Open-Unmix

About

Reference implementation for music source separation from Sony/Inria. 3-layer bidirectional LSTM with fully-connected layers operating on spectrograms. Designed for reproducibility and modularity, not SOTA performance. Published in JOSS. The most hackable/trainable separation framework — easy to modify architecture and train on custom data.

Relevance

Best candidate for bluegrass-specific fine-tuning. If Nat can create even a small set of labeled bluegrass stems, Open-Unmix can be retrained to separate banjo/mandolin/fiddle instead of the generic "other" category. The simple architecture and clean codebase make it the most practical platform for custom separation research.

Mentions

• ../entities/demucs — higher quality, less hackable
• ../entities/asteroid — PyTorch toolkit including Open-Unmix recipes
• ../concepts/permutation-invariant-training — training technique

Links

• GitHub: https://github.com/sigsep/open-unmix-pytorch (1.5k stars)
• Install: pip install openunmix