HTDemucs → Core ML

Convert Meta's Hybrid Transformer Demucs (HTDemucs) into a Core ML .mlpackage you can drop into a macOS or iOS app and run with MLModel.

The hard part of converting HTDemucs to Core ML is not the network itself — it is the STFT/ISTFT and the multi-head attention around it. This repo contains a single-file converter (convert.py, ~600 LoC) that solves the three blockers you hit otherwise:

1. Core ML doesn't support complex64 → real-valued STFT/ISTFT.
2. coremltools can't trace nn.MultiheadAttention → manual decomposition.
3. Core ML's 1D scatter_add is fragile → pre-computed OLA index buffer.

The result is a stand-alone .mlpackage that takes raw stereo audio and outputs four stems (vocals, drums, bass, other) at 44.1 kHz.

Why another conversion?

There is one prior public Core ML conversion of HTDemucs by john-rocky/CoreML-Models at 7.8 s segments / 80 MB. This repo offers:

• Longer segments (10 s by default) → fewer overlap-add boundaries on long files.
• CLI flags for segment length, FP16 quantization, compute-unit selection.
• Source order reordered to [vocals, drums, bass, other] (DJ/UI convention).
• Documented workarounds so you can reproduce or adapt the pipeline for other audio models (Spleeter, OpenUnmix, MDX-Net).

Quick start

git clone https://github.com/dexxdean/htdemucs-coreml
cd htdemucs-coreml
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt

# default: 10 s segments, FP32, ~400 MB
python convert.py

# half size, ~200 MB, slight numerical drift but inaudible in practice
python convert.py --fp16

# shorter segments if you want lower latency / smaller buffers
python convert.py --segment 7

The output is HTDemucs_CoreML.mlpackage (or HTDemucs_CoreML_FP16.mlpackage).

Usage in Swift

import CoreML
import AVFoundation

// 1. Load the model.
let url = Bundle.main.url(forResource: "HTDemucs_CoreML", withExtension: "mlpackage")!
let config = MLModelConfiguration()
config.computeUnits = .cpuAndGPU   // see "Compute units" below
let model = try MLModel(contentsOf: url, configuration: config)

// 2. Feed a (1, 2, 441000) Float32 MLMultiArray named "audio".
//    Output is a (1, 4, 2, 441000) Float32 array named "sources",
//    in the order [vocals, drums, bass, other].

A more complete example with chunking, overlap-add, and AVAudioEngine playback is in examples/swift/StemSeparator.swift.

Model I/O

Compute units

HTDemucs is not stable on the Apple Neural Engine. Use .cpuAndGPU (the default baked into the model). Forcing .all or .cpuAndNeuralEngine may produce silent garbage on some shapes — the validation step in convert.py will warn if numerical drift is large.

File sizes

How it works

See CONVERSION_NOTES.md for the technical deep-dive on the three workarounds (real STFT, manual MHA, OLA scatter).

License & attribution

This repo is MIT-licensed — see LICENSE.

The converted model derives from facebookresearch/demucs © Meta Platforms, Inc., MIT-licensed. The pre-trained HTDemucs weights are downloaded by the demucs Python package at conversion time from Meta's official release. You must comply with Demucs' MIT license when redistributing the resulting .mlpackage — keep the attribution in ATTRIBUTION.md alongside the model and in your app's about/legal screen.

This project is not affiliated with Apple, Meta, or Demucs. The package name HTDemucs_CoreML.mlpackage was chosen to avoid any confusion with Apple-internal model names (e.g., MusicSourceSeparation).

Citation

If you use this in academic work, please cite the original Demucs papers:

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