Why captions end up out of sync, and how to check

Play a few seconds near the start and a few near the end, and watch when captions arrive against the voice. What you see will usually fit one of three patterns:

A known cause of a constant offset is a clip trimmed before upload. A video file stamps every frame and every slice of sound with a time, and after some trims those stamps keep the numbering of the longer original, so your clip's first frame might be marked one second in rather than at zero.

That gives two clocks. The file's clock says the video starts at one second; the browser's video player generally counts from zero at the first frame it shows. Measure word timings on the file's clock and show them against the player's, and every caption lands a second late, in the preview and the export alike.

AutoCap now measures both from the player's zero: the audio sent for transcription is laid out from it, and the exported MP4 draws each caption against it. Saved projects are not rewritten, so one transcribed before this change keeps its old timings; if every caption in it is late by the same amount, a single shift fixes it.

Step through a clip slowly and you may see a word light up just before you hear it. That is deliberate: AutoCap draws captions against a clock running 0.11 seconds ahead of the video. Pop and rise entrances take roughly a quarter of a second to reach full size, so one that began exactly on the sound would peak after the word was heard; leading by about half the animation puts the peak on the sound.

The lead is added only when captions are drawn, in the preview and the exported MP4, and every style gets the same amount; saved timings and SRT and VTT files don't include it. When checking sync, count a caption that is a hair early as correct.

Captions also linger briefly after their last word so they don't blink off. On every style except One word, the next caption waits for that when there is no pause, so in fast speech it can land just after its first word begins.

'Captions early or late?', above the caption list in the Captions panel, has one tool for each of the first two patterns.

For a constant offset, type an amount in milliseconds and press 'Shift all'; a negative number moves captions earlier. Every caption moves by that amount, including ones you typed, and none can be pushed before the start of the video. The field steps in 50-millisecond increments, so try a value, play and adjust.

For drift that grows, press 'Re-sync to audio'. It reads the sound in your video file again and lays the words back onto the speech, judging by where the voice starts and stops rather than recognising each word. Nothing is transcribed, so it uses no transcription minutes, but it needs the video on the device you're editing on: without the file, the button doesn't appear. Undo it if the result looks worse.

A few habits make checking quicker and less of a guess:

A shift is right only when the error is the same everywhere; on drift, it moves the worst of it from one end of the video to the other. It also moves everything, so a title you placed on a particular shot needs setting back with its In and Out times or 'To playhead'.

'Re-sync to audio' does less than its name suggests. It is built for words whose timings were estimated by spreading a phrase across its words, which is how some transcriptions arrive; the editor says whether yours did right after transcribing. Words that came with their own timings give it little to correct, though it may still pull their edges in towards the sound. Captions you added by hand are never moved, and where speech is hard to pick out, as under loud music, in a whisper or on a poor microphone, it changes little or nothing.