A multiresolution audio restoration algorithm

Abstract

This paper presents a method of restoring noisy audio signals based on a multiresolution signal representation-the multiresolution Fourier transform (MFT). This is done by using a filter that relies on detecting musically significant features in a prototype signal, which is a different performance of the same musical piece as the signal to be warped. This filter's time-scale is then warped using a multiresolution warping algorithm to match the time-scale of the signal to be restored (the target). The warped filter is then applied to the target signal allowing the signal to pass and rejecting the noise. A method is described that allows the exploitation of the multiresolution nature of audio signals by combining restoration results from more than one scale. Results are presented showing how the warping algorithm works and the gain in SNR for the restoration of one musical piece with differing amounts of additive Gaussian noise present.