Subtitle generator with private AI

Making a subtitle file, start to finish

Open your video or audio and press Transcribe on this device. Once there is a transcript, the Subtitles panel opens under it. Every transcript line is already a cue with a start and an end time.

Read the cues against the recording and correct the words in place. Then look at the panel's checks, repair what it flags, and download the file as SubRip (.srt) or WebVTT (.vtt).

Load the file beside your video. Most players and editors accept an .srt next to the video file, and web players generally take WebVTT.

The subtitle files are written from what is on screen, so a corrected word in the transcript is a corrected word in the cue. Nothing is uploaded at any step.

What the Subtitles panel measures

The panel checks every cue against four guides. A cue should show no more than 17 characters a second and hold at most 84 characters. It should stay on screen for one to seven seconds and never overlap the next cue.

A cue that breaks a guide is flagged, and the panel offers a repair. Split divides a long cue in two. Merge joins a short cue with the one after it. The timing buttons hold a cue a tenth of a second longer.

These repairs change timing and layout only. They never change a word and never change a speaker. That keeps the checking job and the editing job separate: you fix what was said, the panel fixes how it is laid out.

Before you get the file, it is opened again and measured. Subtitle files are parsed back by something that never saw the transcript, and if a check fails there is no download at all.

Reading speed, in plain terms

A viewer reads a subtitle while the picture moves. Too many characters in too little time, and the viewer is still reading when the cue has gone. The 17 characters a second guide is the panel's way of catching that.

The long cue is the usual culprit. When a speaker runs a sentence on, one cue becomes a block of text. Splitting it into two gives each half its own time and its own line.

The short cue is the opposite problem. A word or two that flashes by in under a second is easy to miss. Merging it with its neighbour, or holding it longer, gives the eye time to land.

None of this makes a subtitle accurate to the speech. It makes the cue readable. The wording is still yours to check against what was said.

What this route does not do

The tool transcribes and never translates. Subtitles come out in the language that was spoken. A French video gives French subtitles, and translating them is a separate job for a person or another service.

It does not burn subtitles into the picture. The output is a separate text file, and you attach it in your player or editor.

It does not separate speakers by itself. Labels are typed by hand. If you add one, it can appear in the subtitle text, so check whether you want it there.

Deciding what goes into the cue text

Each cue is a line of the transcript, so the cue text is whatever that line says. If you typed a speaker label, it can appear at the start of the cue, as in the example on the home page. Decide whether a viewer should see it.

Correct the wording before you repair the layout. A correction that lengthens a line can push a cue past the reading-speed guide again, and repairing it twice is wasted effort.

Play the video with the file loaded before you publish it. The panel measures timing against its guides, but only a viewer can tell whether a cue appears where the speech does.

More questions about subtitles here

Is there a limit on how long the video can be?

The limit is your device: a file up to 2 GB on a computer, and shorter on a phone. Try the first five minutes before you run the whole file.

Which format should I choose, SRT or WebVTT?

SubRip (.srt) is the one most players and editors take beside a video file. WebVTT (.vtt) is the format used by web video players. Both are written from the same cues, so you can download both.

Why did the panel flag a cue I think is fine?

The guides are reading-speed rules, not judgements about wording. A cue is flagged when it runs past 17 characters a second, passes 84 characters, or falls outside one to seven seconds. You may keep it as it is.

Can I fix a wrong word after exporting?

Correct it in the transcript and export again. The page does not keep an unsaved transcript after a reload, so finish the corrections before you close the tab or save the transcript to your account.