HT-Demucs 4-stem — ONNX (fp16 weights)
Meta's HT-Demucs — drums, bass, other, vocals — as one ONNX graph, mirrored unchanged so that Chopmatic does not depend on anyone else keeping their copy up.
This is a mirror, not our export. The file is byte-for-byte
htdemucs_fp16weights.onnx from
StemSplitio/htdemucs-onnx
at commit d54ed9eb60e258ea82131c6ee14578628816456a. Their card has the
export details and a reference infer.py.
| file | size | SHA-256 |
|---|---|---|
htdemucs_fp16weights.onnx |
165,612,636 bytes | d05c269d0178d2a72ad484b10b11dd370193fc923201c3b27a99f848745db70a |
Shape
- input
mix:[1, 2, 343980]— 7.8 s of 44.1 kHz stereo - output
stems:[1, 4, 2, 343980]— sources in the order drums, bass, other, vocals
Longer audio is split into overlapping segments and crossfaded; the
reference infer.py overlaps by a quarter with a linear ramp.
Measured here
On onnxruntime 1.20, CPU provider, an M4 MacBook Air (16 GB):
sum(4 stems) vs mix, relative RMS |
0.016 |
| Four-minute stereo song, end to end | 3.6–4.7 minutes |
| Peak resident memory, whole app | 4.35 GB |
Licence
MIT. HT-Demucs is © Meta Platforms, Inc. and affiliates — see LICENSE.
The ONNX export is StemSplitio's, also published under MIT.
The training data is stated here rather than glossed. Meta's README
lists htdemucs as trained on MUSDB18-HQ plus 800 further songs, and
MUSDB18's tracks are licensed for non-commercial research. Whether that term
reaches weights trained on them is unresolved; the code and the weights are
released under MIT, and nothing upstream says otherwise. Anyone shipping this
commercially should make that call knowingly.