The objective of audio inpainting is to fill a gap in a signal, either to be meaningful or even to reconstruct the original signal. We propose a novel approach applying sparse modeling in the time-frequency (TF) domain. In particular, we develop a dictionary learning technique which deforms a given Gabor frame such that the sparsity of the analysis coefficients of the resulting frame is maximized. A suitable modification of the SParse Audio Inpainter (SPAIN) allows to exploit the obtained sparsity gain and, hence, to benefit from the learned dictionary. Our experiments demonstrate that our methods outperforms several state-of-the-art audio inpainting techniques in terms of signal-to-noise ratio (SNR) and objective difference grade (ODG).
Download Now (1.2 MB)