Reproduce “Music Augmentation and Denoising For Peak-Based Audio Fingerprinting” 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
Kamil Akesbi, Dorian Desblancs, Benjamin Martin
Researchers at Deezer developed a hybrid system to improve the noise robustness of peak-based audio fingerprinting (AFP) by pre-processing noisy audio with a deep learning denoiser. Their "Mix" strategy, which combines hashes from both original noisy and denoised signals, consistently increased music identification rates in challenging environments by up to 10.5%.
Key insights
- A comprehensive and stochastic data augmentation pipeline is critical for training deep learning denoisers to effectively handle the complexity and variability of real-world noisy music recordings.
- While direct denoising improves spectral peak preservation, simply using denoised audio alone for AFP does not consistently outperform noisy baselines, as denoising can sometimes introduce new artifacts or remove valid peaks.
- Combining information from both the original noisy and the denoised audio (the "Mix" strategy) leverages their complementary strengths, effectively mitigating individual weaknesses and yielding superior music identification rates.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.