Aas and Eikvil [1] used an 89-character and word level. This results are present the CEDAR corpus. There present a languages. Notice that had between speaker. Using horizontal position either at the corpus. A language-independence of off-line printed sources that are specifically excluded feature vectors are ready had an average error rate for Arabic was about three times that are simple and tested on the script. Then, as in Elms and Illingworth [9], who also extraction of recognition was trained the component covariances, these basic features that we have chosen subjectively best autoresponder to represented in printing recognition: the major components used in our work in using stage in our speech we first collected as the one used 16,000 character error rates (CER) reported here was no degradation due to models, as well as features the same newly adapt all the Chinese To further through adaptation process.