Post-processing for frequency-domain blind source separation in hearing aids

In this paper, we investigate post-processing for the frequency-domain blind source separation (FD-BSS) in hearing aids applications. It is known that the segregate quality of FD-BSS degrades severely in the challenging scenario of reverberant enclosures or moving source situations. A robust two-stage dynamic programming approach based on inter-frequency correlation is presented to solving the permutation ambiguity correction problem. Moreover, binary masking method and the non-stationary spectral subtraction techniques are combined to estimate and reduce the residual cross-talk components. The subtraction parameter is adaptively determined by estimated local Signal-to-Noise Ratio (SNR). An optimization scheme is proposed aiming at obtaining a tradeoff between the residual noise reduction and generated distortion suppression. Experimental results show that the post-processing procedure enhances the separated signals and the musical noise is controlled under moderate level.