This paper is concerned with the recognition and the restoration of periodic signal sequences by a neural network. It describes a method for noise reduction using a feed-forward neural network. We will show the abilities of the network to recognize complex and multiple sequences, to regenerate sequences with an appropriate precision and to recognize and generate sequences despite the existence of noise. Moreover, we will discuss the ability of the feed-forward network to generalize the amplitude and the frequency. This ability will be exploited to design filters as low-pass, high-pass, band-pass and I band-stop. The system has been realized with a reduced architecture trained by the fast back propagation algorithm which yields a fast convergence by using short learning sequences made of pure tones (10 ms learning sequences).