Underdetermined BSS with multichannel complex NMF assuming W-disjoint orthogonality of source

This paper presents a new method for underdetermined Blind Source Separation (BSS), based on a concept called multichannel complex non-negative matrix factorization (NMF). The method assumes (1) that the time-frequency representations of sources have disjoint support (W-disjoint orthogonality of sources), and (2) that each source is modeled as a superposition of components whose amplitudes vary over time coherently across all frequencies (amplitude coherence of frequency components) in order to jointly solve the indeterminacy involved in the frequency domain underdetermined BSS problem. We confirmed experimentally that the present method performed reasonably well in terms of the signal-to-interference ratio when the mixing process was known.

I. Introduction

BLIND Source Separation (BSS) is a technique for reconstructing source signals from microphone inputs on the assumption that the source signals and transfer characteristics between sources and microphones are unknown.