You can deploy this model and start earning money today!

smank/
mel-band-roformer-vocals-onnx

Name: mel-band-roformer-vocals-onnx
Availability: InStock

onnx

audio

music-source-separation

Views

No views yet

Mel-Band RoFormer (vocals) — ONNX conversion

ONNX conversion of the KimberleyJensen Mel-Band RoFormer vocals model (SDR 10.98 on MUSDB18HQ — among the strongest public vocal isolators), packaged for SonicDemixer (on-device stem separation for macOS).

melband_roformer_vocals.onnx — input mix [1, 2, 352800] float32 planar waveform → output sources [1, 2, 2, 352800] where source 0 is vocals and source 1 is other (mix − vocals). Opset 17. The graph embeds a conv-based STFT (n_fft 2048, hop 441), the mel-band gather, the model's mask estimation, mask scatter/averaging, complex multiply, and a conv-based iSTFT — no complex dtypes, no external pre/post-processing needed. Numerically verified against the reference PyTorch pipeline (max |diff| ≈ 6e-8).

Integrity: LFS-tracked; the LFS OID is the SHA-256 that consuming apps verify after download.

License & attribution

MIT. This is a derivative conversion of:

Mel-Band RoFormer implementation / no-STFT export variant — © ZFTurbo (Music-Source-Separation-Training, MSS_ONNX_TensorRT) building on lucidrains' implementation of the ByteDance paper (arXiv:2310.01809), MIT. https://github.com/ZFTurbo/Music-Source-Separation-Training
Conv-based STFT/iSTFT modules — Mixxx Development Team demucs fork (GSoC 2025, Anmol Mishra), MIT. https://github.com/mixxxdj/demucs

Conversion script: tools/model-conversion/convert_melroformer.py in the SonicDemixer repository.

smank/mel-band-roformer-vocals-onnx

Mel-Band RoFormer (vocals) — ONNX conversion

License & attribution

smank/
mel-band-roformer-vocals-onnx