SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsvoice-cast

voice-cast

Agents & models

View repository

Updated 10-29 15:01

About this skill

voice-cast

Every line in the game is voiced. The script is the source of truth: a line lives in a Dialogue Manager file with a static [ID:…], and that ID names its audio file, its manifest entry, its lock entry and its translation key. Nothing is generated twice for the same text.

Proven in Past: extract, diff, report, post, the envelope table and the runtime (Voice autoload, Say holding lines, stall fallback, browser click). Not proven yet: casting, generation (voice) and review (check). Past's plan 003 slice 2 waits on the ElevenLabs plan and credit ceiling. Those steps are written in references/designed-process.md as the designed process, not a measured one. Run them on a small first batch and measure before any volume.

Read RULES.md before generating anything. Commands are in WORKFLOW.md.

The loop

script ([ID:], [#vo=])  ->  extract  ->  manifest  ->  diff vs lock  ->  voice (new + changed only)
                                                                              |
   game line + envelope  <-  post  <-  masters  <-  check + operator listen  <-+
  1. Lines first. Every spoken line has a static [ID:…] and, where it matters, direction as a tag: [#vo=whispers]. Direction never goes into the spoken text. tools/test.sh's lint fails on a missing or duplicate ID.
  2. Extract. bun tools/audio/run.ts extract reads content/dialogue/** into audio-source/voice/manifest.json: ID, speaker, text, direction, file:line and a text hash. Responses (- text) are spoken by the hero, named in tools/audio/config.json as player.
  3. Diff. bun tools/audio/run.ts diff lists new, changed and orphaned lines against audio-source/voice/lock.json. Only those cost credits.
  4. Cast and generate (designed, not proven): one Voice Design voice per character, one model per character, one request per line, cached by hash.
  5. Review (designed): word match against the script, then the operator listens. Claude cannot hear.
  6. Post. bun tools/audio/run.ts post trims each master (80 ms lead, 200 ms tail), sets one gain to -18 LUFS capped at -1 dBTP, writes a 24 kHz mono WAV to assets/audio/voice/en/<ID>.wav, and writes the 30-a-second envelope table assets/audio/voice/en/envelopes.gd. Only changed masters are posted again.
  7. Import and play. tools/test.sh (or the editor) imports the WAVs as QOA. Voice.play(id) streams the line on the Voice bus; Say holds the subtitle for it; the hero's body eases toward Voice.level().
  8. Prove it in both builds with the voice QA scenario (see ship-check).

Quick start

bun tools/audio/run.ts extract          # manifest from content/dialogue
bun tools/audio/run.ts diff             # what would cost credits
bun tools/audio/run.ts report           # credits used of the limit (needs ELEVENLABS_API_KEY)
# masters land in audio-source/voice/masters/<ID>.wav (generation: designed-process.md)
bun tools/audio/run.ts post             # game lines + envelopes.gd
tools/test.sh                           # imports as QOA, runs GUT and the lint

Per-batch checklist

  1. Every new line has an [ID:…], a speaker that matches an actor's speaker_name or one of its speaker_aliases, and [#vo=…] on turning points.
  2. extract runs clean (a repeated ID is an error) and diff shows only the lines you meant to change.
  3. report before and after the batch, and spend stays under 80% of the operator's credit ceiling.
  4. Masters are kept in audio-source/voice/masters/ and never overwritten without a reason recorded in the lock.
  5. post reports each line's seconds and gain; no line hits the -1 dBTP cap by surprise.
  6. tools/test.sh passes, and --qa=voice passes in the exported desktop app and in Chrome.

Do not

  • Do not put emotion in the spoken text. It is read aloud. Use tags.
  • Do not change an ID to regenerate a line. Change the text; the hash marks it.
  • Do not use Voice Library voices for the cast. Owners can withdraw them.
  • Do not mix models within one character.
  • Do not request Opus. Godot cannot import it.
  • Do not ship WAVs or loose audio files. The export carries QOA inside the pack.
  • Do not analyse audio at runtime. Envelopes are computed offline.
  • Do not print or commit ELEVENLABS_API_KEY. It lives in .env.

Next: ship-check proves the voiced lines in both builds.