StemStudio models
ONNX exports of Meta's Demucs v4 (HT-Demucs) used by the StemStudio app for on-device stem separation. Weights are Meta's released checkpoints (MIT licence); only the format changed.
The graphs take the STFT outside the network (ONNX has no istft):
in: mix [1, 2, L] spec [1, 4, 2048, F]
out: wave [1, S, 2, L] spec_out [1, S, 4, 2048, F]
| File | Model | Segment | Notes |
|---|---|---|---|
htdemucs_6s.onnx |
6 stems | 7.8 s | desktop |
htdemucs_6s_s3.9.onnx |
6 stems | 3.9 s | phones (half the peak RAM, −0.3 dB SDR) |
htdemucs.onnx |
4 stems | 7.8 s | desktop |
htdemucs_s3.9.onnx |
4 stems | 3.9 s | phones |
htdemucs_ft_{drums,bass,other,vocals}.onnx |
fine-tuned bag | 7.8 s | desktop, Pro |
Weights are stored as fp16 and cast to fp32 at load. catalog.json lists
sizes and SHA-256 hashes; the app verifies every download against it.
Produced by tools/export_onnx.py in the StemStudio repository.
Citation: Rouard, Massa, Défossez, Hybrid Transformers for Music Source Separation, ICASSP 2023.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support