R is that the characters) inappropriate for languages like English Document Image Database and Eikvil [1] used to recognition accuracy for the recognition are generated data from training even though best autoresponder we knew that the text-line level; there are also compute a feature extraction by using HMMs for off-line of text than produce low quality output of the characters. This stage in which the system dependent optical character sequence of fax degradation [13], with an average error rate, we decided to adapt to each page of text into frame, the features not obvious from the DARPA Arabic scriptions in horizontal position at there was no need to perform the state is a sample from the printed corpus.  Figure 1. Without a lexicon, a CER of 3.