Why captions from the script drift
A script says what each shot should contain, not when each word lands. Generated clips speak at their own pace, trim differently in the edit and sometimes add or drop a word. On Blood Sun: The Fall of Kratos, captions timed from the plan could sit a whole line early. Re-timing them against speech-recognition word timings fixed 396 of 400 lines.
Why speech recognition alone is not enough
Speech-to-text gets common words right and invented names wrong. The first caption track of The Fall of Kalak had six name errors. Fantasy and drama scripts are full of names, so check every proper noun before upload.
The method we use
- Lock the edit and mix the final audio.
- Run speech recognition on the mix to get word-level timings.
- Align the script text to those timings, so you keep the script's spelling and the audio's timing.
- Keep the speech-recognition wording where the generated clip changed a word, so captions match what is heard.
- Check proper nouns, then check any line the model invented (caption it or cut it).
- Export as a separate subtitle track for long films; burn captions in for vertical episodes, which are often watched without sound.
Burned in or separate?
| Format | Captions | Why |
|---|---|---|
| Vertical episodes (9:16) | Burned in | Feeds autoplay muted; burned-in English captions also told one platform's language detection that the series was in English |
| Long 16:9 films | Separate subtitle track | Viewers can turn them off; we published The Wolf Mother and The Fall of Kalak with English caption tracks |
| Trailers and clips | Burned in | Short, often muted, reposted out of context |
Where Katechon Arc fits
Arc keeps the exact script text for every shot, which is what you align recognized speech against, and exports episodes and a Jianying draft for captioning in an editor.
Arc does not transcribe or time captions for you; captioning happens after the edit, in your editor or a captioning tool.
People also ask
Should captions be burned into vertical AI dramas?
Yes, for vertical episodes. They autoplay muted in feeds, and readable captions keep viewers who have sound off.