Prior-Guided Source Separation with Direct Update of Back-Projected Demixing Vectors

Kukuru Koiso, Taishi Nakashima, Nobutaka Ono

Abstract: In this study, we propose a blind source separation (BSS) method that incorporates prior knowledge of demixing vectors by directly optimizing their back-projected versions. In conventional BSS frameworks, the demixing matrix is first optimized under scale ambiguity and then rescaled through backprojection, which yields the actual demixing matrix used for separation. While some existing approaches have utilized prior information as regularization in the optimization of demixing vectors, a mismatch in scale between the optimized vectors and the prior vectors, which are often defined in the back-projected domain, makes it difficult to control the strength of the prior influence, potentially degrading performance. To address this issue, we build upon our recent formulation that reparameterizes the optimization to directly update the back-projected demixing matrix. By incorporating the prior information in this domain, our method allows for more consistent and effective use of prior knowledge. Simulated experiments demonstrate that the proposed method achieves superior separation performance compared to conventional approaches.

External IDs:dblp:conf/apsipa/KoisoNO25