BS-RoFormer-SW (float16 safetensors)
A 6-stem music source separation model (Band-Split RoFormer), converted to float16 safetensors, to runs on Apple silicon with MLX Swift.
- Stems, in output order: bass, drums, other, vocals, guitar, piano
- Audio: 44.1 kHz stereo
- STFT: n_fft 2048, hop 512, window 2048; the DC bin is zeroed (
zero_dc) - Model: dim 256, depth 12 (one time and one frequency transformer layer each), 8 heads of 64, 62 bands
(
freqs_per_bandsin the config), mask estimator depth 2 - Inference: chunks of 588,800 samples (13.35 s), 2 overlaps (step of half a chunk)
Files
| File | Bytes | SHA-256 |
|---|---|---|
bs_roformer_sw.safetensors | 349,521,144 | 6c8303a829575d03f21562ea185be7b6b23e922052883dec1b9518ca00a920fc |
bs_roformer_sw_config.json | 1,145 | 193f6ef7fcd03f6144eac5017b2b3ec4b1808dbc5ee8dd94f484b72b1e014fdd |
The weights keep their original PyTorch names (the MSST BSRoformer module). The config holds the model and audio
settings, the stem order and the band layout. LumaBeat Monitor pins this repository's revision and both hashes, and
refuses files that do not match.
Origin
This is a format conversion only: no training or fine-tuning was done here.
- Source checkpoint:
BS-Roformer-SW.ckpt, SHA-25624e7d35ee9c64415673d3fd33e06a67cac2c103c5df6267ba1576459c775916e(699 MB), re-hosted by jarredou asjarredou/BS-ROFO-SW-Fixed(no longer available) and mirrored atlainlives/audio-separator-models(revision3b39120). - Conversion: the checkpoint was loaded with
torch.load(weights_only=True)(no code in it runs), the rotary frequencies were checked to be the defaults (they are recomputed at load time), and the tensors were written as float16 safetensors. The conversion is deterministic. - Checked against the PyTorch reference (ZFTurbo's Music-Source-Separation-Training): the float32 port matches within 1.5e-5, and float16 output is 53–66 dB SDR from the float32 output.
License
Unknown. The original trainer is unknown and no license was stated with the checkpoint; the re-host owner has said they were not involved in training and have no provenance information. This repository states that plainly rather than assigning a license it cannot grant.
The architecture code this model is built for is MIT-licensed: lucidrains/BS-RoFormer and ZFTurbo/Music-Source-Separation-Training. The BS-RoFormer architecture is from Music Source Separation with Band-Split RoPE Transformer (Lu et al., 2023).
If you trained this model or hold rights to it, please open a discussion on this repository: we will credit you, add your license, or take the files down.