Equivariance-based self-supervised learning for audio signal recovery from clipped measurements

Abstract--In numerous inverse problems, state-of-the-art solving strategies involve training neural networks from ground truth and associated measurement datasets that, however, may be expensive or impossible to collect. Recently, self-supervised learning techniques have emerged, with the major advantage of no longer requiring ground truth data. The present work contributes to extending equivariancebased y = A(x) + ϵ. (1) Declipping is a common nonlinear operator may be incomplete, i.e., m < n. Regularization distortion, typically occurring with analog-to-digital strategies based on incorporating prior information have been (ADC) converters when the dynamic range of the original widely studied and shown to have performance that strongly (analog) signal is too high. To overcome this limitation, most strategies are based on learning an inverse A. Related works operator of A on X, f Supervised learning were proven effective, with a dictionary trained from learning however suffers from two main limitations: (i) training time windows of clipped signal y. Other prior information sets can be difficult or impossible to obtain, e.g., in medical inspired on human perception [10], or leveraging the presence of data from multiple channels [11], [12], can be also used to LJ's research is supported by the FRS-FNRS (QuadSense, T.0160.24).