Localize a video

The viewer should hear a narrator who talks the way people in their country talk, saying exactly the same facts, in time with the same pictures. Start from the final script (after the fact-check) and the final render: a later change to the source means re-translating and re-voicing every changed paragraph.

$SKILL below is this skill's folder.

Files

dub/<lang>/script.md       translation, same paragraph ids (P01, P02, ...) as script.md
dub/<lang>/tts.md          the text sent to the voice engine (numbers spelled out), if it differs
dub/<lang>/voice/P01.wav   one audio file per paragraph, named by its id
dub/<lang>/narration.wav   the paragraphs joined into one track (for the voice check and YouTube Studio)
dub/<lang>/captions.srt    captions timed to the dub
qa/voice-check-<lang>.json, qa/dub-fit-<lang>.json
upload.json                "localizations": {"<lang>": {"title": ..., "description": ...}}

1. Translate for meaning, not word by word

• Rebuild each sentence the way a native narrator would say it: word order, idiom, form of address and rhythm of the target language. Keep the creator's questions, doubts and emphasis; never invent experiences or evidence.
• Keep exactly: every number, unit and date in facts.json; the currency (a US company's dollars stay dollars); the country and institutions of the events; hedges ("about", "reportedly", "the company says"); product names and commands.
• Keep each sentence in its own paragraph: every paragraph has to fit its own slot in the picture.
• Length targets per language, the number table and the regional variant: $SKILL/references/translation.md. Shorten by cutting filler and repetition, never a number, a name, a direction (up/down, profit/loss) or a hedge.
• Title and description say the same promise in the target language, with the same chapter times and links.

2. Native-listener audit

Read every paragraph twice, first against the source for meaning, then aloud for naturalness:

• Hedges kept: "can keep" did not become "keeps".
• Time direction right: "years ahead" did not become "years ago".
• No subject drift: who sells, who pays, who owns is still clear.
• No translationese: calqued idioms, source word order, bureaucratic verbs.
• Names, terms and numbers read the way this language says them.

A model-only naturalness check is not human review. Unless a native speaker listened, write "model audit, not a native speaker" in the report.

3. Voice

Read the dubbing station from ~/.tubecrew/stack.json (run the setup skill if it is missing):

• cosyvoice: local, free, one job at a time on the GPU. Use a clean reference clip of the creator's own voice.
• elevenlabs: costs credits. Estimate them first (about one credit per character of tts.md), tell the creator the number and the remaining balance, and generate only after a yes.
• none: stop after translated captions, title and description.

Only clone a voice whose owner agreed. Detail and retry rules: $SKILL/references/voice-and-timing.md.

4. Check each language

Back-transcribe the dub with the check-voiceover skill, then check the timing:

python <check-voiceover>/scripts/voice_check.py dub/es/script.md dub/es/narration.wav --language es \
  --out qa/voice-check-es.json
python $SKILL/scripts/dub_fit.py script.md qa/voice-check-transcript.json dub/es/voice --lang es

voice_check.py takes the paragraph files joined in order (or the assembled track). dub_fit.py needs the original narration's word-timed transcript (voice_check.py saves one when it transcribes audio). It FAILs a paragraph that needs more than 1.25x speed-up or starts more than 4 s from where the original paragraph started, and says about how much shorter to rewrite it. Fix by rewriting the translation, never by speeding the voice past the cap; then re-voice only the changed paragraphs and run both checks again. A high back-transcription score does not prove natural delivery: list the lines that need a human listen.

5. Package

• Captions follow the dubbed timing; a number stays on one line with its unit.
• Write each language's title (at most 100 characters) and description (at most 5000 bytes) into upload.json localizations; upload-private sends them with the video.
• YouTube's API cannot attach extra audio tracks: the creator adds each dubbed track in YouTube Studio. Nothing here publishes anything.

Report back

Per language: status (text ready, audio ready or blocked), the voice_check and dub_fit verdict lines, the paragraphs re-voiced, lines that need a human listen, credits spent, and who did the native audit.