Dictionary Learning for Sparse Audio Inpainting

The objective of audio inpainting is to fill a gap in an signal. This ideally done by reconstructing the original signal or, at least, inferring meaningful surrogate We propose novel approach applying sparse modeling time-frequency (TF) domain. In particular, we devise dictionary learning technique which learns from reliable parts around with goal obtain representation increased TF sparsity. based on basis optimization deform given Gabor frame such that sparsity analysis coefficients resulting maximized. Furthermore, modify SParse Audio INpainter (SPAIN) for both and synthesis model it able exploit and—in turn—benefits learning. Our experiments demonstrate developed methods achieve significant gains terms signal-to-distortion ratio (SDR) difference grade (ODG) compared several state-of-the-art techniques.