Open-Unmix

About

Reference implementation for music source separation (JOSS/ISMIR 2019). 3-layer bidirectional LSTM with fully-connected layers, pure spectrogram-domain. Designed for reproducibility and hackability, not SOTA performance. GitHub: sigsep/open-unmix-pytorch (1.5k stars). Install: pip install openunmix.

Relevance

Most trainable/hackable separator. Since it's designed as a reference implementation, it's the best candidate for fine-tuning on bluegrass stems if labeled training data can be created. Simple architecture is easy to modify and retrain.

Mentions

• ../entities/demucs — higher quality but more complex
• ../entities/spleeter — faster but less hackable
• ../concepts/synthetic-mixing-pipelines — training data strategy