mlboydaisuke/HTDemucs-CoreML

HTDemucs — Core ML

Audio Source Separation

Split music into 4 stems: drums, bass, vocals, other. 44.1 kHz stereo, FP32.

Core ML conversion of adefossez/demucs for on-device inference on iPhone, iPad and Mac. Converted with coremltools; the packages are stateless, so all sequencing and buffering lives in your Swift code.

Task	audio to audio
Upstream	adefossez/demucs
Packages	1
Download size	75 MB
Minimum iOS	17.0
Peak RAM	~1000 MB

Files

File	Size	Compute units	SHA-256
HTDemucs_SourceSeparation_F32.mlpackage.zip	75 MB	cpuOnly	0fbb941e15a5b2fa…
Total	75 MB

compute_units is not a suggestion -- it is the configuration the conversion was verified against. Moving a package to a different compute unit can silently change the numerics (FP16 attention overflow) or crash on the GPU.

Download

hf download mlboydaisuke/coreml-zoo --include "demucs/*" --local-dir ./demucs
unzip './demucs/demucs/*.zip' -d ./demucs

Use in Swift

import CoreML

let config = MLModelConfiguration()
config.computeUnits = .cpuOnly   // as converted — see the table above

// Unzip the .mlpackage, drop it into your Xcode target and Xcode compiles it
// at build time:
let model = try HTDemucs_SourceSeparation_F32(configuration: config)

// ...or compile a downloaded .mlpackage at runtime:
let compiled = try await MLModel.compileModel(at: mlpackageURL)
let model = try MLModel(contentsOf: compiled, configuration: config)

Demo

• Sample app — sample_apps/DemucsDemo, a standalone SwiftUI project.
• Models Zoo — this model is downloadable and runnable inside the Models Zoo app on the App Store, no build required.

Conversion

• Script: convert_htdemucs.py
• Pitfalls hit during conversion (FP16 overflow, ANE buffer limits, stride handling): docs/coreml_conversion_notes.md
• Model index: CoreML-Models

License

The conversion inherits the upstream license: MIT.

Credits

• Upstream authors: adefossez/demucs, 2021
• Core ML conversion: john-rocky (Daisuke Majima)