demusic

demusic removes the background music that AI video models (Seedance, Kling, Veo, Hailuo, Wan, ...) bake into generated clips, and keeps the dialogue, sound effects and ambience. It finds where the music is and only replaces the audio there, so the rest of the clip is untouched. Local, free, CLI + MCP server.

Read in Chinese · How detection works

Why

Native-audio video models often add a music bed nobody asked for: a few seconds of score at the start of a shot, a pad under the dialogue. Cut ten such clips together and you get ten pieces of music fighting each other, and you can't lay your own score over it.

The usual fix - run a vocal remover and keep only the vocals - throws away footsteps, doors, rain and room tone along with the music, and it is wrong on every clip that had no music to begin with. demusic is what we run on every clip of our AI dramas at Dawnwood:

1. Separate the audio into stems (Demucs htdemucs_6s, locally or on your fal key).
2. Is it loud enough to be a bed? The melodic stems must be audible (> -50 dBFS).
3. Is it actually music? An AudioSet classifier must hear music on the accompaniment for 2+ seconds. Engine hum, gunfire and room tone pass step 2 but fail this one.
4. Strip only those spans: there the audio becomes dialogue + percussion/impacts, with a 0.25 s crossfade; everywhere else the original audio stays. The picture is stream-copied (no re-encode).

It is deliberately conservative. On 164 clips with no music it wrongly flagged 2; on 87 clips a person had flagged by ear it found 33. The misses are very quiet beds - lower --threshold or use --mode force when you know a clip has music. Calibration details.

Install

Requires Python >= 3.10 and ffmpeg on PATH.

pip install "demusic[demucs]"     # local separation (pulls in PyTorch; CPU is fine)
# or, no torch, separation on fal.ai with your own key:
pip install "demusic[fal]" && export FAL_KEY=...

The music classifier (~90 MB ONNX, BSD-3) downloads once to ~/.cache/demusic.

Quickstart

# where is the music?
demusic check shot_012.mp4
# shot_012.mp4: MUSIC  spans 0-2.8s  melodic -42.1 dBFS  p=0.34

# write shot_012.demusic.mp4 with the music removed
demusic strip shot_012.mp4

# a whole folder; clean copies of every file (unchanged ones copied) into clean/
demusic strip renders/*.mp4 --out-dir clean/

Exit codes: 0 ok (for check: no music), 1 check found music, 2 usage, 3 failure. demusic check in CI or a render pipeline = fail the build when a clip has a music bed.

Speed: separation is roughly real time on a laptop CPU (a 10 s clip ~10-20 s), much faster on a GPU. Detection itself is a few hundred ms.

Python

from demusic.core import process
r = process("shot.mp4", "shot.clean.mp4")          # strip
r = process("shot.mp4", strip=False)                # check only
print(r["music"], r["spans"])

AI agents (MCP)

claude mcp add demusic -- demusic mcp            # Claude Code
# or any MCP client: command "uvx", args ["--from", "demusic[demucs]", "demusic", "mcp"]

Tools: check_music (read-only) and remove_music (writes a new file, never overwrites the input). Set DEMUSIC_SEPARATOR=fal in the server env to use fal. There is also a Claude skill and AGENTS.md.

Limits

• Music that is meant to be in the scene (a band on screen, a radio) is removed too when it is detected. Skip those clips.
• Inside a stripped span, non-percussive effects (a running tap, wind) go with the music; Demucs has no separate SFX stem. Outside the spans nothing changes.
• Sung vocals land in the vocals stem and are kept.

Credit

Apache-2.0. demusic writes cleaned with demusic by Dawnwood into each output file's comment tag. If it saves you time, please keep that and mention "demusic by Dawnwood" - a request, not a condition (see NOTICE).

Made by Dawnwood, an AI drama studio. More tools we use in production: dramapack (short-drama delivery packs and QC) · all our open source.