| name | extract-vocals |
| description | Use when the user wants to quickly extract vocals (or isolated instrumentals) from a song or recording. Convenience wrapper around 2-stem `demucs` mode. Fastest stem separation for the common case. |
Extract Vocals (demucs 2-stem)
Isolate vocals from a mixed audio file using demucs in 2-stem mode. Fast and focused — produces vocals.wav and accompaniment.wav. Optionally drop the accompaniment if the user only wants the vocal stem.
When to use
- The user wants vocals isolated from a song (e.g., for editing, re-recording, or remastering).
- Preparing a karaoke track — extract the accompaniment, discard vocals.
- Extracting a vocal guide from a podcast with music bed.
- Quick turnaround on a single stem (faster than 4-stem
separate-stems).
Do not use this skill for:
- Full stem separation (drums, bass, other) → use
separate-stems.
- Batch processing many files → use
separate-stems with GPU for throughput.
Inputs
- Input file — required. MP3, WAV, FLAC, M4A supported.
- Output directory — defaults to
vocals_<date>/ next to the input.
- Keep accompaniment — boolean flag. Default: keep both stems. If
false, delete the accompaniment stem and keep only vocals.wav.
- GPU flag — optional. Pass
--gpu to use CUDA; default is CPU.
Procedure
-
Check if demucs is on PATH (which demucs). If missing, point the user at install-deps and note the ~2 GB download on first run.
-
Build the command:
demucs \
--two-stems vocals \
-d <device> \
-o "<output-dir>" \
"<input-file>"
Set -d cuda for GPU, -d cpu for CPU. Default is CPU.
-
Monitor progress and capture stderr.
-
After processing:
- If
keep_accompaniment is false, delete the accompaniment stem: rm "<output-dir>/accompaniment.wav".
- Else, leave both stems in place.
-
Verify output: confirm vocals.wav is present and non-zero size. If accompaniment was kept, confirm that too.
Output
- Two WAV files (or one if accompaniment was dropped):
vocals.wav — isolated vocal track.
accompaniment.wav — (optional) instrumental/music bed, kept only if requested.
- Summary:
<input> → <output-dir> (vocals + accompaniment, 2-stem mode) or (vocals only) if accompaniment was dropped.
Notes
- 2-stem mode is faster than 4-stem; suitable for real-time turnaround on single files.
- GPU (CUDA) gives 5–10× speedup. Check
nvidia-smi for availability.
- Output is always 44.1 kHz WAV, matched to the input sample rate where possible.
- The accompaniment includes drums, bass, and all non-vocal elements mixed together.
Dependencies
Install via pipx install demucs. Requires PyTorch. Audio format support via ffmpeg.