ISMIR 2026

RelFx

Beyond Dry References: Learning Relative Audio Effects Representations via Contrastive Distance Learning

RelFx learns the effect transformation between two audio signals and transfers that transformation to new musical content. This page presents case-matched audio comparisons under the Fx-Encoder++ MUSDB18 protocol.

Paper · arXiv GitHub · Source code
Xinlu Liu, Huibin Lin, Weixing Wei, Zhenhai Yan

Galaxy Audio Effect Team
Tencent Music Entertainment

27th International Society for Music Information Retrieval Conference

Matched audio comparisons

Within each case, all methods share the source, processed reference, seven-processor chain, and seven parameter initializations, and are evaluated against the same ground truth. Oracle additionally receives the reference dry signal.

Loading matched audio examples...

Relative representation

Encoding the change, not just the sound

Two weight-shared branches compare original and processed audio. Cross-attention exchanges information between the branches, and Diff-Gate fusion produces a compact representation of their effect transformation.

  • Content-aware comparison The paired encoder can separate the shared musical content from the processing difference.
  • No dry reference during representation training Training can use already-processed audio by comparing different transformations.
  • Useful for downstream production tasks The learned transformation guides effect retrieval, parameter optimization, and style transfer.
RelFx architecture with two CNN branches, cross-attention blocks, Diff-Gate fusion, and a projection head.

Self-ref

Forms an identity target from the processed reference. It does not require the reference dry signal.

Standard

Uses the source signal as a shared anchor when matching the reference and optimized transformations.

Oracle

Uses the ground-truth dry and wet reference pair. This analysis-only condition is not a deployable setting.

People

Authors and affiliations

Xinlu Liu · Huibin Lin · Weixing Wei · Zhenhai Yan

Xinlu Liu, Huibin Lin, and Zhenhai Yan are with the Galaxy Audio Effect Team at Tencent Music Entertainment. Weixing Wei is an independent researcher.

Contact: fayelliu@163.com

Source code: github.com/TMEGalaxyAudioEffect/relfx-ismir2026-release

Citation

@inproceedings{liu2026relfx,
  title     = {Beyond Dry References: Learning Relative Audio
               Effects Representations via Contrastive Distance Learning},
  author    = {Liu, Xinlu and Lin, Huibin and Wei, Weixing and Yan, Zhenhai},
  booktitle = {Proceedings of the 27th International Society for
               Music Information Retrieval Conference (ISMIR)},
  year      = {2026}
}