Exemplar-based Assignment of Large Missing Audio Parts using String Matching on Tonal Features

We propose a new approach for assigning audio data in large missing audio parts (from 1 to 16 seconds). Inspired by image inpainting approaches, the proposed method uses the repetitive aspect of music pieces on musical features to recover missing segments via an exemplar-based reconstruction. Tonal features combined with a string matching technique allows locating repeated segments accurately. The evaluation consists in performing on both musician and nonmusician subjects listening tests of randomly reconstructed audio excerpts, and experiments highlight good results in assigning musically relevant parts. The contribution of this paper is twofold: bringing musical features to solve a signal processing problem in the case of large missing audio parts, and successfully applying exemplar-based techniques on musical signals while keeping a musical consistency on audio pieces.