Reproduce “Learning to Denoise Historical Music” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yunpeng Li, Beat Gfeller, Marco Tagliasacchi, Dominik Roblek
Researchers from Google developed an automated, multi-scale deep learning model for denoising historical music, overcoming the lack of clean ground truth by creating a synthetic dataset from real noise samples and modern clean music. The approach achieved superior audio quality, yielding a 3.4 dB SNR gain and producing outputs that human listeners found perceptually indistinguishable from clean originals on synthetic data.
Key insights
- Synthetically generated training data, derived from real historical noise and modern clean music, effectively enables supervised deep learning for a task where clean ground truth is otherwise unattainable.
- A multi-scale U-Net processing STFT representations with full phase information is crucial for achieving high-quality music denoising, significantly outperforming models that bypass phase or operate solely in the waveform domain.
- While adversarial loss (GANs) can improve perceptual quality, models trained purely with reconstruction loss can achieve outputs statistically indistinguishable from clean ground truth, suggesting its strong utility for high-fidelity restoration in certain contexts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.