Summary

Demucs is a state-of-the-art music source separation model developed by Meta AI, capable of separating drums, bass, vocals, and other stems from songs. The latest version (v4) features a Hybrid Transformer architecture combining spectrogram and waveform domains, achieving competitive separation performance.

Features

• Hybrid Transformer architecture with cross-domain self-attention and cross-attention layers.
• Separates multiple sources including drums, bass, vocals, guitar, and piano (experimental).
• Based on a U-Net convolutional architecture inspired by Wave-U-Net.
• Achieves high signal-to-distortion ratio (SDR) scores on the MUSDB HQ test set.
• Supports models with and without fine-tuning; sparse attention kernels for extended receptive fields are under development.
• Trained on MUSDB HQ dataset plus an additional 800-song training dataset.
• Provides command-line options to activate different model versions.
• Open-source with source code and models available on GitHub.

Additional Information

The project is no longer actively maintained by Meta; a fork by a community member exists for ongoing bug fixes. An experimental 6-source model includes guitar and piano extraction, though piano results currently show artifacts. Detailed metrics and comparisons with other source separation models are provided in the papers linked in the repository.