Performance improvement of filterbank ICA with sound localization

Abstract

In an attempt to improve the performance of ICA based blind source separation methods, the stationarity of the mixing environment is exploited. The similarity between the structural modality of filterbank ICA and human auditory pathway is utilized to develop a binaural system for sound localization and speech enhancement. The developed system includes a filterbank and blind signal separation is conducted on each filtered mixtures which may be decimated for reducing computational complexity. The filterbank architecture is also advantageous to estimate sound directions from noisy multi-source mixtures. The estimated time delays are collected at each subband and are utilized to localize the multiple sources from convolved mixtures. The source arrival time differences are imposed as an additional constraint on filterbankbased ICA approaches. Initialization of subband separation network according to the source directions in localized space results in faster convergence. Though initialization results in faster convergence, no improvement is observed in the final separation performance. However, directional constraint along with initialization on the separation network provides robust behavior to the learning of subband demixing filters even if the mixtures are corrupted with additive noise. Constraining the subband demixing network to the location of source signals in constant mixing environment results in faster convergence and effective permutation correction for noisy observations, moreover, final performance improvement is also observed.