Stem separation models (Core ML)
Core ML conversions of community stem-separation models, used by
Lightspeed Separator on macOS. The app downloads each model the first time
it's needed, pinned to a specific commit of this repo and checked against the
SHA-256 listed in manifest.json.
| Model | Trained by | Download size | License as published |
|---|---|---|---|
| BS-Roformer SW (6 stems) | jarredou | 328 MB | No license published by the author |
| DrumSep MDX23C (6 stems) | aufr33 & jarredou | 201 MB | No license published by the authors |
| BS-Roformer Leap (vocals + instrumental) | unwa (pcunwa) | 111 MB, 111 MB | No license published by the author |
| Karaoke BS-Roformer (anvuew) | anvuew | 95 MB | GPL-3.0 |
| Karaoke BS-Roformer (frazer & becruily) | frazer & becruily | 95 MB | No license published by the authors |
| MVSep Mega 53 Stems v1 | ZFTurbo / MVSep | 71 MB, 60 MB, 708 MB | No separate license for the weights (code is MIT) |
Licenses
These models were trained and published by the people credited above, not by SunnymeadSound. Each folder states the license exactly as its author published it. Where an author published no license, none is granted here — whether and how you may use those weights is for you to decide. See LICENSES.md.
Format
Each file is a zipped .mlpackage (fp16 mlprogram). They need macOS 14 or
later, except Mega 53, which needs macOS 15. Inputs and outputs
are spectrogram tensors — the STFT/iSTFT happen in the app — so each folder's
card notes the shapes.