Fix Auto-Generated Subtitles

Where machine subtitles go wrong

Speech recognition is strong on common words in clear audio and weaker on everything else. Its mistakes cluster in a few places:

• Names of people, places, products and brands, which the model may never have seen written down.
• Numbers, dates, prices and units, which can appear as digits in one line and as words in the next.
• Words that sound alike, such as their and there or accept and except, and technical terms heard as everyday phrases.
• Punctuation and sentence boundaries, which have to be guessed from pauses and intonation.
• Speech over music, two people talking at once, and accents the model handles less well.

During silence or music, recognisers can also produce words nobody said, a known quirk explained in Whisper hallucinations. None of this makes the draft poor. It tells you where the review time should go.

Two sources of automatic subtitles

The editor handles both kinds you're likely to be correcting.

AI subtitles from this editor

Leave "Write the subtitles automatically with AI" on when you choose a video, or click AI subtitles once it's loaded. Your browser sends only the audio to mydubly's server, where Whisper transcribes it and detects the language. The lines come back short: one line each, about 32 characters at most, up to roughly 2.6 seconds, starting on the first spoken word. The cost is 1 credit per minute of video, minimum 5 credits.

YouTube auto-captions

If the video is yours, YouTube Studio lets you download its automatic captions as a file. Pick SRT or VTT, then use Import SRT / VTT here. Formatting tags and inline word timings are stripped on import, leaving plain text.

YouTube's automatic captions often come without punctuation or capital letters. Some VTT copies, especially ones saved with download tools rather than from Studio, also repeat the previous phrase at the start of each cue so the text rolls upward. If your import shows every phrase twice and a crowd of very short lines, that's why; the cleanup is the same, just longer. Editing YouTube auto captions covers the YouTube side.

The review order

Work in passes, one kind of error per pass. A focused pass goes faster than one slow read trying to notice everything at once.

1. Names and terms. Before pressing play, jot down the names, brands and jargon the video uses. Then skim the line list for near misses. A misheard name is usually wrong the same way every time, so once you've seen the pattern it jumps out.
2. Numbers. Check every figure against the audio and make the style consistent: digits for prices, times and measurements, words where the speaker is talking loosely. Numbers in transcripts has the conventions.
3. Misheard words. Now play at normal speed and read along, listening for lines that are grammatical but mean the wrong thing. A spell-checker never catches these.
4. Punctuation and capitals. Add full stops, question marks and capitals where sentences really end. Short AI lines often split a sentence across several subtitles, so punctuate across them as one sentence.
5. Timing, last. Only once the words are right, fix lines that arrive late, hang on too long or flash past. Doing timing first wastes effort, since changing the words changes how long a line needs.

Let the warnings point the way

Each row in the line list shows its warnings underneath, flagged blocks are marked on the timeline, and the Subtitles tab shows how many lines have a problem. On automatic subtitles these are the ones that matter:

Shown for under 0.8 s

Often a one-word fragment or a duplicate from a rolling caption file. Delete it if it repeats the line before, or Merge it into the next line.

Too fast to read

Over 21 characters per second for most scripts, spaces not counted, with lower limits for Japanese, Chinese, Korean and Thai. Cut filler, give the line more time, or Split it.

A line longer than 42 characters

It will wrap, which is fine for two lines but can point to a run-on. Split at a natural pause or shorten it.

More than 3 lines

Usually the result of merging too much. Split it again.

No text

A line you emptied while editing. It's skipped in the export and the render; delete it to keep the list tidy.

Overlaps

Two lines on screen together. Right for crosstalk, a mistake almost everywhere else.

Clear the warnings that matter, but don't chase every one blindly: a short exclamation shown briefly can be exactly right.

Cleaning up in the editor, step by step

1. Open the editor and choose the video. Keep the AI option on to generate subtitles, or switch it off and use Import SRT / VTT for a YouTube file.
2. Press Space to play and Up or Down to move between lines. Selecting a line jumps the video to it.
3. Correct words in each line's text box. Bursts of typing are grouped into single undo steps, so a bad edit is easy to reverse.
4. Split lines that carry two sentences at the playhead, and Merge fragments into the line after them.
5. Drag a line's edges on the timeline, or type exact times, where it starts late or ends early.
6. Work through whatever warnings remain in the Subtitles tab.
7. Export SRT or VTT to keep the corrected file, and render the video if you want the subtitles burned in.

Undo reaches further than you might expect. If you click "Replace with AI subtitles" and then prefer your edited version, Ctrl+Z or Cmd+Z brings it back, although the credits for the new run are spent. A draft is also saved in this browser moments after each change, so closing the tab doesn't lose your corrections to that video. For the mechanics of dividing and joining lines, see splitting and merging subtitles.

Three cleanups in practice

What automatic subtitles leave out

Automatic subtitles are words only. There are no speaker names and no descriptions of sound such as music or laughter, which deaf and hard-of-hearing viewers rely on. If the video needs closed captions or SDH, add them by hand during the review: a name before a line when the speaker is off screen, a bracketed note for sounds that carry meaning. SDH subtitles and captioning sound effects and music explain the conventions.

Translated AI subtitles are timed by text length rather than matched to each spoken word, so allow a slightly longer timing pass for them. And when the source audio is poor, cleaning the recording before transcription saves more time than any amount of correcting afterwards.

Before you publish

• Every name and brand spelled the way its owner spells it.
• Numbers correct and written consistently.
• No line that reads well but says the wrong thing.
• Sentences punctuated, even when they run across several subtitles.
• No warning left that you haven't decided to keep.
• One last watch at normal speed, with sound, start to finish.

Proofreading an AI transcript applies the same thinking to full transcripts. If you only need a corrected subtitle file and not a video, the AI subtitle generator produces one you can bring here for this review.

About the subtitle editor on this page

Opens

MP4, MOV, M4V, WebM, MKV videos your browser can decode

Subtitles from

AI (from the speech, optionally translated into one of 21 languages), an SRT or VTT file, or typing them in

You download

An MP4 with the subtitles drawn into the picture at the original resolution and frame rate, plus SRT and VTT files

Cost

Editing, styling, rendering and SRT/VTT export are free, with no account and no watermark. AI subtitles cost 1 credit per minute (5-credit minimum per video); signing in with Google gives 100 free credits and accounts get 10 free credits a day

Privacy

The video never leaves your device and is rendered in your browser. Only for AI subtitles is the audio sent, over HTTPS, and it is deleted within 30 minutes of delivery

Rendering needs

Chrome or Edge 94+, Safari 16.4+, or Firefox 130+ on a computer. Editing and SRT/VTT export work in any current browser