silverdaw/mel-band-roformer-vocals-onnx

Silverdaw Vocal Quality Pack — Mel-Band RoFormer (ONNX)

A host-STFT ONNX export of a Mel-Band RoFormer vocal-separation model, used by Silverdaw as its optional higher-quality vocals model. Silverdaw produces the isolated vocal stem with this model; the instrumental side is reconstructed as the residual (mix − vocals).

The torch.stft / torch.istft are stripped out of the graph: the ONNX core consumes a precomputed STFT and returns per-bin complex masks, and the host application runs the STFT, applies the masks, and runs the iSTFT.

This repository is a re-host of musetric/vocal-separation-roformer-onnx so Silverdaw can serve the file from a stable, self-owned location. The weights and the export are unchanged (byte-identical; same SHA-256).

Files

File	Size	Notes
syhft_core_folded_fp16_webgpu.onnx	~5 MB	The neural-core graph.
syhft_core_folded_fp16_webgpu.onnx.data	~707 MB	External fp16 weights. Must sit beside the .onnx — the graph references it by exact filename; ONNX Runtime loads it automatically.

I/O contract (host pipeline)

• STFT — n_fft = 2048, hop = 441, 44.1 kHz, periodic Hann window, reflect ("centre") padding by n_fft/2. One chunk = hop × (frames − 1) samples (≈ 11 s, frames = 1101).
• Tensor — [1, 2050, 1101, 2] = (batch, (n_fft/2 + 1) × channels, frames, complex), with packed bin index 2 × freq + channel.
• Mask — the model returns a complex mask of the same shape; the host applies it by complex multiplication onto the STFT, then runs an envelope-normalised iSTFT and a chunk-level overlap-add. Track-level recombination is a Hamming overlap-add (8 s step over the ~11 s window) with peak normalisation.

Provenance & license

License: MIT.

• Weights — © Kimberley Jensen / SYH99999 (SYH99999/MelBandRoformerBigSYHFTV1Fast; "Kim Vocal 2" / SYHFT lineage).
• ONNX export — © musetric (musetric/vocal-separation-roformer-onnx).
• Architecture — Mel-Band RoFormer (Wang, Lu, Won; arXiv:2310.01809) / lucidrains/BS-RoFormer (MIT).

Note on training data: the upstream training-data provenance is undocumented; the weights are distributed by their author under MIT. This re-host treats them as MIT-licensed accordingly, the same posture taken by other openly-distributed separation models.

Intended use

Designed for offline (non-real-time) vocal separation inside a desktop DAW. It is not a real-time audio-callback model — run it on a worker/background thread.