Determined Blind Source Separation Using Metric Projection and Proximity Operator of Log-Det Function Under Projection-back Constraint

Kazuki Matsumoto, Kohei Yatabe

Session: General

Keywords: Audio source separation, time-frequency domain, demixing matrix estimation, affine constraint, alternating direction method of multipliers (ADMM)

Abstract:

Determined blind source separation (BSS) formulated in the time-frequency domain usually applies pre- and post-processing called whitening and projection back.
They convert the BSS problem into an easier optimization problem,
and the existing methods have relied on them.
However, these pre- and post-processing steps modify the optimized signals,
which prevents us from directly modeling the source signals.
In this paper, to directly handle the signals in BSS, we define the projection-back constraint set and derive the metric projection and the proximity operator related to it.
We also propose two algorithms using the alternating direction method of multipliers (ADMM) and experimentally confirm their correctness.

Submission Number: 121