This thesis presents the development of an artificial neural network system for adaptive subband noise reduction, CHN3R (Cascaded Hybrid Neural Network for Noise Reduction). In particular, the proposed system focuses on noise reduction of corrupted speech signals in realistic noisy environments. The intended end user of CHN3R is human listener, for such applications as digital hearing aids and personal communication devices. A literature survey of the myriad noise reduction/cancellation approaches briefly introduces the reader to the current state of the art. CHN 3R is developed for the frequency domain, combining the self-organizing map and radial basis network structures. Its performance is evaluated using three different measures: segmental signal-to-noise ratio, a perceptually weighted distortion measure, as well as an informal subjective listening test. It is also compared with several classical and commercial algorithms. The performance of CHN3R is found to be comparable to some commercial algorithms in terms of the objective measures, but is found to be subjectively preferable over all the other algorithms with which it is compared.