1) We have introduced a variety of new models for separating individual instruments:
| 1) MVSep Harp (Demo) | 2) MVSep Double Bass (Demo) | 3) MVSep Mandolin (Demo) |
| 4) MVSep Trombone (Demo) | 5) MVSep Tambourine (Demo) | 6) MVSep Oboe (Demo) |
| 7) MVSep Clarinet (Demo) | 8) MVSep Digital Piano (Demo) | 9) MVSep Synth (Demo) |
| 10) MVSep Electric Guitar (Demo) | 11) MVSep French Horn (Demo) | 12) MVSep Banjo (Demo) |
| 13) MVSep Marimba (Demo) | 14) MVSep Glockenspiel (Demo) | 15) MVSep Timpani (Demo) |
| 16) MVSep Harmonica (Demo) | 17) MVSep Triangle (Demo) | 18) MVSep Sitar (Demo) |
| 19) MVSep Harpsichord (Demo) | 20) MVSep Tuba (Demo) | 21) MVSep Bassoon (Demo) |
| 22) MVSep Congas (Demo) | 23) MVSep Bells (Demo) | 24) MVSep Ukulele (Demo) |
| 25) MVSep Dobro (Demo) | 26) MVSep Wind Chimes (Demo) | 27) MVSep Accordion (Demo) |
2) Two models have been updated:
- A new saxophone model based on the BSRoformer architecture has been added. It demonstrates significantly improved metrics compared to the previous version, with the SDR increasing from 7.13 to 9.77. This model is available in "MVSep Saxophone (saxophone, other)" under the "BS Roformer (SDR saxophone: 9.77)" option.
- A new version of the MVSep Organ (organ, other) model has been released: "BS Roformer (SDR organ: 5.08)". The SDR has increased from 3.05 to 5.08.
3) Current separation scheme can be found below:
4) We have added two new Karaoke models to "MVSep Karaoke (lead/back vocals)":
- The first was trained by our team. It's available under the option "BS Roformer by MVSep Team (SDR: 10.41)". Unlike other Karaoke models it returns three stems: "lead", "back" and "instrumental". (Demo) (Metrics).
- The second karaoke model was trained by @anvuew. It is available under the option "BS Roformer by anvuew (SDR: 10.22)". (Metrics).
5) A model by baicai1145 was added to "Apollo Enhancers (by JusperLee, Lew, baicai1145)" under the name "Universal Super Resolution (by baicai1145)". (Original model link). A new option has been added for Apollo Enhancers (by JusperLee, Lew, baicai1145): Cutoff (Hz). Sometimes it can be useful to cut higher frequencies before applying the model.
6) We added a new DeReverb model by @anvuew. It's available in Reverb Removal (noreverb) under the option DeReverb room by anvuew (BSRoformer). It works only for vocals. Since it is a mono model it processes stereo channels independently. (Demo)
7) We added a new model "SCNet XL IHF (high instrum fullness by bercuily)" to the "SCNet (vocals, instrumental)" algorithm. It's a high-fullness version o instrumental stem prepared by @becruily.
| Model | Vocals fullness | Vocals bleedless | Vocals SDR | Vocals L1Freq | Instrum fullness | Instrum bleedless | Instrum SDR | Instrum L1Freq |
| SCNet XL IHF | 17.98 | 28.31 | 11.11 | 37.91 | 28.87 | 40.37 | 17.41 | 38.54 |
| SCNet XL IHF (high instrum fullness by becruily) | 22.70 | 25.48 | 10.89 | 38.18 | 32.31 | 38.15 | 17.20 | 38.43 |
8) We have released the beta version of the MVSep App for iOS: https://testflight.apple.com/join/Js7BwfEC It supports iOS 15.6+. Feel free to test it and report any bugs to us.
