Compressing Audio Signals with Inpainting-Based Sparsification

Abstract

Inpainting techniques are becoming increasingly important for lossy image compression. In this paper, we investigate if successful ideas from inpainting-based codecs for images can be transferred to lossy audio compression. To this end, we propose a framework that creates a sparse representation of the audio signal directly in the sample-domain. We select samples with a greedy sparsification approach and store this optimised data with entropy coding. Decoding restores the missing samples with well-known 1-D interpolation techniques. Our evaluation on music pieces in a stereo format suggests that the lossy compression of our proof-of-concept framework is quantitatively competitive to transform-based audio codecs such as mp3, AAC, and Vorbis.