Skip to content

konradre/video-production-skills / skills/spot-audio-assembly

spot-audio-assembly

Verified

Assembles the SOUND of a generated-video piece: voice-over takes (one voice and model per role), dubs for refused lines in the character's cloned voice, off-screen lines, sfx with the syllable map confirmed before the cut, music cues cut to markers with ducks, room tone, captions, the stem, and the read of the delivered audio. Use when a line must be voiced or re-voiced, a line was refused, sfx or music must be sourced or cut, captions must be built, a take's audio carries an artifact, or the delivered audio must be verified. Triggers — "generate the VO", "same voice as before", "dub the line", "the off-screen voice", "the sting", "the syllable map", "the music is too loud", "duck the music", "room tone", "captions", "is the vo on time", "musical artifact at the start". Not for placing lines on the timeline or the beat gate — use video-edit-edl. Not for the master render or the delivery encode — use video-finish-qc. Not for LUFS mastering of a finished mix — use mastering-audio.

FileKindSize
SKILL.mdskill md22.0 KB
references/CAPTIONS.mdmarkdown5.4 KB
references/MIX-AND-QC.mdmarkdown13.4 KB
references/SFX-AND-MUSIC.mdmarkdown13.2 KB
references/VO-PIPELINE.mdmarkdown6.6 KB
scripts/audio_head_scan.pyexecutablescript3.9 KB
scripts/beat_grid.pyexecutablescript30.1 KB
scripts/build_captions.pyexecutablescript8.8 KB
scripts/build_vo_stem.pyexecutablescript5.7 KB
scripts/gen_vo.pyexecutablescript7.0 KB
scripts/music_loop.pyscript7.6 KB
scripts/music_take_read.pyscript28.2 KB
scripts/qc_vo_placement.pyexecutablescript4.2 KB
scripts/roomtone_synth.pyexecutablescript3.4 KB
scripts/sfx_onsets.pyexecutablescript7.2 KB
scripts/sound_table.pyexecutablescript33.5 KB
scripts/vo_word_times.pyexecutablescript6.4 KB
scripts/voice_ab_reel.pyscript3.9 KB

references/SFX-AND-MUSIC.md

View on GitHub
# SFX and music — sourcing, the syllable map, cues, ducks

## SFX sourcing ladder

1. **The take's own sound** — the knock, the thud, the tear, the shout: cut from the take at the frame
   and placed as an sfx where the picture needs it (`video-edit-edl` § carried sound), the native muted
   where it played. Free, and it is the room.
2. **A library sound** — Freesound through a domain-scoped search plus the constructed HQ preview URL
   (the site's search pages are a JS login shell); read the licence line per sound: CC0 = no credit,
   CC BY = a credit line in the post copy, which is the operator's send-time call.
3. **Generated sound** — ElevenLabs sound generation as the paid fallback (a fixed credit cost per
   generation → cost line + GO).
4. "Something zanier" from the operator = a sourcing round, not a filter tweak.

## Reel → syllable map → cut

A pick from a reel is confirmed as a **syllable map before cutting**: which onsets, in which order, the
bitten-off one excluded — the map is read back to the operator and confirmed before the cut.
`sfx_onsets.py` lists every onset by ordinal and time so the operator's ordinal ("the second last one
that starts at 0:43") is a sample-exact `--cut a:b`. Cuts carry 5 ms fades; every sfx is 48 kHz 24-bit
WAV.

