Blind deconvolution using temporal predictability

A measure of temporal predictability is defined, and used for blind deconvolution of sound signals. The method is based on the observation that physical environments act as smoothing filters, and therefore increase the predictability of signals. These smoothing effects can be reversed by a deconvolution filter which minimises a measure of temporal predictability. This filter is obtained as the closed form solution to an eigenvalue problem which scales as O(N-3), where N is the number of filter coefficients. It is proven that the method minimises mutual information in a gaussian channel with feedback. Results are presented for sound signals. (C) 2002 Elsevier Science B.V. All rights reserved.