Design of new auxiliary function for fully blind spatially regularized independent low-rank matrix analysis

Le résumé fourni par la source

A representative method for blind source separation (BSS) is independent low-rank matrix analysis (ILRMA). Spatially regularized ILRMA (SR-ILRMA) utilizes prior information about the acoustic transfer system, such as steering vectors (SVs) for each source, as a regularizer in ILRMA. Although it has been reported that SR-ILRMA achieved higher separation performance than ILRMA in an experiment, SR-ILRMA is not a fully blind method; i.e., SVs should be known in advance. In our previous study, we proposed a fully blind SR-ILRMA that simultaneously estimates SVs and other parameters in SR-ILRMA on the basis of the majorization–minimization (MM) algorithm. Since an auxiliary function is not unique in the MM algorithm, it may be possible to design a better auxiliary function that achieves faster convergence. In this paper, we design a new auxiliary function for deriving the update rule of SVs, motivated by the conjecture that an auxiliary function that better approximates the cost function would lead to faster convergence. In a numerical experiment, we confirm that the update rule based on the new auxiliary function achieves faster convergence than that based on the conventional one.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.