This study describes a hidden Markov model using a grapheme neural networks approach designed to recognize off-line unconstrained Arabic handwritten words. After pre-processing, a word image is segmented into characters or pseudo-characters called graphemes and represented by a sequence of observations. Each observation consists of a set of global and local features that reflect the geometrical and topological properties of a grapheme accompanied with information concerning its affiliation to one of five predefined groups. Within its group, the classification of a grapheme is done by a neural network trained with fuzzy class memberships rather than crisp class memberships as desired outputs because it results in more useful grapheme recognition modules for handwritten word recognition. The experimental results on a test database are presented to demonstrate the reliability of this study.