A frequency domain blind deconvolution algorithm in acoustics

Abstract

It is common in acoustics to measure a signal that has been corrupted by an unknown filtering function during propagation from an unknown source. Blind deconvolution is a technique for learning and applying the inverse of the unknown channel impulse response in order to recover the original source signal. One approach to accomplishing this task is based on an adaptive nonlinear algorithm using mutual information as a cost function [A. J. Bell and T. J. Sejnowski, Neural Comput. 7, 1129–1159 (1995)]. A new frequency domain implementation of this algorithm is presented which greatly reduces computational cost. The frequency domain approach allows adaptive learning rates to be applied individually to each frequency bin of the inverse filter. This technique can lead to improved convergence times for filters with a large spread of frequency response magnitudes. Preliminary results suggest that a factor of two reduction in convergence time and a factor of ten reduction in computational cost can be attained. Experimental results for several simple acoustical systems are presented comparing the performance of the pre-existing time domain algorithm and the new frequency domain implementation. [Work supported by Dr. David Drumheller, ONR Code 333, Contract No. N00014-00-G-0058.]