Moises
Diff-VS
- Diffusion-based vocal separation built on the EDM framework
- Achieves results on par with many discriminative models in both subjective and objective evaluations
- Requires as few as 7 steps to reach optimal performance
Objective Evaluation
Train and evaluate on MUSDB
| Model | Type | Params | cSDR ↑ |
|---|---|---|---|
| HDemucs [6] | disc | 42 M | 8.13 |
| TFC-TDF V3 [7] | disc | 70 M | 9.59 |
| BSRNN [5] | disc | 37 M | 10.01 |
| BS-RoFormer-6L [3] | disc | 72 M | 10.66 |
| SCNet-L [2] | disc | 42 M | 10.86 |
| Diff-VS | gen | 57 M | 10.12 |
Train on MUSDB + MoisesDB and evaluate on MUSDB
| Model | Type | Params | cSDR ↑ |
|---|---|---|---|
| SCNet-L [2] | disc | 42 M | 11.11 |
| SGMSEVS [1] | gen | 65 M | 8.63 |
| Diff-VS | gen | 57 M | 10.88 |
Subjective Evaluation
We use MERT-MSE as our proxy subjective metrics [1]
| Model | Type | Params | MERT MSE ↓ |
|---|---|---|---|
| SCNet-L [2] | disc | 42 M | 0.096 |
| SGMSEVS [1] | gen | 65 M | 0.089 |
| Mel-Roformer [3] | disc | 105 M | 0.071 |
| Diff-VS V2 | gen | 54 M | 0.083 |
Listening Samples
Listening samples are randomly chosen from MUSDB test set
Little Chicago’s Finest - My Own (SDR 15.41)
PR - Happy Daze (SDR -0.58)
The Long Wait - Dark Horses (SDR 11.62)
Timboz - Pony (SDR 5.2)
We Fell From The Sky - Not You (SDR 8.44)
Teaser
Ongoing work on Diff-VS v2 can achieve on par results with BS-RoFormer and SCNet
Train and evaluate on MUSDB
| Model | Type | Params | cSDR ↑ |
|---|---|---|---|
| SCNet-L [2] | disc | 42 M | 10.86 |
| BS-RoFormer-12L [3] | disc | 93 M | 11.49 |
| Mel-Roformer [4] | disc | 105 M | 12.08 |
| Diff-VS V2 | gen | 54 M | 11.46 |
Train with MoisesDB + MUSDB and evaluate on MUSDB
| Model | Type | Params | cSDR ↑ |
|---|---|---|---|
| SCNet-L [2] | disc | 42 M | 11.11 |
| SGMSEVS [1] | gen | 65 M | 8.63 |
| Diff-VS V2 | gen | 54 M | 11.85 |
Little Chicago’s Finest - My Own
PR - Happy Daze
The Long Wait - Dark Horses
Timboz - Pony
We Fell From The Sky - Not You
Reference
[1] Bereuter, Paul A., et al., “Towards Reliable Objective Evaluation Metrics for Generative Singing Voice Separation Models”, IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), 2025.
[2] Tong, Weinan, et al. “Scnet: Sparse compression network for music source separation.”, IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024.
[3] Lu, Wei-Tsung, et al. “Music source separation with band-split rope transformer.”, IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024.
[4] Wang, Ju-Chiang, Wei-Tsung Lu, and Minz Won. “Mel-band roformer for music source separation.” arXiv preprint arXiv:2310.01809.
[5] Luo, Yi, and Jianwei Yu. “Music source separation with band-split RNN.” IEEE/ACM Transactions on Audio, Speech, and Language Processing 31 (2023): 1893-1901.
[6] Rouard, Simon, Francisco Massa, and Alexandre Défossez. “Hybrid transformers for music source separation.”, IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023.
[7] Kim, Minseok, Jun Hyung Lee, and Soonyoung Jung. “Sound demixing challenge 2023 music demixing track technical report: Tfc-tdf-unet v3.” arXiv preprint arXiv:2306.09382 (2023).