Captions that lag the voice are read as sloppiness, but the cause is almost never carelessness — it’s a data model. Most caption drift comes from one of three specific things, and two of them are baked in before you ever touch the timing.
Cause 1 — Timing lives on the sentence, not the word
The classic subtitle format stores a start time, an end time and a block of text. That was designed for film subtitles, where a line sits on screen for a couple of seconds and nobody expects it to track the actor’s mouth. It is the wrong resolution for short-form video, where the whole point is that the words move with the voice.
With sentence-level timing, any per-word effect is interpolated: the tool spreads five words evenly across the line’s duration and hopes. Real speech isn’t even — you pause, you rush, you stress a word — so the highlight is ahead on some words and behind on others. Viewers can’t articulate what’s wrong; it just feels cheap.
Cause 2 — You edited after captioning
This is the one that produces the dramatic version, where captions are seconds off by the end. Timestamps are relative to the timeline they were computed on. Cut two seconds out of the middle and everything downstream is now two seconds early. Cut three more places and you’ve got a staircase.
- The rule: caption last. In a timeline editor, finish every cut before the caption pass.
- Or use a tool where the cut and the transcript are the same object. If deleting a sentence from the transcript is the edit, there’s nothing to fall out of sync — the words that remain still carry their own times.
- Watch out for silent trims. Auto-remove-silence features are edits too, and they’re the most common invisible cause of this.
Cause 3 — The audio was hard to transcribe
Boundaries get vaguer as confidence drops, and confidence drops for boring reasons: loud music under the voice, two people talking over each other, a strong accent, heavy room reverb, a phone mic across a table.
- Transcribe from a clean voice track, before you mix music under it, if you can.
- Handing over the finished mix — or worse, the whole video file — is the low-quality path. Extracting the audio first is both faster and more accurate.
- Where a word is genuinely ambiguous, fix it in the transcript rather than nudging its timing. A corrected word usually re-aligns; a nudged one drifts again next time.
Fixing it, in order of effort
- Correct the text first. A large share of “timing” complaints are actually mis-transcribed words that make the line feel wrong.
- Check for a constant offset. If everything is late by the same amount, one global shift fixes it. If the error grows through the clip, an edit happened after captioning.
- Re-transcribe from clean audio rather than nudging thirty lines by hand. On any clip longer than about twenty seconds this is the faster option.
- Move to word-level timing if you want per-word effects at all. Every style in the field guide depends on it; interpolated highlighting is what makes those styles look off-the-shelf.
Why it doesn’t come up in Moonshot
Two design decisions, both upstream of the problem. Audio is extracted and transcribed on its own, and every word carries its own start and end — so highlighting is read from the data rather than interpolated. And because you edit the transcript rather than a timeline, deleting a sentence is the cut: the remaining words keep their own timestamps, so there’s no downstream shift to re-sync.
Quick answers
Why are my captions slightly behind the audio?
Usually because the timing is stored per sentence rather than per word. The line appears when the sentence starts and sits there for its whole duration, so any word you look at is at best approximately timed — and the mismatch is most obvious at the start and end of each line.
Why did my captions go out of sync after I trimmed the video?
Because the timestamps are relative to the original timeline. Cut two seconds from the middle and everything after that point is two seconds early — a constant offset that compounds with each additional cut. Caption last, or use a tool where the cut and the transcript are the same object.
What are word-level timestamps?
Timing information attached to every individual word rather than to each line. It's what makes karaoke-style highlighting possible, and it's also what makes captions survive edits: each word knows when it happens, so re-flowing lines doesn't require re-timing anything.
Does music make transcription timing worse?
Yes. Loud backing music, overlapping speakers and heavy accents all reduce confidence, and lower confidence shows up as vaguer boundaries. Extracting and transcribing a clean audio track, rather than handing over the finished mix, meaningfully improves it.
Captions that track the voice
Word-level timestamps from a clean audio pass, and a transcript you edit directly — so trimming a sentence can't push everything after it out of sync.
Open Moonshot free