Audio Soft Declipping Based On Weighted L-1-Norm

This paper addresses soft clipping in audio and speech signals, in which the distortion can be approximated by an invertible polynomial. Recent proposals have shown that the sparsity of the original signal is a useful prior information for performing blind compensation of such form of distortion. In this paper we introduce a weighted l(1)-norm cost function that explores both the sparsity and the spectrum profile of typical undistorted audio/speech signals. In our fully blind declipping proposal, the weights are calculated solely from the distorted signal, which renders the method adaptive to different signal characteristics. Our tests show that the resulting soft declipping solution outperforms recent sparsity-based strategies for both audio and speech signals.