Generate
Upload a video and CaptionFX transcribes the speech into captions that are already aligned to the timeline, with a start and end time on every line and on most individual words.
CaptionFX is an AI video caption generator that turns speech into time-aligned captions, then hands them to you as editable text over the playing video. You fix what the model got wrong, style the caption layer, and export a new MP4 with the captions burned into the frames.
CaptionFX is an AI video caption generator. You upload a video, it transcribes the speech into time-aligned captions, you correct any word over the playing video and style the captions, then you export a new MP4 with the captions burned in. Your first project, up to five minutes, exports free.
Last updated
Upload a video and CaptionFX transcribes the speech into captions that are already aligned to the timeline, with a start and end time on every line and on most individual words.
The captions open as editable text over the playing video. Fix a word, retime a line, bleep or cut something, or mark words for emphasis - all of it reversible until you export.
The deliverable is a video, not a subtitle file. CaptionFX renders the caption layer into the frames of a new MP4 and leaves your original upload untouched.
Automatic speech recognition is good enough to save you the typing and not good enough to publish unread. On clean English audio a modern model lands somewhere around 95% of words correct. That sounds close to solved until you count it out: a 60-second clip is roughly 150 spoken words, so 95% means about seven or eight wrong words in every minute you post. They are rarely random, either - the errors cluster on exactly the words that matter most. Names, brands, product terms, numbers, places and jargon are the tokens a general-purpose model has the least evidence for, and they are the tokens your audience notices first.
That is the gap most caption tools leave open. They generate, then hand you a transcript in a text box, disconnected from the footage. You end up scrubbing the video in one window and reading a wall of text in another, trying to work out which line belongs to the moment you just heard. CaptionFX is built the other way round: the captions stay attached to the video while you work on them, so you read a line, hear the audio under it, and fix it in place.
The practical consequence is that a draft is useful even when it is imperfect. You are not proofreading a document - you are watching your own video with the captions on, stopping where something looks wrong, and correcting that one thing.
A transcript is text. Captions are text with timing, and the timing is what makes every downstream feature possible. CaptionFX keeps a start and end time on each caption line, and on individual words wherever the speech model measured them directly rather than estimating.
That distinction matters more than it sounds. A word whose timing was measured from the audio can safely anchor an operation that cuts or censors audio, because the boundaries are real. A word whose timing was interpolated cannot - acting on a guess produces a bleep that lands half a syllable late, which is worse than no bleep. CaptionFX tracks which words are which and refuses to export an edit that depends on timing it does not trust, instead of quietly shipping something misaligned.
Timing is also what lets you cut. Removing a word from the captions can remove the corresponding audio and video, re-time every caption after the cut, and leave you with a shorter clip whose captions still line up. That is only possible because the words carry real timestamps rather than a word count.
There are two ways to put captions on a video. A sidecar file (.srt, .vtt) rides alongside and is rendered by whichever player loads it - if it loads it at all. Burned-in captions, also called open captions or hardcoded subtitles, are drawn into the pixels of the video itself.
For social video the second one is the only reliable option. A clip uploaded to Reels, Shorts, TikTok, a LinkedIn feed, an embedded player or someone's re-share does not carry your subtitle file with it, and most of those surfaces autoplay muted. Captions that depend on the player being cooperative are captions that are frequently absent.
CaptionFX treats the captioned video as the product. The caption data, the styling and the edits are all working state in your browser; the thing you download is an MP4 with the words already in the frames. A subtitle file is an export intermediate here, never the deliverable.
CaptionFX routes speech to the model that fits the language it detected rather than pushing everything through one general-purpose English model. That routing is the whole reason Hindi and other Indian languages work here at a usable level instead of coming back as approximate phonetic mush.
Being specific about the numbers: clean, well-recorded English sits around 95% word accuracy. Clean Hindi is lower, roughly 85-90%. Code-mixed Hinglish - the register most Indian creators actually speak in - is harder than either, and CaptionFX treats that output as a draft that wants a review pass rather than something to publish straight through.
Published benchmark figures are measured on clean studio audio, so real recordings do worse than any number you will see quoted, including these. Background music, overlapping speakers, a phone mic at arm's length and a room with hard walls all cost you accuracy. This is the reason the editor exists and the reason it is the centre of the product rather than a feature bolted on afterwards.
Yes. The final download is an MP4 with the captions burned into the video frames, ready to post without a separate subtitle file. A subtitle file is only an intermediate step in the render here, not the product.
Yes, and this is the point of the tool. Captions stay editable in the browser until you export - text, timing, styling, bleeps, cuts and emphasis are all working state, and none of it touches the source video until you render.
On clean English audio, roughly 95% of words. Clean Hindi is around 85-90%. Code-mixed Hinglish is harder and should be treated as a draft to review. Published accuracy figures are measured on clean studio audio, so real recordings with music, background noise or several speakers will do worse.
MP4 and the common consumer video formats. CaptionFX inspects each upload before processing it and will tell you if a file has no usable audio track, which is the most frequent reason captioning cannot start.
Yes. A new account can create one project, generate captions, edit them and preview the styled result at no cost. Credits are spent when you download the clean exported MP4.
No. CaptionFX runs in the browser. Transcription and the final video render happen on the server, so your machine is not doing the encoding and you do not need a plugin or a desktop editor.