| # Step 1 — Demucs Setup |
|
|
| **Status: DONE** |
|
|
| ## What we did |
|
|
| Installed Demucs and ran it on a local audio file to confirm the model works before building anything around it. |
|
|
| --- |
|
|
| ## What was installed |
|
|
| ```bash |
| # CPU-only PyTorch (saves ~2GB vs CUDA build — no GPU on this machine) |
| pip install torch torchaudio --index-url https://download.pytorch.org/whl/cpu |
| |
| # Demucs model |
| pip install demucs |
| |
| # Required for saving output audio files |
| pip install torchcodec |
| ``` |
|
|
| ## How to run it manually |
|
|
| ```bash |
| source venv/bin/activate |
| demucs --two-stems=vocals -n mdx_extra "path/to/audio.mp3" |
| ``` |
|
|
| - `-n mdx_extra` — lighter model, faster on CPU |
| - `--two-stems=vocals` — splits into two files: voice and background |
|
|
| ## Output |
|
|
| Results land in: |
| ``` |
| separated/mdx_extra/<track-name>/ |
| ├── vocals.wav # isolated voice |
| └── no_vocals.wav # background / everything else |
| ``` |
|
|
| --- |
|
|
| ## Notes |
|
|
| - Processing time: ~40 seconds for a 33-second clip on ThinkPad X270 |
| - Fan noise mixed into voice is not fully removed — Demucs is a music separator, not a speech denoiser |
| - Future fix: swap Demucs for **DeepFilterNet** (better for speech + background noise) |
| - Model weights cached at `~/.cache/torch/hub/checkpoints/` — no re-download on next run |
|
|
