In order to reduce the effects of noise mixed with a speech, a method of noise reduction is proposed. The method is composed of a combination of neural networks and vector quantization. A multi-stage neural networks is proposed. Each stage has a small number of output categories, and its performances are evaluated. The scheme is effective for noise reduction in the range of SNR better than 0 dB» and developed to be practicable by the introduction of reject and re-try process.