Bayesian blind MIMO deconvolution of nonstationary autoregressive sources mixed through all-pole channels

Blind deconvolution is fundamental in signal processing applications and still remains a challenging problem. In particular, blind dereverberation is necessary for applications set in acoustic environments. In this setting, a temporally-correlated observed signal whose signal-value has infinite support is modelled as the convolutive mixture of unknown source signals with an unknown channel. Multi-channel blind deconvolution is tackled by extending a method that has previously been successfully applied to the single-channel scenario. To avoid any channel-source identification ambiguities, each nonstationary source is modelled by block stationary AR process, and each channel path by a stationary subband all-pole filter. Robust and accurate estimates of the channel are obtained using Bayesian techniques, and an estimate of the original signal is obtained by inverse filtering the observed convolved signal. Simulation results are included, and it is expected that further results is presented at http://www-sigproc.eng.cam.uk/jrh1008.