A Causal Frequency-Domain Implementation of a Natural Gradient Multichannel Blind Deconvolution and Source Separation Algorithm

Natural gradient adaptation is useful for developing block-based adaptive solutions to convolutive blind source separation tasks, in which frequency-domain fast convolution methods can be exploited for computational simplicity. To maintain causal operation, previouslydeveloped natural gradient convolutive blind source separation algorithms employ approximations to the natural gradient that ultimately limit the separation performance s of these schemes. In this paper, we derive a novel causal frequency-domain implementation of a natural gradient algorithm for convolutive blind source separation tasks. Simulations with convolutive speech mixtures indicate that the proposed method provides robust convergence performance for relatively-short separation filters without creating preor post-echo artifacts in the extracted sources.