Blind Source Separation Using Incomplete De-mixing Transform with a Precise Approach for Selecting Constrained Sets of Frequencies

This paper presents a modification of the Natural Gradient algorithm for Independent Vector Analysis estimating incomplete de-mixing transform and performing its completion. Incomplete de-mixing transform is obtained when it is estimated only on subsets of most active frequencies of the sources. The transform is then completed using methods for sparse reconstruction. In previous works, the incomplete subset of frequencies was the same for all separated signals. In this paper, we propose a new approach in which the subsets are source-dependent. Experiments conducted on the CHiME-4 dataset show that the proposed approach improves the separation performance.