The goal of this paper is to describe an off-line segmentation-free Arabic handwritten words recognition system. This system is based on a semi-continuous 1- dimensionnal hidden Markov models (SCHMMs) with explicit state duration of different kinds (Gauss, Poisson and Gamma). First preprocessing is applied to simplify the feature extraction process, then the word image is analyzed from right-to-left, by using a sliding window approach, in order to extract from it a vectors sequence of statistical and structural features. The extracted sequence is submitted to a SCHMMs classifier based on a likelihood criterion for identifying the word using a modified version of the Viterbi algorithm. Several experiments were performed using the IFN/ENIT benchmark database, they showed, on the one hand, a considerable improvement in the recognition rate when SCHMMs with explicit state duration of either discrete or continuous distribution are used instead of standard SCHMMs (i.e. with implicit state duration), on the other hand, the Gamma distribution for the state duration, that have given the best recognition rate (90.89 % in top 1), seems more suitable for the SCHMMs based modeling of Arabic handwriting.