Deep Audio Waveform Prior

TL;DR

This paper reveals that state-of-the-art neural network architectures for audio source separation inherently contain deep priors that can be exploited for unsupervised audio restoration tasks such as denoising, inpainting, and dereverberation using raw waveforms.

Contribution

It demonstrates that existing SOTA audio models have deep priors in raw waveform domain, enabling unsupervised restoration without specialized architectures or spectrogram inputs.

Findings

Neural networks can generate cleaner audio signals before fitting the corrupted data.

Deep priors are effective for removing background noise, reverberations, and filling gaps in audio.

Raw waveform models contain inherent priors useful for various audio restoration tasks.

Abstract

Convolutional neural networks contain strong priors for generating natural looking images [1]. These priors enable image denoising, super resolution, and inpainting in an unsupervised manner. Previous attempts to demonstrate similar ideas in audio, namely deep audio priors, (i) use hand picked architectures such as harmonic convolutions, (ii) only work with spectrogram input, and (iii) have been used mostly for eliminating Gaussian noise [2]. In this work we show that existing SOTA architectures for audio source separation contain deep priors even when working with the raw waveform. Deep priors can be discovered by training a neural network to generate a single corrupted signal when given white noise as input. A network with relevant deep priors is likely to generate a cleaner version of the signal before converging on the corrupted signal. We demonstrate this restoration effect with several corruptions: background noise, reverberations, and a gap in the signal (audio inpainting).