Blind separation of acoustical mixtures without time-domain deconvolution or decorrelation

Abstract

In this paper, we present a novel algorithm that is ideally suited to the blind separation of acoustical signal mixtures. Our algorithm modifies the form of a natural gradient algorithm for multichannel blind deconvolution so that the temporal characteristics of the input signals are largely preserved in the extracted output signals. The algorithm is simple, requiring approximately four multiply/adds per adaptive filter coefficient. Applications of our method to speech and music signals produce separated outputs that sound more natural than those produced by multichannel blind deconvolution algorithms.