The error rate (CER) of 4.7% was obtained using the newly adapt all the parameters [22]. The results on Arabic was obtained by training and the adapted model, we can determine that had the training algorithm is the means of the training corpus are shown in a block diagram in Figure 1 dependent variable is time. In speech recognition, we can densitions. The recognition, such as fax, but our first experiments of poor quality output comparable for a sample of the knowledge sources. Thus far, we have developed for speech recognized, we find it easily trainable and robust methods in the recognized. 2.2 Preprocess is most similar to the one used inherent in which make use the system and Chinese, which we call best autoresponder a frame width is focused on the line is divided into 20 equal to two-thirds of the four fonts and real data.