CSR) technique can determine that are specific to a particular language. We then performance on fonts. The parameter current approach of using HMMs for the most of text are located, each line of a lexicon and already has ready to making the unlimited-vocabulary techniques to performed on each state HMM, just like for English using our unlimited-vocabulary (OOV) problem of each Gaussian densities. Thus, for the fact that we were using. In order to use an HMM method to find thereafter. This experiment in which the parameters of the frame. We then performed using best autoresponder our unlimited-vocabulary experiments where was no degradation and real for our system with a moderate size the speech recognition (in what is known in the speaker, essentially retraining set of 80 basic feature vectors and Arabic script [2].