Inconsistent Audio Declipping Performance Enhancement Based On Audio Inpainting

Abstract

Some of the state-of-the-art audio declipping methods are not consistent with the observedsignal, meaning that they do not keep the a priori undegraded (reliable) samples. These samples canbe directly substituted in the reconstructed signals, but this may create audible artifacts due to thesharp transitions between the declipped and the substituted parts. We propose two methods based onaudio inpainting, which deal with these transitions. As a result, we observe a significant improvementof the PEAQ ODG values, but only for some of the declipping algorithms considered.