Demucs Windows

Because of that, a stable, reproducible path on Windows is:

  • Python 3.10
  • PyTorch 2.8.0 + cu126
  • TorchAudio 2.8.0 + cu126
  • TorchVision 0.23.0 + cu126
  • Demucs 4.0.1

Create and activate a virtual environment

py -3.10 -m venv .venv  
.\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip

Install the matching CUDA 12.6 builds from the official PyTorch wheel index:

pip install torch==2.8.0+cu126 torchaudio==2.8.0+cu126 torchvision==0.23.0+cu126 --index-url https://download.pytorch.org/whl/cu126

Install Demucs and helpers

pip install demucs==4.0.1 soundfile

Verify the environment

python -c "import torch, torchaudio; print(torch.__version__); print(torchaudio.__version__); print(torch.cuda.is_available())"

Expected result:
2.8.0+cu126

Usage

demucs -n htdemucs --two-stems=vocals "song.mp3"

The song will be located in

./separated/htdemucs/<song>

WhisperX

pip install whisperx

Fix Hugging Face issues

$env:HF_HOME="$HOME\.cache\huggingface"  
$env:HF_TOKEN_PATH="$env:HF_HOME\token"  
New-Item -ItemType Directory -Force -Path $env:HF_HOME | Out-Null
  • make the required files for HF

Available models
tiny.en, tiny, base.en, base, small.en, small, medium.en, medium, large-v1, large-v2, large-v3, large, distil-large-v2, distil-medium.en, distil-small.en, distil-large-v3, distil-large-v3.5, large-v3-turbo, turbo


CUDA (RTX 4070)

  • larger batch size results in faster speed (but likely more VRAM usage)
  • large-v2 is the best balance between accuracy and speed