Algorithm Demucs3 splits track into 4 stems (bass, drums, vocals, other). The winner of the Music Demuxing Challenge 2021.
Link: https://github.com/facebookresearch/demucs/tree/v3
Quality table
| Algorithm name | Multisong dataset | Synth dataset | |||||
| SDR Bass | SDR Drums | SDR Other | SDR Vocals | SDR Instrumental | SDR Vocals | SDR Instrumental | |
| Demucs3 (Model A) | 9.50 | 8.97 | 4.40 | 7.21 | 13.52 | --- | --- |
| Demucs3 (Model B) | 10.69 | 10.27 | 5.35 | 8.13 | 14.44 | 9.78 | 9.48 |
Note: For version A only MUSDB18 training data was used for training, so quality is worse than Demucs3 Model B. Demucs3 Model A and Demucs3 Model B has the same architecture, but has different weights.
