DrumSep v2 — 10 s window, fp16-weights ONNX (browser build)
A browser-friendly re-export of DrumSep v2. It splits a drum stem into kick, snare, cymbals (hats + crash/ride) and toms. It is used by the in-browser Stem Studio from NoDAW (Liminal / CoProducer), running in onnxruntime-web.
What this is (and is not)
- Not new weights. The weights are those of
gridshiftstudio/drumsep-onnx(v2), a fine-tune ofinagoy/drumsep, an HDemucs-style drum separator. All model credit goes to those authors, and to the HDemucs / Demucs authors (Alexandre Défossez et al., Meta AI Research; MIT). - Rebuild. The upstream ONNX has a fixed 40 s window (4–6 GB RAM, not usable in a tab). The HDemucs was rebuilt in PyTorch from the ONNX initialisers and re-exported with a 10 s window (
[1,2,441000]at 44.1 kHz). Weights are stored in fp16 with a Cast back to fp32; activations stay fp32. File size 168 MB (vs 335 MB fp32). - Interface: inputs
mix [1,2,441000]andmag [1,4,2048,431]; outputsfreq_out,time_out. The STFT / iSTFT are done outside the graph (in JS in Stem Studio), so you must supply the complex-as-channels magnitude input yourself. - Parity (measured on the rebuild): vs the upstream 40 s ONNX, 66.6 dB SNR (float noise). 10 s ONNX vs PyTorch rebuild, 86.5 dB. Browser wasm vs Python onnxruntime on the same window, 83–91 dB.
- Quality is that of upstream. Third-party reports, which I have not re-measured: kick ≈ 21 dB SDR, snare ≈ 12–13 dB, cymbals ≈ 5 dB, and toms are weak. Chunked 10 s inference differs from the single 40 s pass.
Licence and provenance — please read
- Weights and code of both upstream repos are MIT, and this file is released under MIT too (see upstream for the copyright notices).
- Training-data provenance is NOT verified. The v2 model card says it was trained on the author's own one-shots plus Magenta Groove MIDI (CC BY 4.0) rendered with them. The upstream v1 training data is not documented in detail. MIT covers the weights file, not any unknown rights in training audio. Check this yourself before using the model in a commercial product.
File
drumsep_v2_w10_f16w.onnx — 168,050,967 bytes — sha256 b05779ea756870f3fe359c63ac501dd12caf3b24d66237b1404bb4dd5b32839a
Pin a specific commit when you load it (/resolve/<commit>/drumsep_v2_w10_f16w.onnx) and verify the SHA-256.
Citation
Défossez et al., Hybrid Spectrogram and Waveform Source Separation (HDemucs); Demucs: https://github.com/facebookresearch/demucs