The slow way to proof captions is to replay the whole video and hope every kind of mistake announces itself. The fast way is six narrow passes. Each pass asks one question, so your eye stops treating plausible text as correct text.
The six-pass check
| Pass | Look for | Review mode |
|---|---|---|
| 1. Names | People, brands, places, acronyms | Read the transcript |
| 2. Numbers | Prices, dates, units, percentages | Listen and compare |
| 3. Timing | Early words, late words, growing drift | Watch at normal speed |
| 4. Lines | Orphans, broken phrases, three-line blocks | Watch with sound off |
| 5. Emphasis | Too much color, wrong hero word | Scan the frame |
| 6. Export | Crop, font fallback, missed edits | Watch the final file |
1. Search for words the model could not know
Begin with proper nouns, not ordinary grammar. A transcript can be 99% readable and still miss the one word that damages trust: the guest’s surname, the product name, the city, the technical acronym.
Search the transcript for every name you expected to say. If the clip mixes languages or scripts, check each switch separately; a model may turn an unfamiliar phrase into fluent but unrelated text. Our Hindi and Hinglish caption guide covers the line-box and transliteration problems that follow.
2. Listen to numbers as data
Numbers are short, context-heavy and expensive to get wrong. “Fifteen” and “fifty” can both produce a grammatical sentence; the viewer cannot infer which one you meant when the caption disagrees with the voice.
- Check prices and currencies together.
- Keep units attached to the number they qualify.
- Use one consistent style for numerals across the clip.
- Recheck any number that appears in the hook or cover.
3. Diagnose timing by the shape of the error
Do not drag every caption block before you know what kind of sync problem you have. If every word is equally late, correct the track offset. If the error grows toward the end, the timing model or media clock is drifting. If one phrase is wrong, fix that phrase rather than moving the entire track.
The full diagnostic is in why captions drift out of sync.
4. Break lines where the thought breaks
Correct words can still be hard to read. Read each caption as a visual phrase and keep modifiers attached to what they modify. Avoid leaving articles, prepositions or pronouns stranded at the end of a line.
- Better: “The fastest fix / is the first cut.”
- Harder: “The fastest / fix is the first cut.”
Preview the longest real line on a phone. Desktop width hides the exact wrapping failure that viewers will see.
5. Spend emphasis on meaning
Auto-emphasis often follows loudness or lexical importance. Review it as a design choice: does the highlighted word carry the contrast, result or object of the sentence? If three neighboring words all demand attention, none of them creates hierarchy.
Remove emphasis before adding more. The scarcity rule in which words to emphasise is a useful final filter.
6. Review the file the viewer will receive
The editor preview is not the deliverable. Export, move the file to a phone, and watch once with sound off. This catches font fallbacks, clipped matras, captions under platform controls, a stale line that missed the last transcript edit, and contrast failures on bright frames.
- Watch the opening twice: once for the words, once for placement.
- Scrub to the visually brightest frame and check contrast.
- Check the longest two-line caption.
- Watch the final ten seconds without touching playback.
Quick answers
What should I check in automatic captions?
Check proper names, numbers and units, word timing, line breaks, emphasis, then watch the exported video once with sound off. Separate passes work better than trying to notice every kind of error in one read.
Why are auto captions wrong even when the audio is clear?
Names, brands, acronyms and code-switched phrases may be rare in the model's language data even when they sound clear to a person. Context can also push a plausible common word ahead of the correct unusual one.
How do I fix captions that are out of sync?
First determine whether every word has shifted by the same amount, the drift grows over time, or only one phrase is late. Those patterns point to an offset, a timing-model problem, or a local transcription error respectively.
Should captions include every filler word?
Keep fillers that carry tone or hesitation; remove ones that only reflect the speaker searching for the next phrase. If you remove a spoken word from the caption without cutting the audio, check that the remaining timing still feels natural.
Correct the transcript once
In Moonshot, changing a word updates the caption track and keeps the edit in sync—so proofing does not become a second timeline job.
Fix a clip free