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/VO-PIPELINE.md

View on GitHub
# VO — voices, takes, dubs, off-screen lines, word times

## Venue and cost

- **ElevenLabs by raw REST** (`gen_vo.py`): `eleven_multilingual_v2`, `mp3_44100_128`. Billed per
  **character**, per call, with no dry run on the vendor side — `--dry-run` prints the characters each
  job would spend; the plan's `character_count / character_limit` is a free read printed before and
  after so the billed delta is on the record. The starter tier carries a monthly character cap and ten
  voice slots (instant clones yes, professional no). Speech-to-text (Scribe) bills per minute of audio.
- 🔴 **The cost line and the operator's GO come before every TTS, STT, sound-generation or clone call**,
  in the form `video-gen-cost-gate` prescribes, however small the amount.
- An existing output is SKIPPED by name: a take is never regenerated by accident; a new take is a new
  job name (`-b`, `-c`).
- Key: `ELEVENLABS_API_KEY` in the environment from the env file that holds it (`KEY=value`, no `export`, no
  quotes — a systemd `EnvironmentFile` rejects `export` and starts a keyless service). Never
  `elevenlabs auth login --with-token` with an `sk_` key (it is filed as OAuth and every call 401s);
  the CLI's 401 text always blames the keyring — probe with `curl …/v1/user -H "xi-api-key: …"` first.

## One voice per role, one model per voice

- The campaign narrator is ONE voice id chosen by the operator, rendered with
  ONE model for the campaign's life: `eleven_v3` renders the same id with a different timbre. Identity
  across regens is audited by ear — same id,
  same model, same settings, and a `seed` when a take must be pinned.
- An **off-screen line is never the narrator**: the client hears the narrator asking a question that
  belongs to the character behind the door. A distinct voice is picked from a reel of shared-library
  voices (added to the account free) against the client's brief (a voice brief in their words), then given the **door treatment** (`--door`: high-pass 110 Hz, low-pass
  1.4–1.8 kHz, a short boxy echo, +1.25 gain).
- A **character's spoken line the venue refused** (moderation on profanity) is
  **dubbed in that character's own cloned voice**: an instant clone from the character's own take lines
  (every sample ≥ 4.6 s), the shot prompted with a similar-mouth word, the real
  line rendered by the clone and placed on the mouth shape (`video-edit-edl`: native muted from the
  onset). The clone id is recorded once and reused. A real
  person's voice as a clone = likeness exposure, stated to the operator ONCE.

## Takes and picks

- Three takes per line (stability / style / seed variants) into a NAMED preview reel with 0.7 s gaps and
  an index (`--preview`); the operator picks **by ordinal from the named file** ("the second one from
  PREVIEW-S03-L2.mp3"). "not bad" is not a pick.
- Steer prosody and language without speaking it: `prev` / `next` = `previous_text` / `next_text`
  (English context — multilingual v2 reads slang as another language without it). Slang is spelled the
  way it must sound (**phonetic prompting**: "lada" / "lotta", never "lot of"), and a spelling spread
  across the three takes finds the reading.
- A quip's variant text changes the captions too; the VO script is a table in the EDL (`text` on every
  line) and is quoted in the review ask.

- **A client line note is a take BATCH, not a fix**: the same words re-voiced (delivery is the note) AND
  the client's alternative wording, context-steered, three seeds per setting, reels to the operator for the
  pick; the picked take's word times spliced in with `--only` so hand patches on the other lines survive;
  the new take placed on the OLD beat (last-word end before the hit; speech start after the cut), and the
  placement verified by envelope NCC (< 15 ms) before the render.

## Where a VO file came from — provenance and the gap floor

- Every placed line records `source`: the TTS job id, the clone id, or `extracted_from: <path> @ <s>`. Lifting a
  line out of a delivered cut is a reasonable act — the client approved that reading — and without the field
  nothing in the record says the file was never separated from the mix it came out of.
- **A VO excerpted from a finished cut is not a stem.** It carries that cut's music bed and the scene's native
  audio inside every line: the new mix shows a second track appearing whenever the voice does, muting the events'
  native audio changes nothing, and every placement instrument validates the file against itself.
- `build_vo_stem.py` reads each line's **gap floor** — the 10th percentile of its 50 ms block levels, L/R
  averaged — before placing it, and refuses a line above `--floor-max`, −40 dBFS by default (measured on one job's
  four lines: −28.9 to −38.0 dBFS lifted from a finished cut, −50.4 to −87.1 for the same lines rendered clean; move
  the threshold only on a new measurement — a recorded voice measured after its repair chain, or another TTS vendor's
  clean files, is that measurement for its project). A line that legitimately carries sound under it — a room
  recording, a treated O.S. line — declares `floor_ok: "<why>"`.

## Word times

- `vo_word_times.py --engine local` — faster-whisper on this machine, free; merges hyphen fragments.
- `--engine scribe` — ElevenLabs Scribe, billed per minute; cost line first.
- **An empty or short timing result stops the run.** A line that comes back with no words, or with fewer words than
  the captions reference, exits non-zero naming the line, and nothing is written — a timing gap is never filled by
  interpolation. Upstream precedent (an explainer kit, 2026-09): a TTS library changed its default boundary type,
  returned no word events, and the captions fell back to character interpolation up to half a second out.
- **`--only L2` when one take is swapped**: a full re-run clobbers hand patches on the other lines
  (a spurious 40 ms "The" came back; three patched splits were undone).
- Whisper mis-times onsets in noisy sections: word times drive the cards; where a line SITS is
  measured by `qc_vo_placement.py` on the delivered file, not by the transcript.

## Recorded voice — the repair order

Generated VO arrives clean; when a RECORDED voice enters a lane (a client's phone take, a room mic), the chain runs in
one order only: surgical repairs first (de-click, de-plosive — an impulse), then the neural denoiser, then loudness.
A network trained on additive noise smears an impulse it is handed instead of removing it, and its output has
arbitrary gain, so loudness measured before it is meaningless (`WolfCut`'s experiments bed, 2026-08, the finding it
kept). Never another order.
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