How to find the words your captions left out

Last updated 2026-09-10 · 8 min read

What is the short version?

If your edit removed pauses, the captions were rebuilt around those cuts, and any caption event that began exactly on a removed boundary can fall through the gap. The video keeps talking and the text goes quiet for a second, which looks like a glitch rather than a missing word.

1. Go to Vizard Agent with the finished cut.
2. Say where the captions go silent while someone is speaking.
3. Ask for an audit against the full transcript.

What do you need before you start?

The rendered video and the transcript it was built from. Vizard Agent compares one against the other, so the audit is only as good as the transcript underneath it — and if the pauses were cut before the captions were made, that mapping matters as much as the words do.

• The finished cut. As delivered.
• The transcript. With word-level timings.
• Where you noticed it. One timestamp is enough.
• Whether pauses were removed. That is usually the cause.
• The caption file. If it was produced separately.

What do you type into Vizard Agent?

Describe it to Vizard Agent as words going missing rather than as a timing problem. A gap in the captions and a caption that arrives late look almost identical on playback, and Vizard Agent will go looking for a timing fault if timing is what you name.

Prompt

Variants worth knowing:

• "Blank while she is still talking." Missing, not late.
• "Against every word." A full audit, not a spot check.
• "Group them into phrases." Restored words read oddly alone.

What does Vizard Agent actually do?

Here is the sequence from a real talking-head edit where two caption windows had gone silent. The interesting part is the root cause, which is not a transcription failure at all — the words were transcribed correctly and then lost when the picture was tightened.

1. Inspects the reported timestamps on the render.
2. Traces the caption events around the silent windows.
3. Checks the source caption entries by their spoken text.
4. Examines the pause-mapping logic that rebuilt the timings.
5. Finds captions that began exactly on a removed boundary.
6. Restores those events and re-renders.
7. Confirms the caption now appears immediately after the cut.
8. Audits every caption event against every transcribed word.
9. Adds the remaining missing words.
10. Checks the restored words for overlap and readability.
11. Regroups them into natural phrases.

Step four is where the real answer lives. It is tempting to assume a missing caption means a missing word in the transcript, but when pauses are cut from the picture the caption timings have to be remapped onto the shortened timeline, and events sitting on the boundary are exactly the ones that mapping gets wrong.

Step eight is the difference between fixing the two you noticed and fixing the ones you did not. A full audit against the transcript is cheap and catches the words nobody happened to be watching for, which on a long edit is usually more of them than you would expect.

Step eleven matters for how it reads. A word restored on its own flashes up as a single item between two proper phrases, so Vizard Agent merges it back into the group it belongs to rather than leaving it floating.

What does the result look like?

Captions that carry every word that was actually said, in phrases that read normally, with no silent windows while somebody is mid-sentence. Vizard Agent will also tell you how many words it restored, which is the number that tells you whether this was a one-off or a systematic mapping problem.

When does this not work well?

An audit compares captions against a transcript, so anything the transcript got wrong stays wrong. Words that were never transcribed cannot be restored by comparing against a record that does not contain them, and heavily accented or overlapping speech is exactly where that happens.

• Words the transcript missed. There is nothing to restore.
• Overlapping speakers. The transcript is unreliable there.
• Deliberately dropped filler. Not every omission is a fault.
• Translated captions. They will not match word for word.
• Heavily stylised caption tracks. Some drop words by design.

How do you fix a result that came back wrong?

Say whether the words are missing or misplaced. Because Vizard Agent has both the transcript and the caption events, a note that names which of the two is wrong resolves in one pass instead of turning into a general caption rebuild.

• "Still blank at 0:31." A second event on that boundary.
• "Now it overlaps the next line." Regrouped and re-timed.
• "That word was cut on purpose." Excluded from the audit.
• "The transcript is wrong there." Re-transcribe that section.

How does Vizard Agent compare to doing it yourself?

By hand the two gaps somebody noticed get patched by typing the words back in, and nobody checks the rest of the video. The mapping bug that caused them is still there, so the next edit of the same footage produces the same problem in different places.

Common questions

Why do captions disappear at cuts? Because events on a removed boundary get lost when timings are remapped.

Is it a transcription failure? Usually not. Vizard Agent checks the transcript before blaming it.

Will it catch every one? Vizard Agent audits every word against the transcript, which is the best available check.

Does it change the style? No. Vizard Agent restores events without touching the look.

What about words cut on purpose? Tell Vizard Agent which and it excludes them from the audit.

Can it check a subtitle file? Yes, if you hand Vizard Agent the file alongside the video.

Will restored words overlap? Vizard Agent checks for that and regroups where they do.

Does it re-render everything? Vizard Agent rebuilds the caption layer only. The picture is untouched.

How many words is normal to lose? None. Vizard Agent treats any loss as a mapping fault worth fixing at its source.

Can it show me what it added? Yes. Vizard Agent lists each restored word with its timing and spoken text.

Does this apply to translated captions? The audit does not, because the words will not match.

What if pauses were not removed? Then the cause is different and Vizard Agent will look elsewhere.

Can it run before delivery? Yes, and asking Vizard Agent for the audit before delivery is much cheaper.

Why did playback look fine? Because a one-second gap in text reads as a pause, not as an error.