## Placement rules

- **A sound is placed by its SYNC POINT on the visual event it belongs to, never by a guessed offset** (mg-styles-15
  `lib/audio/README.md` § 1): a one-shot or an impact syncs on its transient, a whoosh on its pass-by peak (its loudest
  moment), a riser, a reverse cymbal or a swell on its END — the hit it drives into. A riser placed by its start lands early
  by however long its build runs. `scripts/sfx_onsets.py <file> --sync transient|peak|end [--src-in --dur] --event <T>`
  reads the point (a one-shot trimmed tight to its hit included) and prints the `at` that puts it on T, the EDL's `sync`
  block and the file's `derived` stamp; a plan sfx with a `sync` and no `at` is placed by it (`edl_build.py`), and
  `edl_check.py` fails a sound whose sync point has left its event — a re-timed marker, an insert.
- **Levels by construction**: bring every effect to one common level before it is placed (the donor normalises each to
  ≈ −18 dB short-term, peak ≤ −1 dBFS, which sits it about 2 dB over its bed), so `vol` states intent only — about −6 dB
  for a subtle sound, +2 for the hero hit — never a repair for a file that arrived at another level.
- **VO never overlaps an sfx** — stagger; a couplet's halves sit either side of the sound.
- **The audio signature sits on the END CARD** and its hit lands on the card's burst: the file starts
  a hair after the card cut (`card + 0.02`) so that the hit inside the file (at +T0, e.g. 0.83 s) meets
  the card's burst frame; the card's burst frame is what moves to meet it (`designed-elements`), found by
  PSNR and verified by region brightness.
- A native artifact — a tiny click where a shot cuts, a distorted onset, a generated drone — is muted by
  `native_audio_from/to`, and the hole is filled with **synthesized room tone** (`roomtone_synth.py`):
  a pasted slice carries the take's artifacts, and a hard mute reads as a hole.
- Designed-sound extras (the name-pop chimes on a product display) are **opt-in, default off**.
- **A creator-style spot has no music under the open** — room tone carries the hook; a cue may enter under the demo,
  ducked under every line; the disclosure line is never ducked over or covered by an sfx; no sign-off signature (the
  sub-genre carries none — `ad-spot-preprod` UGC-GRAMMAR.md § The beats, MIX-AND-QC.md § The phone-mic register).

## Sound direction — silences, suck-outs, impacts, the low end

lemo-opuscar judges a film by its sound first (`DIRECTOR.md` §§ 6, 11; `styles/satellite-briefing/STYLE.md` § Mix; the
suck-outs in its `demo/mix.py` files). These hold for every genre unless a line says otherwise.

- **Two real silences before the turn — films and music videos** (a 15 s ad already carries its one
  silence, before the hit). At least two silences or near-silences before the turn or the emotional peak, and the first
  sound after one is among the most important sounds in the piece. A suck-out is its own rule, never one of the two. The
  EDL names the turn's marker in `qc.turn`.
- **A suck-out before each designed impact**: 0.25–0.6 s in which the music and the foley drop 25–30 dB, reached by a
  50–80 ms ramp and released ON the impact (lemo's windows: 0.27–0.63 s at −26 to −30 dB, ramps of 50 and 80 ms). The
  music's half is a duck carrying its own ramps, `[from, to, 0.03, 0.06, 0]` — −30 dB, a 60 ms ramp in, full again AT `to`,
  the hit (`video-edit-edl` EDL-CONTRACT.md § audio). The foley's half is placement: an sfx ends at the suck-out's start
  (its `dur`), a take's own sound at `native_audio_to` or a cut on the hit, and a riser or sweep that drives INTO the hit is
  synced on it and stays. `edl_check.py` warns on any sfx, take sound or line that plays inside a suck-out.
- **The two or three designed impacts are the loudest moments of the mix.** The master is a static sum, so nothing
  compresses them; the read is the delivered file's loudest 400 ms windows and what the EDL places there.
- **The low end stays lean**: 20–60 Hz about 15 dB under the full band, music and foley high-passed around 36–44 Hz (lemo's
  mixes: 35–46 Hz). A cue that arrives heavier is filtered before it is placed.
- **Every row of the sound table is in the mix, and the EDL's audio block is the table.** On the delivered file: each sfx at
  its `at`, read around its sync point when it has one (`qc_deliverable.py`'s sfx row), each sync point on its event
  (`edl_check.py`), and each music cue's presence across its span, each suck-out's depth, the silences before the turn, the
  loudest moments and the low end, plus the loudness steps, the entries after a near-silence, the loudness range and what
  the delivery limiter took inside each designed impact (`scripts/sound_table.py` — INFO rows only until the reads are
  measured on two real deliverables; MIX-AND-QC.md § Loudness steps and entries).

