In this paper, we propose a model to generate fundamental frequency (F0) contours using neural networks. A learning procedure is proposed as an alternative to synthesis-by-rules. The generation of correct fundamental frequency contour is one of the important issues in the naturalness of automatic text-to-speech conversion systems. The proposed approach is based on a standard feed-forward multi-layer network that produces global F0 contours of sentences, directly from encoded linguistic features of standard Arabic language. Our model does not need syntactic information to produce suitable declarative intonation. TD-PSOLA synthesizer is used for validation of our results.