Abstract: Speech is the most natural and widespread form of human communication. That's why speech synthesis has interested researchers for decades. But, it is turned out that developing an unlimited text-to-speech system is an enormous task. The traditional methods (synthesis by rule and synthesis by concatenation of pre-recorded sounds) used for this haven't given good results. In such situation, neural networks (NNs) have the potential to give better results thanks to their property of interpolation and their capacity of generalisation. In this paper, we present a synthesis system for Arabic language. The choice of parameters which will be used to drive the NN is very important and have an effective influence on quality of produced speech. Till now, works have been done to evaluate different methods based on linear predictive coding; speech resulted was machine-like and not intelligible. We suggest to use CELP to drive the NN, which provides a high quality speech.