M. Kowalski (L2S): Social Sparsity: application to audio inpainting.
Date(s) : 28/03/2013 iCal
14h00 - 15h00
Social Sparsity: application to audio inpainting.\n\nBy Matthieu Kowalski\, L2S.\n\nAbstract:\nAudio inpainting problem is under consideration\, using iterative\nthresholding algorithms build on the « social sparsity » principle.\nFirst\, we present new shrinkage operators\, allowing one to take into\naccount the neighborhood of time-frequency coefficients. Then\, the\naudio declipping problem is formulated as a unconstrained convex\noptimization problem\, but taking into account an inportant hypothesis\nof audio declipping: reconstructed samples must be greater than the\nclipping threshold. The structured thresholding operators\, such as the\nwindowed group-Lasso or the persistent empirical wiener\, are embedded\ninto iterative algorithms\, and we show on experimental results the SNR\nimprovement compared to a more conventional Lasso. We also compare the\nresults to the state of the art audio declipping.\n\nDownload slides
Catégories
