In a recent publication [1], we have introduced a neural predictive system for on-line word recognition. Our approach implements a Hidden Markov Model (HMM)-based cooperation of several predictive neural networks. The task of the HMM is to guide the training procedure of neural networks on successive parts of a word. Each word is modeled by the concatenation of letter-models corresponding to the letters composing it. Successive parts of a word are this way modeled by different neural networks. A dynamical segmentation allows to adjust letter-models to the great variability of handwriting encountered in the words. Our system combines Multilayer Neural Networks and Dynamic Programming with an underlying Left-Right Hidden Markov Model (HMM). In this paper, we present an extension of this model to off-line word recognition. We use on-line data in these off-line experiments, generating a binary image from trajectory data. The feature extraction module then turns each binary image into a sequence of feature vectors, called ‘frames’, combining low-level and high-level features in a new feature extraction paradigm. Some results for word recognition are presented.