cjm-demucs-v4 0.0.1

Inference-only fork of Demucs v4 that provides audio source separation with TorchCodec replacing torchaudio I/O.

cjm-demucs-v4

Inference-only fork of Demucs v4 (Hybrid Transformer) for Python 3.12+ with TorchCodec replacing torchaudio I/O.

What Changed

• TorchCodec migration: Replaced torchaudio.save() and torchaudio.load() with TorchCodec's AudioEncoder and AudioDecoder
• Removed audio_legacy.py: torchaudio 2.1 compatibility shim no longer needed
• Stripped training code: Removed training, evaluation, augmentation, distributed training, and experiment grid files
• Updated dependencies: torchcodec replaces torchaudio; requires torch>=2.0, python>=3.12

Installation

pip install cjm-demucs-v4

Or from source:

pip install git+https://github.com/cj-mills/cjm-demucs-v4.git

Requirements

• Python >= 3.12
• PyTorch >= 2.0
• FFmpeg (for audio reading)
• CUDA GPU recommended (CPU inference works but is ~10-20x slower)

Usage

Python API

from demucs.api import Separator, save_audio

separator = Separator(model="htdemucs", device="cuda")
origin, separated = separator.separate_audio_file("input.mp3")

# Save just the vocals
save_audio(separated["vocals"], "vocals.wav", samplerate=separator.samplerate)

CLI

demucs input.mp3 -o output_dir
demucs input.mp3 --two-stems vocals  # vocals + accompaniment only

Available Models

Model	Description
htdemucs	Hybrid Transformer Demucs v4 (default)
htdemucs_ft	Fine-tuned version (better quality, 4x slower)
htdemucs_6s	6-source variant (adds piano, guitar)
hdemucs_mmi	Hybrid Demucs v3 retrained
mdx_extra	MDX architecture with extra training data
mdx_extra_q	Quantized MDX (fastest)

All 4-stem models separate into: drums, bass, vocals, other.

License

MIT License (same as upstream Demucs).

Credits

Based on Demucs by Alexandre Defossez / Meta Platforms, Inc.