Phase aliasing correction for robust blind source separation using DUET

Abstract

The Degenerate Unmixing Estimation Technique (DUET) is a technique for blind source separation (BSS). Unlike the ICA-based BSS techniques, DUET is a time–frequency scheme that relies on the so-called W-disjoint orthogonality (WDO) property of the source signals, which states that, statistically, the windowed Fourier transforms of different source signals have disjoint supports. In addition to being computationally very efficient, one of the advantages of DUET is its ability, at least in some cases, to separate n⩾3 source signals using only two mixtures. However, DUET is prone to phase wrap-around aliasing, which often leads to incorrectly separated sources with artifacts and distortions. This weakness limits the effectiveness of WDO and DUET as a robust tool for BSS. In this paper we present a method for correcting the phase wrap-around aliasing for WDO-based techniques such as DUET using over-sampled Fourier transform and modulo arithmetic. Experimental results have shown that this phase aliasing correction method is very effective, yielding a highly robust blind source separation for speech mixtures.