Auxiliary-Function-Based Steering Vector Estimation Method for Spatially Regularized Independent Low-Rank Matrix Analysis

Blind source separation (BSS) is a technique to separate each source signal from observed mixtures without prior information, and independent low-rank matrix analysis (ILRMA) is one of the state-of-the-art BSS methods. As not a fully blind method, spatially regularized ILRMA (SR-ILRMA) achieves higher separation performance by using prior information about the demixing matrix or steering vectors (SVs) as a regularizer in ILRMA. In this paper, we propose a new fully blind method that models SVs using the time differences of arrival (TDOAs) and simultaneously estimates TDOAs, the demixing matrix, and other parameters. Then, we derive the update rule of TDOAs on the basis of the majorization-minimization algorithm, which guarantees the monotonic non-increase in the proposed cost function. We also propose the method of initializing TDOAs based on impulse responses for the proposed cost function. Numerical experiments confirmed that the proposed method achieves better separation performance than ILRMA, and initialization based on impulse responses is effective for the proposed method.