README.md

metadata

license: mit
tags:
  - mlx
  - audio
  - source-separation
  - music
  - apple-silicon
base_model:
  - facebook/demucs
library_name: mlx

MLX Demucs Weights

Converted weights for mlx-demucs, an Apple Silicon port of Meta's Demucs audio source separation models using the MLX framework.

All models achieve <0.04% relative error vs the original PyTorch weights.

Models

Model	Stems	Notes
htdemucs	drums, bass, other, vocals	Hybrid Transformer Demucs
htdemucs_ft_drums	drums, bass, other, vocals	Fine-tuned for drums
htdemucs_ft_bass	drums, bass, other, vocals	Fine-tuned for bass
htdemucs_ft_other	drums, bass, other, vocals	Fine-tuned for other
htdemucs_ft_vocals	drums, bass, other, vocals	Fine-tuned for vocals
hdemucs_mmi	drums, bass, other, vocals	Hybrid Demucs, no transformer
htdemucs_6s	drums, bass, other, vocals, guitar, piano	Experimental 6-stem model

Usage

Install mlx-demucs — weights are downloaded automatically on first use:

pip install mlx-demucs
mlx-demucs song.wav
mlx-demucs song.wav -m htdemucs_6s

Or use the Python API:

from mlx_demucs.utils.loader import load_model

model = load_model("htdemucs")       # 4-stem
model = load_model("htdemucs_ft")    # 4-stem ensemble (best quality)
model = load_model("htdemucs_6s")    # 6-stem (experimental)

Performance

~38x realtime on Apple Silicon.

License

Weights are derived from facebook/demucs and released under the MIT License. Copyright (c) Meta Platforms, Inc. and affiliates.