Diffusion-based denoising of historical recordings

Abstract

In the context of audio restoration, background noise removal in historical music recordings is a relevant topic for which the use of signal processing and traditional supervised deep learning methods has been previously studied. In this work, a generative approach for removing perceptually distributed noise is investigated by adapting diffusion conditional sampling for zero-shot background noise removal in digitized classical solo piano 78 RPM recordings. Existing diffusion conditional sampling methods used in audio restoration typically require a deterministic model of the degradation being removed, which would not be possible for background noise removal in 78 RPM recordings. The proposed method overcomes this by using a set of noise examples to simulate 78 RPM defects during conditional sampling, which also makes the method potentially generalizable to any random additive degradation. Experiments using both real historical recordings and artificially generated examples show that diffusion models have comparable restoration performance to state-of-theart supervised deep learning methods in most settings. However, there is a tradeoff: while supervised deep learning methods tend to remove all noise at the cost of cutting off high-frequency signal components, the proposed diffusion approach preserves signal integrity, but leaves some residual noise.