A new method is used for character segmentation from cursive Arabic words. The method is based on statistical approach which uses Normalization and rectification, coordinate transformation and clustering to extract ligatures. The output is then filtered to extract start, overlapped and end segment errors. After applying the filter the characters are completely isolated and ready for recognition. The system, when testing the segmentation on 5 different Arabic sentences and by 20 different users, scored 98.03%, 94.82, 97.33 and 90% for normal segment, starting segment, last segment and overlapped segment.