This paper describes a system for off-line recognition of handwritten Arabic words. It uses simple and easily extractable features to construct feature vectors for words in the vocabulary. Some of these features are statistical, based on pixel distributions and local pixel configurations. Others are structural, based on the presence of ascenders, descenders and diacritic points. The system is evolved based on vertical and horizontal Hidden Markov Models and Dynamic Bayesian Network. Our strategy consists of looking for various architectures and selecting those which provide the best recognition performance. Experiments on handwritten Arabic words from IFN/ENIT database and ancient manuscripts strongly support the feasibility of the proposed system. The recognition rates achieve 91.89% (IFN/ENIT) and 94.61% (ancient manuscripts).