conception, the development and the test of an online unconstrained handwriting word recognition system for an omni-writer application. The proposed system is based on a hybrid architecture including on the one hand, a neural convolutional network (TDNN and/or SDNN), and on the other hand Hidden Markov Models (HMM). The neural network has a global vision and works at the character level, while the HMM works on a more local description and allows the extension from the character level to the word level. The system was first dedicated for processing isolated characters (digits, lowercase letters, uppercase letters). This architecture has been optimized in terms of performances and size. The second part of this work concerns the extension to the word level. In this case, we have defined a global training scheme directly at the word level. It allows to insure the global convergence of the system. It relies on an objective function that combines two main criteria: one based on generative models (typically by maximum likelihood estimation) and the second one based on discriminant criteria (maximum mutual information). Several results are presented on MNIST, IRONOFF and UNIPEN databases. They show the influence of the main parameters of the system, either in terms of topologies, information sources, and training models (number of states, criteria weighting, duration).