## Music

- **The AceDataCloud Suno API is the DEFAULT music route** (house rule, 2026-09-12). `POST
  https://api.acedata.cloud/suno/audios`, `Authorization: Bearer $ACEDATACLOUD_API_TOKEN` sourced from the
  env file that holds it and never placed on a command line. Model **`chirp-v5-5`** — the vendor's own
  client ships an `.env.example` pinning `chirp-v4-5`, so never copy that file into place. `instrumental: true` for a bed; custom mode
  (`custom: true` with `lyric`, `title`, `style`, `vocal_gender`) when the cue is written rather than
  described. **Always async:** the call returns a task id; poll `POST /suno/tasks` with `{"id": "<task id>"}`
  every 3–5 s, ≈ 2 min 20 s to finish. **One call ≈ 0.56 credits ≈ $0.08 and returns TWO takes**
  (≈ $0.04 a take) — a billed generation, so it goes behind the cost line and the operator's GO like every
  other venue. `duration` (10–360 s) applies only to a custom-mode call and is a target, not a cap: an
  unguided generation returned 5.5 minutes. Downloads are **Opus inside `.m4a`** — Resolve rejects them, so
  transcode before anything reads them: `ffmpeg -i <cue>.m4a -ar 48000 -c:a pcm_s24le <cue>.wav`. The vendor
  exposes no balance endpoint — track spend from the receipts. No music client ships here: the call above is
  the whole contract, and the vendor publishes its own skill for it.
- **A music take is measured before it is trusted** (nicceasy/3d-product `references/sound-design.md` §§ 5, 8, 10, MIT).
  Generators ignore the requested key and tempo, so before a sweetener is tuned or a cue is cut,
  `scripts/music_take_read.py <take>` reads:
  - the key, with its confidence (r > 0.7 solid) and a near-tie WARN when the runner-up sits within 0.05 (a relative key
    shares the notes; the tonic is a listening call);
  - the tuning in cents from A440, and its spread (≤ ~3–5 c is one tuning);
  - the tempo drift between the first and last third (≤ ~2–3 %), with no verdict when the onsets are weakly periodic (a
    piano ballad);
  - the tail: the commonest fault sits in the decay — energy above 6 kHz rising, the 100–500 Hz channel correlation
    falling or going negative, the mid band turning to noise.

