[REQ] BS-RoFormer separation model(s)

[MarcoRavich]

Hi there,
I've recently tested the model (which seems to outperform the best Demucs version (htdemucs), according to this MVSEP's - vocal - leaderboard (about 3 sdr points: 11.8938 > 8.7869).

My (lossless) demos:

• https://mvsep.com/result/20260318182458-252cd55a1d-un-cavalier-di-spagna.flac
• https://mvsep.com/result/20260318193400-a786943fd2-le-corsaire.flac

It would be interesting to exploit this so here's some - maybe - useful links:

HF: https://huggingface.co/models?sort=modified&search=BSRoformer

ONNX: https://huggingface.co/xycld/BS-RoFormer-ONNX

CPP version by @chenmozhijin: BSRoformer.cpp

Thanks in advance.

[RyanMetcalfeInt8]

Hey @MarcoRavich,

On our next branch, we've enabled MelBand Roformer (The @KimberleyJensen edition) -- https://github.com/intel/openvino-plugins-ai-audacity/blob/next/model_cards/music_separation/mel_vocals_kimberley_jenson.md

As far as I can tell, this is the highest quality voice separation model which has openly available weights.

SDR of 10.98 according to this list: https://github.com/ZFTurbo/Music-Source-Separation-Training/blob/main/docs/pretrained_models.md

The BS Roformer which shows an SDR of 11.89 from your leaderboard link I think is some MVSep-trained variant, which from what I can tell is proprietary as the weights aren't posted anywhere.

[MarcoRavich]

MVSep Mega 53 Stems model has been just released.

[RyanMetcalfeInt8]

MVSep Mega 53 Stems model has been just released.

Cool, thanks for the heads up. Have you found metrics (SDR, etc.). for this model?