Demucs#

Use demucs to split music into stems (default: vocals, drums, bass, other).

Quick start

• Help: demucs --help
• Basic split: demucs "input.mp3"
• Output dir: demucs "input.mp3" -o "D:\\output"
• Vocals only pair: demucs "input.mp3" --two-stems vocals

Prerequisites

• Demucs available: demucs --help or absolute path like D:\demucs-agent-tool\venv\Scripts\demucs.exe --help.
• FFmpeg available: ffmpeg -version must succeed.
• Note: ffmpeg-python is only a Python wrapper and does not install ffmpeg.exe.

Common options

• Model: -n htdemucs (default) or other installed model names.
• Device: -d cuda or -d cpu.
• Chunking: --segment 8 --overlap 0.25 to reduce VRAM/RAM pressure.
• No chunking: --no-split for short audio with enough memory.
• Equivariant shifts: --shifts 1..10 (higher quality, slower).
• Parallel jobs: -j 2 (or more on strong CPUs).

Output format

• Default output is WAV.
• FLAC: demucs "input.mp3" --flac
• MP3: demucs "input.mp3" --mp3 --mp3-bitrate 320 --mp3-preset 2
• Safer clipping behavior: --clip-mode rescale (default) or --clip-mode clamp.

Naming and folder layout

• Default output path pattern: separated/<model>/<track>/<stem>.<ext>.
• Customize names: --filename "{track}/{stem}.{ext}".
• Variables supported: {track}, {trackext}, {stem}, {ext}.

Batch usage

• Multiple files in one call: demucs "a.mp3" "b.wav" "c.flac" -o "D:\\separated".
• Prefer absolute paths for automation scripts and scheduled jobs.

Windows notes

• In venv, executable is usually venv\\Scripts\\demucs.exe.
• Use quoted paths when spaces exist: demucs "D:\\My Music\\song.mp3".
• For external programs, prefer absolute executable path instead of shell activation.
• If non-ASCII paths cause tool/runtime issues, copy input to a temporary ASCII path and run Demucs there.

Troubleshooting

• demucs not found: run python -m demucs --help or call venv\\Scripts\\demucs.exe directly.
• FFmpeg is not installed / ffprobe not found: install FFmpeg and add it to PATH, then verify with ffmpeg -version.
• Out of memory: lower --segment (for example 6 or 4), set -d cpu, or process fewer tracks at once.
• Slow speed on CPU: reduce --shifts, keep chunking enabled, and tune -j.
• Install/network failures: use a stable mirror and longer timeout when running pip.