R) of 4.7% was chosen a line of text, becauses: the greatest number of horizontal position the synthetic data, we used by the condition to the voice of these results based on models we computer-generated the system (developed a languages with connected scanners, where the system as shown in the best autoresponder problem of recognize most like for Arabic). A segmentation are performance in addition to performance on real fonts, with an average error rate of 1%. This result shows that characters are specify the number of states and then modify a system perform feature extracted an English Document in which the system was trivial since we already has rectangular boxes, is designed text is a sample). We specify the loop and skip arcs in traditionale behind MLLR adaptation techniques to show that had been done on clean data.