sound-mix
Skill0 starsChanged 18 days agoMITView on GitHub
Sound for short AI videos - checks in numbers for what an agent can't hear (hum, clicks at cuts, chatter between lines, dull "cheap mic" generated voices, loudness and true peak), effects on their frames, music placement, and mastering without loudnorm. Use before anyone hears a cut, when adding effects or music, mixing dialogue, mastering to a loudness target, or when a mix sounds like a fan, clicks, pumps or "sounds cheap".
In this repository4 other files it gives its agents.
- blind-reviewSkill
- product-sizeSkill
- recurring-castSkill
- scene-setSkill
The file
We split this skill into the rules it gives an agent
6 rules. About 0 are common, 6 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈6
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 16
Run it on the mix without music: music hides hum and pauses.
- Distinctive: line 37
clicks: put every in and out point in silence, with a 30 ms fade.
- Distinctive: line 49
Every effect plays over the shot that shows its source.
- Distinctive: line 58
A dialogue comedy: no music under the lines; a 3-5 s button after the punchline into the end card.
- Distinctive: line 69
Always deliver a no-music master too, and the stems (voices, effects, music) as WAVs from 0 s, when an editor might re-cut it.
- Distinctive: line 73
No taste: whether the music fits, whether an effect lands on its frame, whether a voice suits the character.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Sound mix
The agent that cuts a video can't hear it. Generated voices come out band-limited, a room-tone bed reads as a fan, a voice changer turns a model's murmur between lines into chatter, and cuts leave clicks. A viewer hears all of it in the first seconds, and the video sounds cheap. These checks measure what can be measured, so a cut that fails never reaches a person's ears. Everything else is the owner's ear.
Check every cut before anyone hears it
uv run scripts/sound_check.py <cut_nomusic.mp4> --words words.json --end <s>
Run it on the mix without music: music hides hum and pauses. Word times come from any Whisper with word timestamps
(openai-whisper, mlx-whisper); without them the pauses check is skipped. --end is where the body ends, before an
end card: a silent card hides a hum, and a card's voice-over is the brand voice, not a generated one. Run the
loudness check again on the final master (--skip hum,clicks,voice_hf).
| Check | FAIL when | Calibrated on |
loudness | integrated loudness more than 1 LU off --target (default -14 LUFS), or true peak above -1 dBTP | social video delivery |
hum | the steady floor (the quietest tenth of 50 ms frames) is above -55 dBFS | a reel a client rejected because the room tone sounded like a fan: -33 dBFS |
clicks | sound above -35 dBFS drops more than 30 dB within 5 ms | a rejected cut with 4, all at edits; the client's published reels had none |
pauses | a pause between words isn't 15 dB under the speech, or a 20 ms burst in it comes within 6 dB of it | chatter a voice changer made of a model's murmur, a laugh under a cut, an ad-lib after the last line |
voice_hf | speech energy above 5 kHz is more than 12 dB under the whole | a generated voice whose "s" was heard as "sh": -14.4 dB; the brand's studio voice-over: -6.7 dB |
A FAIL stops the cut: fix it, or send the owner one fix-or-accept question with the time. A FLAG means the number
can't decide: hum flags a cut where over a tenth is digital silence (gaps between clips), because the floor then
reads -120 and hides any bed under the rest. Listen to it. voice_hf judges generated speech. A dark TTS voice-over
can sit under -12 on its own: check it by ear once, then run with --skip voice_hf.
When a check fails
- hum: find the bed (an ambience track, a room tone, a generated clip's own noise) and drop it. A varied location ambience (wind, birds, a distant road) low under the lines works; a constant room tone doesn't.
- clicks: put every in and out point in silence, with a 30 ms fade. Never move a cut only to turn a check green: move it to where it sounds right.
- pauses: clean each generated take with a speech enhancer (DeepFilterNet) first; raw generated audio never reaches a mix. End each voice window at the last scripted word, with a short fade.
- voice_hf: revoice the take. One voice-changer call per character (join their lines with short gaps, convert once, cut back), so the timbre holds across takes. No second enhancer pass after it: every pass costs high end.
Effects
- One sound per thing the picture shows, placed on its frame. Find each footstep, contact or crack frame by frame and place one generated hit there. A 5 s "walking" texture over a shot with one step reads as a mechanical rustle.
- A still or push-in insert gets no action sound, at most one on the cut.
- Every effect plays over the shot that shows its source. A sound cue the script names gets its shot.
- Set an effect's level by its 50 ms peak in the master (~10 dB under the voices). LUFS reads a 0.1 s thud 15-20 dB under its peak.
Music
- Library or trending tracks work. Clients rejected AI-generated music more often than not; if you generate it, ask for "no lead melody" and show raw beds first. Avoid models with non-commercial weights for paid work.
- A talking head: the bed sits under the voice, ducked hard (sidechain), so it fills every pause. A track with a big dynamic jump gets compressed flat first.
- A dialogue comedy: no music under the lines; a 3-5 s button after the punchline into the end card.
- Land the peak on the payoff: the music's peak sits on the onset of the final line's first word, both measured on the waveform. Offset = the track's peak time - the line's time in the video.
- Take a button from where the track really ends. A generated 20 s track can go silent seconds before its end.
- Measure music level in the master, not by the mix number: a limiter lifts music several dB against limited voices.
Mastering
uv run scripts/master.py <mix> <out.wav> [-14]: a linear gain plus a 4x-oversampled true-peak limiter, corrected
until it lands on target, then measured. Social video: -14 LUFS, true peak ≤ -1 dBTP. Never master with ffmpeg
loudnorm: it silently switches to a dynamic mode when a linear gain would break its peak target, and the mix pumps.
Always deliver a no-music master too, and the stems (voices, effects, music) as WAVs from 0 s, when an editor might re-cut it.
Limits
- No taste: whether the music fits, whether an effect lands on its frame, whether a voice suits the character.
humandvoice_hfread the whole body up to--end. A deliberate ambience bed failshum: skip it and judge it by ear.- The thresholds were set on a handful of 9:16 dialogue reels. Re-check them on your own approved and rejected cuts.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

