Abstract
Source separation remains a challenging problem, made even more challenging when the mixtures are distorted. We present a novel framework for the source separation from mixtures affected by clipping distortion. Our method combines ℓ1 minimization with knowledge about the geometry of clipped mixtures when the sources are disjoint in time. Our algorithm recovers the sources by solving a convex optimization problem, which is constrained by the clipping geometry. Comparative evaluation experiments show a significant increase in objective recovery performance of our proposed joint method compared to sequential approaches for speech and synthetic signals.