SonicMaster: research on controllable music mastering
By Sequential AI · Updated
SonicMaster is a research model that can restore and master a piece of music in one step, steered by plain-language instructions or run automatically. It was written by Jan Melechovsky, Ambuj Mehrish, Abhinaba Roy and Prof. Dorien Herremans, and presented at ICML 2026. The paper is on arXiv.
What problem does SonicMaster address?
Recordings can go wrong in many ways at once: too much reverb, uneven tone, squashed dynamics, a narrow stereo image. Fixing each usually takes a different tool and a trained ear. SonicMaster is presented as a single model that handles many of these problems together, and that can follow a written instruction such as "reduce the reverb" or run in an automatic restoration mode.
How does it work?
- Training data. The authors built a dataset of paired tracks, a degraded version and a high-quality version, by simulating common problems with nineteen processing functions in five groups: equalization, dynamics, reverb, amplitude and stereo.
- The model. A flow-matching model learns to turn the degraded audio back into the high-quality audio, guided by a text prompt describing the fix.
- Two ways to use it. Give it a text instruction for a specific change, or leave it to restore the track automatically.
What did the authors find?
Objective quality metrics improved across every category of problem the authors tested, and in listening tests participants preferred SonicMaster's output over the other systems it was compared with. The paper has the full numbers and the comparison systems.
How does this relate to Sequential AI?
The paper comes from the research team behind Sequential AI. This page describes the paper; for what the Sequential AI Studio does for your tracks today, see how it works.
Where can I read the paper or hear examples?
How do I cite it?
@misc{melechovsky2025sonicmaster,
title = {SonicMaster: Towards Controllable All-in-One Music Restoration and Mastering},
author = {Melechovsky, Jan and Mehrish, Ambuj and Roy, Abhinaba and Herremans, Dorien},
year = {2025},
eprint = {2508.03448},
archivePrefix = {arXiv}
}