The quick answer: auto captions in four steps
Your video is finished and it needs captions. Typing them line by line means pause, type, rewind, for the whole video, and many free caption sites want your file uploaded before they start. Auto Subtitles transcribes the speech on your own device instead, so the video never leaves the tab.
- Drop in the video (MP4, MOV, WebM or MKV, or an audio file) and pick the spoken language. Choose captions in that language, or translated to English.
- Generate. The speech model downloads the first time, then runs on your device. Captions appear as each part is transcribed.
- Fix and style. Click any line to jump the video to it and correct the words, then choose a look.
- Export an MP4 with the captions burned in, or download SRT, WebVTT or a plain-text transcript.
Burned in, or a separate subtitle file?
This choice decides who controls the captions after you publish, so make it before you export.
| Export | Best for | Viewers can switch off | Good to know |
|---|---|---|---|
| MP4, captions burned in | Reels, TikTok, Shorts, ads, LinkedIn, any look you want to control | No | Part of the picture, so a typo means exporting again |
| SRT | YouTube, Vimeo, course platforms, editing software | Yes | The most widely supported subtitle format |
| WebVTT | Web players and the HTML track element | Yes | Can carry cue positions and comments |
| Text transcript | Show notes, blog posts, quotes | Not applicable | Words only, no timing |
You can do both: burn short-form captions into the vertical cut and keep an SRT for the long version. To upload one to YouTube, open the video in YouTube Studio, go to Subtitles and choose Upload file (menus change, so check the current labels). Once the subtitles are in, chapters are the next thing that makes a long video easier to navigate. Style, size and timing choices are covered in captions people actually read.
How speech recognition runs on your device
A speech recognition model is a trained neural network that turns sound into words with timings. ClipLime uses Whisper, an open-source one. Rather than sending your audio to a server, your browser downloads the model once, keeps it, and runs it locally: on your graphics card through WebGPU when it is available, otherwise on your processor.
You pick a model size. Fast is for quick drafts, Balanced for most videos, Accurate for accents, jargon or background noise. The download sizes below are what the tool uses at the time of writing.
| Model | Download on a processor | Download with a graphics card |
|---|---|---|
| Fast | About 41 MB | About 120 MB |
| Balanced | About 77 MB | About 206 MB |
| Accurate | About 249 MB | About 586 MB |
After the first download your browser caches the model, so the next video starts almost immediately, until you clear your browser data. With a capable graphics card, transcription is often faster than real time, though speed depends on your device and the model size. A processor is slower but still practical, and very long recordings take more time and memory, so use a desktop or laptop for them.
Languages, and what translation means here
The model handles more than 30 spoken languages, including English, Spanish, Portuguese, French, German, Italian, Hindi, Urdu, Arabic, Turkish, Russian, Japanese, Korean and Chinese. Pick the spoken language yourself before you generate.
Translation goes one way: non-English speech can be turned into English captions. There is no translation into other languages, and accuracy varies with the language and how clearly people speak. Expect to proofread.
Fix mistakes before you export
No speech recognition is perfect, least of all with names, brands and numbers, which are the words viewers notice. The editor makes the fixes quick:
- Click a line to jump the video to it, then type the correction.
- Type exact times, merge lines or delete them.
- Use Find & replace to fix a misheard name everywhere at once.
- Try Accurate if the first pass is rough.
Do the edit first. Captions made after you cut dead air, for example with the silence remover, match the final timing, and you proofread only once. The method is in removing pauses from talking videos.
Already have subtitles? Convert, re-time and tidy
Import an existing SRT or VTT into Auto Subtitles to restyle and burn it in without transcribing. To change the file itself, use the Subtitle Fixer. It opens SRT, VTT, YouTube SBV and ASS/SSA files, and saves SRT, VTT or TXT as UTF-8. Converting SRT to VTT keeps italics, bold and underline. VTT to SRT drops the positions and comments SRT cannot hold.
Out-of-sync subtitles follow three patterns, and the pattern tells you the fix:
| What you see | Fix in the Subtitle Fixer |
|---|---|
| Every line is early or late by the same amount | Shift everything |
| Fine at first, wrong after a certain point | Shift from the first wrong cue onward |
| Fine at the start, steadily worse towards the end | Sync to two points, or change the frame rate (23.976 against 25 fps is the classic cause) |
Its checks also flag overlaps, empty cues and cues shown for too short a time to read (more than 20 characters per second), with one-click fixes. A common guideline is lines of about 42 characters and two lines at most.
Check the placement before you post
Burned-in captions are part of the picture, so the order matters. If the video is going vertical, reframe it first with Resize for Social, then caption it, so the captions are laid out for the final frame instead of being cropped later (sizes are in the vertical video guide).
Then use the position, size and font controls, and choose Sentences for classic subtitles or Short-form for big words a few at a time. You can switch between the two without transcribing again. Finally, open the exported video in the safe zone checker to see whether the TikTok, Reels or Shorts buttons cover anything. The zones are approximate, so leave some margin.
What goes wrong, and what it cannot do
- Gibberish or the wrong script. The spoken language was set wrong. Set it to what is actually spoken and generate again.
- The same name wrong ten times. Use Find & replace, and try Accurate next time.
- Accented letters look broken in an old SRT file. Older files are often Windows-1252 rather than UTF-8. The Subtitle Fixer reads them correctly and saves UTF-8.
- A long video is slow or runs out of memory. Use a desktop or laptop, and cut the video down first with the video trimmer.
Honest limits: it does not translate into languages other than English, it will not get every name right, and burned-in captions cannot be turned off by viewers.