StemFX AI learns mixing styles by predicting FX chains on separated stems

StemFX addresses a core challenge in AI-assisted music production: capturing a mix engineer's style through FX chain decisions (choice, ordering, parameterization of effects per stem). Previous approaches either worked on stereo mixes without per-stem modeling, required differentiable effects, or relied on scarce multitrack data. The authors solve this by training on pseudo-stems extracted from a large corpus of ~105K songs via source separation, then prompting a Transformer decoder to predict variable-length tokenized FX chains. A band-split multi-band CNN encoder with FiLM conditioning captures spectral structure per stem, while the MultiAFx toolkit (unifying 85 effects from 7 Python libraries) enables data augmentation and scalable training.

On mixing style retrieval, StemFX outperforms all baselines across all tested chain lengths. For paired mixing style transfer, it achieves the best spectral fidelity and highest listener preference, and it is over 4000 times faster than iterative optimization methods. The model does not require differentiable effect implementations, making it practical for real-world DAW integration. By learning from existing mixes, StemFX opens the door to style-consistent auto-mixing, AI-assisted remixing, and educational tools that explain FX chain decisions. Accepted to ISMIR 2026, the work signals a shift toward expressive, data-driven mixing style modeling.