HT-Demucs (single-file 4-stem) — ONNX

The first ONNX export of the standard htdemucs (non-FT) model on the Hugging Face Hub. Runs in onnxruntime on CPU out of the box, and on CoreML / CUDA / DirectML with a one-line provider change. No PyTorch required at inference.

This repo is the single-file companion to StemSplitio/htdemucs-ft-onnx. You get all 4 stems out of one 316 MB .onnx file (htdemucs.onnx), or 166 MB if you grab the fp16weights variant. The FT bag is higher quality; this single model is ~30% faster and uses 1 session instead of 4.

---

TL;DR

# 316 MB fp32 model:
pip install onnxruntime numpy soundfile
python infer.py your-song.mp3 ./out/ --write-all-stems
# writes ./out/{drums,bass,other,vocals}.wav at 44.1 kHz stereo

# 166 MB fp16weights variant (same runtime cost):
python infer.py your-song.mp3 ./out/ --small --write-all-stems

The repo contains:

• htdemucs.onnx — 316 MB, opset 17, parity-verified vs PyTorch fp32.
• htdemucs_fp16weights.onnx — 166 MB, fp16-stored weights, same runtime memory / latency.
• infer.py — pure-numpy reference inference (~200 lines, no torch).
• requirements.txt — three small packages, no PyTorch.

---

Quality

The official htdemucs model is the precursor to htdemucs_ft — same architecture, single set of weights instead of 4 specialist sub-models. On MUSDB18-HQ:

Parity vs PyTorch fp32 (random input, 7.8 s segment):

• htdemucs.onnx max abs diff: 6.62 × 10⁻⁴
• htdemucs_fp16weights.onnx max abs diff (vs fp32 weights): 4.6 × 10⁻⁵

Both well within the 1e-3 publish threshold.

---

Performance

Single 7.8 s segment, Apple M4 Pro CPU:

CUDA / DirectML / CoreML EPs are typically ≥ 5× faster on real GPUs.

---

Quick start

Python

import soundfile as sf
import infer

audio, sr = sf.read("your-song.mp3", dtype="float32", always_2d=True)
stems = infer.separate(audio.T, sr,
                       model_path=infer.DEFAULT_MODEL,
                       providers=["CPUExecutionProvider"])
for stem, arr in stems.items():
    sf.write(f"{stem}.wav", arr.T, sr)

CLI

python infer.py your-song.mp3 ./out/ --write-all-stems
python infer.py your-song.mp3 ./out/ --providers coreml   # macOS arm64
python infer.py your-song.mp3 ./out/ --providers cuda     # Linux + NVIDIA
python infer.py your-song.mp3 ./out/ --providers dml      # Windows + DX12
python infer.py your-song.mp3 ./out/ --small              # 166 MB variant

Mobile / Web (after pip install onnxruntime-mobile or onnxruntime-web)

// iOS / Swift
import onnxruntime_objc
let opts = try ORTSessionOptions()
try opts.appendCoreMLExecutionProvider(with: ORTCoreMLExecutionProviderOptions())
let session = try ORTSession(env: env,
    modelPath: Bundle.main.path(forResource: "htdemucs", ofType: "onnx")!,
    sessionOptions: opts)

// Browser / web
import * as ort from "onnxruntime-web";
const sess = await ort.InferenceSession.create("htdemucs_fp16weights.onnx", {
  executionProviders: ["wasm"],
});
const t = new ort.Tensor("float32", audioBuffer, [1, 2, 343980]);
const out = await sess.run({ mix: t });   // out.stems is (1, 4, 2, 343980)

For a turnkey browser demo with file-picker + chunked overlap-add, see demucs-onnx browser-demo.

---

Input / output spec

For longer audio, chunk with overlap-add — see infer.py::separate for a working 60-line implementation.

---

Tooling — demucs-onnx Python package

This model can be run (and re-exported from PyTorch) via the open-source demucs-onnx Python package on PyPI. It auto-downloads from this repo on first use, so you don't have to clone or wrangle file paths.

pip install demucs-onnx

# Single-file 4-stem flavor (this repo):
demucs-onnx separate song.mp3 stems/ --model htdemucs

# Python API:
python -c "from demucs_onnx import separate; \
  print(separate('song.mp3', model='htdemucs').keys())"

To re-export your own fine-tune:

pip install 'demucs-onnx[export]'
demucs-onnx export htdemucs out/htdemucs.onnx

---

How it was built

The export pipeline lives in the open-source demucs-onnx package at demucs_onnx/export/. It applies four patches to make torch.onnx.export work on htdemucs:

1. Complex-typed torch.stft outputs → Conv1d with sin/cos kernels.
2. model.segment fractions.Fraction → plain float.
3. random.randrange in transformer pos-embedding → hardcoded shift=0.
4. aten::_native_multi_head_attention (no ONNX symbolic) → drop-in nn.MultiheadAttention.forward built from Linear/bmm/softmax.

These are the four blockers every previous community attempt at "demucs onnx" stalled on. See the README of the demucs-onnx package for the full write-up with code references.

---

Related work

Sibling ONNX repos from the same export pipeline:

Full benchmark across every popular open-source separator: StemSplitio/stem-separation-benchmark-2026.

---

Skip the infrastructure — use the StemSplit API

Don't want to bundle a 316 MB model in your app, manage a GPU pool, or write overlap-add chunking? Use the StemSplit API instead — same model under the hood, hosted for you, with credits and a dashboard.

• 🌐 stemsplit.io
• 📘 Developer docs
• 🔌 API reference

Or use the no-code tools that ship the same model family:

• 🎤 Vocal Remover
• 🎶 Karaoke Maker
• 🎙️ Acapella Maker
• 📺 YouTube Stem Splitter

---

License & attribution

This repo is MIT-licensed, matching the original HT-Demucs.

@inproceedings{rouard2023hybrid,
  title     = {Hybrid Transformers for Music Source Separation},
  author    = {Rouard, Simon and Massa, Francisco and D{\'e}fossez, Alexandre},
  booktitle = {ICASSP},
  year      = {2023}
}

• Original PyTorch model: facebookresearch/demucs
• ONNX export, parity verification, and packaging by StemSplit
• Search keywords: htdemucs onnx, demucs onnx single file, demucs ios, demucs android, music source separation onnx, stem separation mobile.