What AI captions get wrong
Speech-to-text tools such as OpenAI's Whisper and the auto-caption features in editing apps get the words mostly right, but their subtitle files are made for a machine, not a reader. The same problems come up again and again:
- Lines that are too long. A whole sentence on one line, far beyond the 42 characters a line that subtitle guidelines such as Netflix's allow, with no line break.
- Overlapping captions. A caption still on screen when the next one starts, so players show two at once or skip one.
- Captions that flash by. "Okay." for 300 milliseconds, gone before anyone can read it.
- Repeated lines. The same caption twice in a row, which speech models sometimes produce on music, silence or a pause.
- Empty captions. Timing with no text.
The sample loaded above has four of these. The cleaner fixes all of them in one pass and tells you how many of each it changed.
What the cleaner changes
Rewrap long lines: each caption's text is split into lines of at most the number of characters you set, as evenly as possible. 42 suits most languages written in the Latin alphabet; some services use 37 or 32 for small screens. Captions that contain styling tags are left as they are so the tags don't break. If a caption still needs more than two lines, the result says so: it's too long to read in one go, and you'll want to split it in your editor.
Fix overlaps: a caption that runs into the next one now ends 80 milliseconds before it starts, about two frames, the usual minimum gap.
Lengthen captions that flash by: captions shorter than the minimum you set (1 second by default) are extended, but never into the next caption.
Merge repeats: a caption with exactly the same text as the one before, starting within a second of it, is merged into it.
The words themselves are never changed, and nothing is uploaded: the file is cleaned in your browser.
A clean workflow for AI captions
- Generate captions in your speech-to-text tool and export SRT or VTT.
- Load the file here and clean it. Download.
- Read it once against the video for wrong words; that part still needs a person.
- If the whole file is early or late, shift it; if it slowly drifts, use the frame rate converter; if a platform wants the other format, convert it.
Questions
How do I fix long lines in Whisper subtitles?
Load the SRT or VTT file and keep Rewrap long lines on. Each caption is split into lines of at most 42 characters, or the number you set, as evenly as possible.
Why do AI-generated subtitles overlap?
Speech-to-text tools time each caption from the audio and don't always leave space before the next one. The cleaner ends each caption 80 milliseconds before the next begins.
Does the cleaner change the words?
No. It only changes line breaks and timing, merges exact repeats and removes empty captions. Checking the words against the video still needs a person.
Is my subtitle file uploaded?
No. The file is cleaned in your browser and never leaves your device.