Sparseness‐based 2ch blind source separation in reverberant environments based on direction‐of‐arrival estimation with a reliability measure

Abstract

We present a novel sparseness‐based 2ch blind source separation (BSS) method based on robust direction‐of‐arrival (DOA) estimation that enables reasonable separation of source signals even in a highly reverberant environment. DOA estimation for BSS is not easy in reverberant environments since reflected sound waves behave as coherent interferences in the observed signals and make ambiguous the sound source direction. Our key idea to solve this problem is to find and integrate clean, reliable time‐frequency fragments through the observation with a reliability measure. For that, (1) based on a diffused sound field model of reverberation, a theoretical correlation matrix is introduced for whitening the coherent interferences, (2) the DOA of each time‐frequency bin is estimated with reliability by utilizing information of neighboring bins, and (3) the reliability‐weighted DOAs are integrated over the whole time‐frequency bins. After that, separated signals are obtained by time‐frequency masking or inverse filtering. Experimental results including comparison with other methods are also reported.