A new method for speech enhancement is presented. The noise removal process is performed by a Neural Network which is fed with the noisy speech signal power spectrum and produces at its output the power spectrum of the enhanced signal. To train the Neural Network, the Auditory Masking Threshold of a rough estimate of the clean signal is used and the algorithm employed is an extended version of the backpropagation. The NN's used in a novel way, since training and operation modes are combined for every short time frame of the speech signal. Preliminary results have shown that the power spectrum of the enhanced signal is very close to the power spectrum of the clean signal and better than the spectrum obtained using the Power Spectral Subtraction method.

Keywords: Speech Enhancement, Neural Network, Masking, Backpropagation.