Mel-Band RoFormer (vocals) — GGUF

GGUF conversion of the Mel-Band RoFormer vocal-separation model for use with CrispASR's --separate task — no Python, no PyTorch at runtime, just the CrispASR binary and this file.

Separates a music mix into vocals and other (instrumental) stems.

Usage

crispasr --separate -m mel-band-roformer-vocals-f16.gguf -f song.flac
# -> song_vocals.wav, song_other.wav  (stereo, 44.1 kHz)

Provenance & license

• Weights: KimberleyJSN/melbandroformer (Kimberley Jensen), released under the MIT license.
• Architecture: Mel-Band RoFormer, from the MIT-licensed lucidrains/BS-RoFormer implementation (Phil Wang). Paper: Mel-Band RoFormer for Music Source Separation.
• This GGUF: converted with CrispASR's models/convert-mel-band-roformer-to-gguf.py. MIT.

The CrispASR C++ port was validated stage-by-stage against the PyTorch reference at cosine = 1.000000 (STFT, band split, all 6 axial RoFormer layers, mask estimator, iSTFT), with the reconstructed waveform bit-exact.

Files

Provenance and EU AI Act Art. 53 note

• Upstream model: KimberleyJSN/melbandroformer — published by KimberleyJSN.
• Upstream licence: mit. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository. No training-content summary was found on the upstream model card at the time of writing; that documentation gap is upstream's and is not filled here.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.