In the context of audio restoration, the need to remove background noise from historical music recordings is a recurring problem, for which traditional signal processing and supervised deep learning methods have been previously applied. In this work, a generative approach that adapts conditional diffusion sampling for removing perceptually distributed noise is investigated, using the particular case of background noise removal from solo classical piano recordings as a proof of concept. The proposed method uses a set of noise examples to simulate perceptually distributed noise with specific characteristics throughout conditional diffusion sampling. Experiments with real historical 78 RPM recordings and clean recordings with added 78 RPM noise and tape hiss demonstrate that diffusion-based audio denoising performs comparably to state-of-the-art deep learning methods.
(See document for exact affiliation information.)
https://aes2.org/publications/elibrary-page/?id=22814
(341KB)
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member Join the AES. If you need to check your member status, login to the Member Portal.
