In this work, we present an approach to multichannel source separation that incorporates prior information, such as approximate demixing vectors, into the optimization process. Our method is built within the blind source separation (BSS) framework. Conventional BSS methods typically optimize the demixing matrix by minimizing an objective function, followed by a postprocessing step known as back-projection to restore the scale of the separated signals. However, this rescaling step causes a mismatch in scale between the optimization target and the available prior information, making it difficult to appropriately control the influence of the prior. This mismatch can degrade separation performance. To address this issue, we propose directly updating the back-projected demixing vectors, allowing both the optimization and the prior information to be handled on the same scale. Simulation experiments demonstrate that the proposed method outperforms conventional approaches, confirming its effectiveness. Work supported by JSPS KAKENHI Grant No. JP25H01150.
Koiso et al. (Wed,) studied this question.