Soft-Mask De-Mixing for Anechoic Mixtures

This paper extends a computationally efficient, soft-mask based source separation (SS) technique called Redress, to anechoic mixing scenarios. SS methods are an integral part of hearing aid research. We call the resulting method D-Redress. In its original form, Redress was intended for instantaneous mixing scenarios. Numerical evaluations demonstrate that soft-mask based techniques reduce the level of artifacts in the separated speech. Monte Carlo trials on 1000 real speech mixtures demonstrate that the D-Redress successfully extends Redress in terms of Overall-Perceptual (OPS), Target-Perceptual (TPS) scores and Human-Ear Intelligibility (HEI).