"Sparse Audio Inpainting via Convex-Concave Saddle Point Optimization"Nenov, RossenAudio inpainting aims to restore missing parts of an audio signal, striving to reconstruct the original or create a plausible alternative. In [1] an innovative problem formulation was introduced that employs sparse modeling within the time-frequency (TF) domain. The method of Alternating Direction Method of Multipliers was applied to solve a relaxed version of the proposed problem. In this contribution we rewrite the proposed formulation as a convex-concave saddle point problem with a nonsmooth coupling function and propose a method for this optimization task. Additionally we state the convergence guarantee of this method and present experiments demonstrating improvements in terms of signal-to-distortion ratio. [1] Tauböck, Georg and Rajbamshi, Shristi and Balazs, Peter. (2020). Dictionary Learning for Sparse Audio Inpainting. IEEE Journal of Selected Topics in Signal Processing. PP. 1-1. |
« back