WildFX: A DAW-Powered Pipeline for In-the-Wild Audio FX Graph Modeling
Accepted as an oral paper at DAFx-26. WildFX appears in Paper Session 5: Sound Design and Effects at MIT in Cambridge, Massachusetts.
Overview
WildFX bridges machine-learning research and professional music-production workflows by giving Python programmatic control over a REAPER digital audio workstation running inside Docker. The pipeline generates multitrack datasets with realistic, heterogeneous audio-effects graphs instead of limiting experiments to isolated or simplified differentiable processors.
What WildFX enables
- Integration with commercial and open-source plugins in VST, VST3, LV2, and CLAP formats.
- Complex signal routing, including sidechains, parallel paths, and multiband processing.
- Scalable, parallelized rendering from compact YAML and JSON project metadata.
- Research on plugin classification, parameter estimation, mixing-graph inference, grey-box modeling, and music-aware source separation.
An earlier version of this work appeared at the AI Heard That! ICML 2025 Workshop on Machine Learning for Audio.