Blind source separation of convolutive audio using an adaptive stereo basis
Abstract
We consider the problem of convolutive blind source separation of audio mixtures. We propose an Adaptive Stereo Basis (ASB) method based on learning a set of basis vectors pairs fromthe time-domain stereomixtures. The basis vector pairs are clustered using estimated directions of arrival (DOAs) such that each basis vector pair is associated with one source. The ASB method is compared with the DUET algorithm on convolutive speech mixtures at different reverberation times and noise levels.
Origin : Publisher files allowed on an open archive
Loading...