Adaptation of MVSep-MDX23 algorithm for Colab, with few tweaks:
Recent changes:
v2.4
- BS-Roformer models from viperx added
- MDX-InstHQ4 model added as optionnal
- Flac output
- Control input volume gain
- Filter vocals below 50Hz option
- Better chunking algo (no clicks)
- Some code cleaning
Full changelog :
- HQ3-Instr model replaced by VitLarge23 (thanks to MVSep)
- Improved MDXv2 processing (thanks to Anjok)
- Improved BigShifts algo (v2)
- BigShifts processing added to MDXv3 & VitLarge
- Faster folder batch processing
- Improved MDXv3 chunking code (thanks to HymnStudio)
- D1581 demo model replaced by new InstVocHQ MDXv3 model.
v2.2.1
- Added custom weights feature
- Fixed some bugs
- Fixed input: you can use a file or a folder as input now
v2.2
- Added MDXv3 compatibility
- Added MDXv3 demo model D1581 in vocals stem multiband ensemble.
- Added VOC-FT Fullband SRS instead of UVR-MDX-Instr-HQ3.
- Added 2stems feature : output only vocals/instrum (faster processing)
- Added 16bit output format option
- Added "BigShift trick" for MDX models
- Added separated overlap values for MDX, MDXv3 and Demucs
- Fixed volume compensation fine-tuning for MDX-VOC-FT
- Updated with MDX-VOC-FT instead of Kim Vocal 2
- Updated with new Kim Vocal 2 & UVR-MDX-Instr-HQ3 models
- Folder batch processing
- Fixed high frequency bleed in vocals
- Fixed volume compensation for MDX models
Original work by ZFTurbo/MVSep : https://github.com/ZFTurbo/MVSEP-MDX23-music-separation-model