A METHOD FOR SIGNAL EXTRACTION FROM NOISE-ADDED SIGNALS

This paper proposes a new method for signal extraction from noise-adde d signals, addressing the problem of segregating two acoustic sources as a model for acoustic source segregation. This method models some co nstraints of auditory scene analysis (ASA) and makes it possible to se gregate two acoustic sources using the amplitude envelope and the phas e deviation (input and output phase) obtained from the output of a wav elet filterbank. Using these three physical clues, the amplitude envel ope and the output phase are determined; then, the input phase is dete rmined by using the physical constraints translated from heuristic reg ularities, changes in an acoustic event and gradualness of change, as proposed by Bregman. As an example of segregation using the proposed m ethod, we seek to provide a solution for the problem of segregating tw o acoustic sources in which a sinusoidal signal is added to a bandpass ed noise. In particular, if the parameters of the proposed model are s et to human auditory properties, it can serve as a computational model for co-modulation masking release, which makes extraction of sinusoid al signals when such signals are added to AM bandpassed noise simpler while making the extraction of sinusoidal signals difficult when such signals are added to bandpassed random noise. (C) 1997 Scripta Technic a, Inc.