demucs-torchcodec 0.1.0

An updated continuation of Demucs for source separation using torchcodec instead of torchaudio.

Demucs-Torchcodec

Demucs-Torchcodec is a modernized fork of the Original Demucs.

While the original repository is no longer actively maintained, this fork continues the project by replacing the deprecated torchaudio decoding backends with the high-performance torchcodec library.

Why use this fork?

• Modern Backend: Uses torchcodec for faster and more robust audio decoding.
• Dependency Lean: Removes reliance on older torchaudio versions that often conflict with modern PyTorch installations.
• Future Proof: Designed to work with Python 3.10+ and the latest PyTorch ecosystems.

---

Installation

1. Requirements

Before installing, ensure you have FFmpeg installed on your system (required by torchcodec):

• Ubuntu/Debian: sudo apt-get install ffmpeg
• macOS: brew install ffmpeg
• Windows: choco install ffmpeg (or download from ffmpeg.org)

2. Install the Package

pip install -U demucs-torchcodec

---

Usage

The entry point for this version is demucs-torchcodec.

# Basic separation (defaults to htdemucs)
demucs-torchcodec test.mp3

# Separate into 2 stems (vocals and accompaniment)
demucs-torchcodec --two-stems=vocals test.mp3

# Use the high-quality fine-tuned model
demucs-torchcodec -n htdemucs_ft test.mp3

---

Supported Models

This fork supports all standard Demucs v4 models, including:

• htdemucs: Hybrid Transformer Demucs (Default).
• htdemucs_ft: Fine-tuned version (Higher quality, slower).
• hdemucs_mmi: Hybrid Demucs v3 retrained.
• mdx_extra: Trained with extra data.

---

Credits & License

Original Author: Alexandre Défossez.

This project is licensed under the MIT License , exactly like the original Demucs. See the LICENSE file for details.

If you use this model in your research, please cite the original paper:

@inproceedings{rouard2022hybrid,
  title={Hybrid Transformers for Music Source Separation},
  author={Rouard, Simon and Massa, Francisco and D{\'e}fossez, Alexandre},
  booktitle={ICASSP 23},
  year={2023}
}