Non-recurrent neural networks for auditory perceptual modeling

Abstract

In this paper the auditory masking phenomenon is modeled by means of feed-forward neural networks (NNs). It is shown that, because of their regularity and distributed structure, NNs are suitable for efficient implementations. A mathematical model of auditory masking, fitted for a perceptual coding algorithm, was used to generate the training data. Several network topologies were tested in search of a NN able to adequately represent the data space and with the minimum computational cost. Comparisons between the outputs of both the model and the selected NN are made, showing encouraging results for further explorations.