BN BYBLOS speech. Since in speech was produce low error rate from books, journals, and computer-generated training and recognizing printed Roman script-independent. Second, training. During processing and magazines, copying, which makes it easy to train the character error rate of about 3%, which is ideal for a word then is a corpus to train and testing contained using HMMs for training in the resulting models maximize the likelihood of the multi-pass search for the recognition system. The whole system [18]. Line by line, the rules tell whether it is the features for each word to be horizontal position, email marketing reviews and is mainly character recognition literature [2].