Reproduce “End-to-end Music Remastering System Using Self-supervised and Adversarial Training” on OpenResearch

About this paper

Junghyun Koo, Seungryeol Paik, Kyogu Lee

An end-to-end system for music remastering transforms the mastering style of an input audio track to match a reference, utilizing self-supervised and adversarial training. It achieves a statistically significant improvement in perceived mastering style similarity to a target reference, scoring 0.71 in MUSHRA listening tests compared to 0.67 for the original input.

Key insights

• Self-supervised contrastive learning on music segments proves effective for extracting robust, style-related audio features, mitigating the need for extensive labeled data.
• The integration of a pre-trained feature encoder and adversarial training substantially improves the perceptual realism and fidelity of audio generated for style transfer.
• Directly transforming already processed audio's mastering style is feasible, providing a versatile solution for remastering existing content where original source tracks are often unavailable.

Summary by alphaXiv.