  A surprising number is checked by a second method before it is acted on: the tempo with `beat_grid.py`, the key by
  ear. **Fixes, cheapest first:** regenerate with clean-ending wording → a tail-only cleanup crossfaded in over ~1 s, with
  everything before it sample-identical → resynthesise the final chord's partials. An AI listener is triage only (its
  passes disagreed on up to ~4 in 10, and it missed a tail the user called crunchy). Two generator traps: a distilled
  ("turbo") model ignored the brief's negative constraints (vocals and keys on an instrumental brief), and some models
  end a piece early, so ask for a longer piece and keep its natural ending. **Whose numbers:** upstream's are the key
  confidence (r > 0.7), the tuning spread (≤ ~3–5 c), the drift (≤ ~2–3 %) and the tail's direction. Ours, set on
  synthetic fixtures and two real takes and uncalibrated, are the near-tie margin (0.05), the drift gate (onset r 0.3),
  the body correlation a decorrelation WARN needs (0.8), the level wobble (3 dB) and the pitch wobble (8 c). All are
  defaults until measured on our takes. On two real Suno piano takes the drift read declined a verdict (onset periodicity
  r 0.10 and 0.03), and `beat_grid.py`, run on one of them, returned BEAT-MAP with a 481 ms residual, so its beats were no
  reliable cross-check there either. On two steady four-on-the-floor Suno takes asked at 124 it gave its first real verdict
  (2026-10-26): −0.05 % and +0.05 % at 125.0 and 126.0 BPM, r 0.68 and 0.75, where `beat_grid.py`'s kicks agree (about
  0.02 % over 90 s). That is the "no drift" side only: catching real drift is still proven on the synthetic ramp alone.
  The onsets do not settle the tempo's OCTAVE. A clap on 2 and 4 made the strongest period two beats, so those takes once
  read 62.5 and 63. A 120 BPM prior now picks among half, same and double (`octave_pick`), and the same prior lifts a
  70 BPM kick with quiet offbeat hats to 140. So whenever the comb without the prior picks another octave, the tempo line
  names it: "125.0 → 124.9 BPM (or 62.5 — the octave is a call)". `beat_grid.py`'s x0.5/x1/x2 rows or the ear decide
  the call. A clap on 2 and 4 at 170 reads 85, unflagged. The drift is a ratio, so it holds in either octave. The
  tail also reads its level wobble around the decay's trend
  (≥ 3 dB) and, in a tonal tail only, its pitch wobble frame to frame (≥ 8 c); every spectrum adds the channels' powers,
  so a take in opposite polarity never cancels a read.
- **Every other music route is a FALLBACK and needs the operator's permission first** — the operator's own
  candidates under `audio/music/candidates/`, or a library bed with its licence line read. Name which route
  and why, then wait. A vendor outage or a refused cue is a reason to ask, never a reason to switch quietly.
- The reseller licence question for a paying-client deliverable is the operator's call, raised once.
- **A cue is cut to the marker it ends on** (the before cue to the HIT) from the head of the candidate;
  the file name carries the length. **A cue shorter than its span is looped on its own beat grid by whole bars, the
  take's ending kept, before a longer cue is priced** — `scripts/music_loop.py` plays the take to a beat A, jumps back a
  bar multiple to a beat B on the same bar phase, crossfades the join (40 ms, equal power) and plays through the take's
  own ending, whose hit then lands exactly the jump later; the operator listens across the join before it enters a build.
  With no `--bpm` the grid comes from `scripts/beat_grid.py`, which CHECKS it: a least-squares line through the tracked
  beats (a tracker's tempo number was 2 % off; its beat times were right), accepted only when the beats sit within ±15 ms
  of it and the kicks drift under 10 ms across the take, its phase taken from the kicks (a tracker locked onto off-beat hats
  put a synthetic grid half a beat off — measured 2026-09-29); 2x or 0.5x only when the kicks say so. A BEAT-MAP verdict
  (the tempo moves) loops on the tracked beats, and an irregular interval is named — a skipped beat breaks the bar phase
  (a 38.8 s take under a 57.9 s spot, zero spend, 2026-09-15). A cue neither looped nor long enough just ENDS mid-shot —
  `edl_check` fails it.
- **The earlier accepted cue wins** — a cue the operator approved is the campaign's cue for that beat.
- A cue slams in ON the reveal marker (`fade_in: 0`); the bed under the turntable and card runs to the
  end; the hit shot and the line after it carry only native sound.
- **Ducks** under speech, per cue: `[[from, to, gain]]` timeline seconds with 0.25 s linear ramps,
  rendered as a per-frame volume expression by the finisher — 0.25 under a disclaimer, 0.5 under a dub or
  a shout, 0.32 under the closer; the music never fights the VO. A duck may carry its own ramp, `[from, to, gain, ramp]`,
  or two, `[from, to, gain, ramp_in, ramp_out]` — a suck-out's `ramp_out` is 0 (§ Sound direction).
- A cue spanning an insert keeps playing under it (`edl_insert.py` moves only its end).
slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack."Stars: 198kUpdated Claude CodeCodexVerifiedInstall
ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt.Stars: 196kUpdated Claude CodeCodexVerifiedInstall