Abstract Because of the harsh deployment environment of the fibers, distributed acoustic sensing (DAS) data usually suffer from the low signal-to-noise ratio issue. Many methods, whether simple but efficient or sophisticated but effective, have been proposed for dealing with noise and recovering signals from DAS data. However, no matter what methods we apply, we will inevitably damage the signals, more or less, resulting in coherent signal leakage in the removed noise. Here, we present a method (SigRecover) for minimizing signal leakage by recovering useful signals from removed noise and its open-source package (see Data and Resources). We apply a robust dictionary learning framework to retrieve the coherent signals from removed noise that can be captured by a pretrained library of atoms (features). The atoms are obtained by a fast dictionary-learning approach from the initially denoised data. The proposed framework is a self-learning methodology, which does not require additional training datasets and thus is conveniently applicable to any input data. We use three well-processed examples from the literature to demonstrate the generic performance of the proposed method. The idea behind this article is inspired by similar methods widely used in the exploration seismology community for retrieving signal leakage and is promising not only for DAS data processing, but also for all other multichannel seismological datasets.