1) We have changed the way of selecting models in the menu. Now, instead of a dropdown menu, there is a list with the ability to display information about the models and statistics. If you wish, you can roll back to the old version of the list.
2) By popular demand, we have added the HQ5 instrumental model to the site for the MDX-B algorithm (vocals, instrumental).
3) We have published weights obtained on the MUSDB18 dataset for the top models BSRoformer, MelBandRoformer and SCNet XL. These weights can be an excellent starting point for training your own models.
4) We added three models from unwa and 2 models from becruily, which are based on the Mel-Band RoFormer architecture. All models are focused on increasing the fullness metric either for vocals or for instrumental. They give a fuller sound but may contain more noise. The new models are available under the names:
- unwa Instrumental v1 (SDR vocals: 10.24, SDR instrum: 16.54)
- unwa Instrumental v1e (SDR vocals: 10.05, SDR instrum: 16.36)
- unwa big beta v5e (SDR vocals: 10.59, SDR instrum: 16.89)
- becruily instrum high fullness (SDR instrum: 16.47)
- becruily vocals high fullness (SDR vocals: 10.55)
The models are located in the "MelBand Roformer (vocals, instrumental)" section. Detailed metrics are available in the table below:
| Model | Vocals fullness | Vocals bleedless | Vocals SDR | Vocals L1Freq | Instrum fullness | Instrum bleedless | Instrum SDR | Instrum L1Freq |
| MelBand Roformer (Kimberley Jensen) | 16.66 | 36.51 | 11.01 | 38.96 | 27.71 | 46.72 | 17.32 | 39.77 |
| MelBand Roformer (ver. 2024.08) | 16.39 | 39.13 | 11.18 | 39.26 | 27.74 | 47.07 | 17.49 | 40.16 |
| Bas Curtiz edition | 16.30 | 38.94 | 11.18 | 39.18 | 27.49 | 47.00 | 17.49 | 40.15 |
| MelBand Roformer (ver. 2024.10) | 16.92 | 37.78 | 11.28 | 39.41 | 27.71 | 47.29 | 17.59 | 40.29 |
| unwa Instrumental v1 (SDR vocals: 10.24, SDR instrum: 16.54) | 15.89 | 27.48 | 10.24 | 36.06 | 35.44 | 38.02 | 16.55 | 38.67 |
| unwa Instrumental v1e (SDR vocals: 10.05, SDR instrum: 16.36) | 14.67 | 26.83 | 10.06 | 34.37 | 38.85 | 35.68 | 16.37 | 38.31 |
| unwa big beta v5e (SDR vocals: 10.59, SDR instrum: 16.89) | 20.78 | 32.02 | 10.59 | 38.53 | 25.65 | 45.90 | 16.90 | 37.31 |
| becruily instrum high fullness (SDR instrum: 16.47) | 15.76 | 30.15 | 10.16 | 35.84 | 33.93 | 40.55 | 16.47 | 38.86 |
| becruily vocals high fullness (SDR vocals: 10.55) | 20.72 | 31.25 | 10.55 | 38.84 | 28.28 | 40.85 | 16.86 | 38.24 |
5) We have added 2 models from lew for Super Resolution task. The first "Universal Super Resolution (by Lew)" - restores high frequencies for music, the second more specialized "Vocals Super Resolution (by Lew)" restores the quality and high frequencies for vocals. They are available for selection in the menu under the item "Apollo Enhancers (by JusperLee and Lew)".
6) We have added a set of models for separating vocals into Male/Female. There are 2 models from Sucial and aufr33. There are also two models trained by the MVSep team based on SCNet XL and MelBand RoFormer. All models available in "MVSep Male/Female separation".
| Algorithm name | Male/Female validation dataset |
|||
| SDR Male | SDR Female | L1_Freq Male | L1_Freq Female | |
| BSRoformer by Sucial (SDR: 6.52) | 6.82 | 6.23 | 40.99 | 40.62 |
| BSRoformer by aufr33 (SDR: 8.18) | 8.47 | 7.89 | 46.65 | 44.73 |
| SCNet XL (SDR: 11.83) | 12.08 | 11.58 | 50.50 | 51.51 |
| MelRoformer (2025.01) (SDR: 13.03) | 13.39 | 12.68 | 57.61 | 56.76 |
7) We have added a new SCNet XL model for bass with a very high SDR: 13.81. In the ensemble, the SDR metric reached 14.07, which is a record. The model is available under the item MVSep Bass (bass, other)
8) We have added the second version of the model for removing the dereverberation effect from Sucial to the Reverb Removal (noreverb) section. Model name: Reverb removal by Sucial v2 (MelRoformer).
9) We have prepared a new model for vocals based on the SCNet XL architecture, it has achieved quite high metrics.
| Algorithm name | Multisong dataset | Synth dataset | MDX23 Leaderboard |
||
| SDR Vocals | SDR Instrumental | SDR Vocals | SDR Instrumental | SDR Vocals | |
| SCNet | 10.25 | 16.56 | 12.27 | 11.97 | --- |
| SCNet Large | 10.74 | 17.05 | 12.89 | 12.59 | --- |
| SCNet XL | 10.96 | 17.27 | 13.08 | 12.78 | --- |
Adding SCNet XL to Mel and BS roformers in the ensemble increased the SDR metric:
vocals: 11.54 -> 11.61
instrumental: 17.84 -> 17.92
10) We have added a new model for organ musical instrument. It is available in the list under the name: MVSep Organ (organ, other).
11) We have updated our API, adding more functionality related to the task queue, rating, and the use of different types of separation, as well as added a Quality Checker to the API. More information is available in the documentation: https://mvsep.com/full_api
12) We are testing an Android application, it will soon appear on Google Play. We will announce this separately.
13) In the near future, we plan to publish examples of using the MVSep API in Python. Both simple console programs and those with a graphical interface.
