Hybrid Transformers for Music Source Separation¶
Summary¶
Extends Hybrid Demucs by replacing the innermost U-Net layers with a cross-domain Transformer encoder. The Transformer applies self-attention within each domain (time, frequency) and cross-attention between them. Achieved 9.20 dB SDR on MUSDB18 with extra training data, establishing a new state of the art. The hybrid architecture lets the model learn whether waveform or spectrogram processing is better for each source.
Key Claims¶
- Cross-domain Transformer attention significantly outperforms LSTM bottlenecks
- Self-attention within time/frequency domains + cross-attention between them captures long-range dependencies
- The model learns to weight waveform vs. spectrogram processing per source (vocals benefit from spectrogram, drums from waveform)
- 4-stem and 6-stem variants available; both outperform all prior work
Related¶
- ../sources/2019-11-27-demucs — original Demucs paper
- ../entities/demucs — the tool (all versions in one repo)
- ../entities/bs-roformer — current SOTA (2023) that surpassed HTDemucs
- ../concepts/permutation-invariant-training — training technique used