Reproduce “Synth-JEPA: Joint Embedding Prediction for Renderer-Free Synthesizer Parameter Search” 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

Ben Hayes, Haokun Tian, Stefan Lattner

Synth-JEPA matches target sounds by searching synthesizer settings in a learned audio–parameter representation, without rendering candidates during inference. It trains on paired settings and rendered sounds, then searches with a hybrid objective. On held-out Surge XT targets, it achieves mean MSS 8.44 versus 12.75 for flow matching, and listeners prefer it in 85.1% of comparisons; results reflect this trained synthesizer setting.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.