A burned-in caption is part of the picture. It cannot be switched off, lost in a re-encode, stripped by a platform, or left behind when somebody downloads your clip and posts it somewhere else. Send in a long video and every clip that comes back has word-by-word subtitles rendered into the frames, timed to the speech, with no sidecar file to attach and nothing to re-sync after a trim.
Real ClipSpeedAI output — every clip below was scored, cut, reframed to 9:16 and captioned automatically from the video above it.
Source video
15:27
Three steps and no subtitle editor at any point.
Paste a YouTube link or upload the file. There is no separate transcription step to run first and no SRT you have to source from somewhere else — the speech is read directly off the audio track of whatever you submit. A paid upload can run to two hours; the free demo takes sources under thirty minutes.
Timing is the part that decides whether captions read as professional or amateur. Every word gets its own timestamp rather than a block appearing for a whole sentence, which is what lets the caption highlight track the voice instead of trailing it. Filler words and dead air are cut out of the clip first, so the subtitle never displays the ums that were removed from the audio.
Pick from eleven caption looks, or pin one of them as your default in the Brand Kit alongside your brand colours and your logo, so the choice stops coming up on every export. The text is rendered into the final vertical frame at export, which means the captions are composed for the crop you are actually posting rather than scaled down from a landscape master and ending up half off screen.
Everything here exists because of a specific way subtitles go wrong.
A hardcoded subtitle is made of the same pixels as the rest of the frame. Re-encode it, download it, hand it to a client, upload it to a platform that has never heard of your subtitle file — the words are still there, in position, in sync. That permanence is the entire reason people search for this rather than exporting an SRT.
Traditional subtitles show a full line for two or three seconds. On short vertical video that reads as sluggish, because the viewer finishes the line and then waits. Highlighting one word at a time keeps the caption moving at the speed of speech, which is why the format took over short-form in the first place.
TikTok stacks a username, a description and a column of buttons over the lower right of every video. Subtitles dropped at the classic bottom-centre position land underneath all of it. Captions here sit inside the region that stays visible across vertical platforms, so you are not discovering the overlap after the post is live.
Style is not decoration on a muted platform; it is legibility. Heavy weights with a stroke survive busy footage, lighter treatments suit clean backgrounds. Choose the style once, save it in the Brand Kit next to your colours and your logo, and every later export inherits all three, which is how a feed of clips starts looking like one channel.
Filler removal and caption rendering are the same pipeline, not two tools stapled together. When a false start or a three-second silence is removed from the audio, the words come out of the caption track too. Bolt-on subtitle burners routinely leave that mismatch behind, and it is instantly visible.
Captions burned into a 16:9 master and then cropped to 9:16 lose their edges. Here the vertical crop and the speaker framing happen first and the text is composited onto the final frame, so caption size and position are correct for the shape you are actually publishing.
Clips cut from a running YouTube, Twitch or Kick broadcast come back with subtitles already burned in. The whole value of a live clip is posting it while the moment is current, and stopping to caption it by hand is exactly what destroys that window.
A caption is only as good as the clip boundary underneath it. Clips are cut on sentence boundaries, so the first subtitle starts on the first word of a thought and the last one completes it. Half a word on screen at second zero is the fastest way to look automated.
If burning captions is a step inside something larger — a content pipeline, a client delivery process — the same engine is reachable through a developer API and an MCP connector. Documentation is at the API docs.
Interview clips are dialogue with almost no visual event in them, so the words on screen are the whole hook. The podcast clip generator covers that source in full.
Ads autoplay muted in a feed. A hardcoded caption is the only version of your copy that still reaches someone who never turns the sound on.
The deliverable is a finished MP4 the client posts themselves. Attaching a subtitle file and hoping it gets used correctly is how captions go missing.
Busy backgrounds punish thin lettering. Stroked, high-contrast styles keep the text readable over explosions and UI clutter.
A definition or a step list is far easier to absorb as text on screen than as speech alone, especially when the viewer is on a train.
With nobody on camera, captions carry the entire performance — more on that on faceless video clips.
Soft subtitles live in a separate file — an SRT or a VTT — that a player reads alongside the video. The viewer can toggle them, translate them and restyle them, and none of that survives being separated from the file. Upload the MP4 without the sidecar and the subtitles simply do not exist.
Platform captions are generated by TikTok, Instagram or YouTube after you upload, and they live inside that platform only. They are convenient and they are trapped. Download your own video back from TikTok and the caption layer either comes with a watermark or does not come at all, which is the exact moment most people discover the problem.
Hardcoded captions are drawn into the video frames during export. There is no toggle and no file, because there is nothing to attach — the words are as much a part of the image as the person speaking. That is the version that travels.
The common workflow is to cut a clip, post it to TikTok, then push the same file to Reels and Shorts a day later. If the captions came from TikTok, that second and third post go out silent and unreadable, or carry a competitor logo across the frame. Everyone learns this once, usually from a clip that was doing well.
The second failure is handoff. You send a file to a client, a colleague, an editor, or a scheduling tool. A sidecar SRT has to travel with it, be recognised at the other end, and be re-synced if anyone trims a second off the front. In practice at least one of those three steps fails, and the version that gets published has no captions at all.
The third is the repost economy. Clips get downloaded and reposted by other accounts, and a burned caption goes with them. Whether you welcome that or not, it decides whether your words survive the trip — an SRT never does.
Burning is wrong for plenty of jobs and pretending otherwise would be dishonest. On a forty-minute YouTube upload, viewers expect to switch subtitles off, and a permanent caption strip across a lecture is an irritation rather than a feature. Long-form belongs to soft subtitles.
Multiple languages need soft tracks too. A single burned track can only ever be one language, whereas a set of sidecar files lets a viewer choose. To be clear about our own scope: we transcribe the language that was spoken and burn that. There is no translation or dubbing in this product, so multi-language delivery is not something to plan around here.
Formal accessibility work is the third case. Where a standard requires the viewer to control caption presentation — size, contrast, on and off — burned text does not satisfy it, because the whole point is that the user cannot change it. Short-form vertical clips are a different context, and there burned captions are the accessible option in practice, since the platforms give viewers no reliable alternative.
Look at it on a phone rather than the monitor you exported from. Text that feels comfortable at desktop size can be marginal at arm's length on a bright train, and that is the only viewing condition that decides anything.
Watch the first second with the sound off. If the opening word is missing, clipped or arriving late, fix the clip boundary instead of the caption — the boundary is nearly always the real cause, since text cannot start cleanly on a sentence that was already halfway through.
Scan for names and jargon. Speech models handle ordinary language well and mangle proper nouns, product names and industry acronyms far more often, and those tend to be the words the clip exists for. Two seconds of attention each time is a fair price for not publishing a misspelled surname.
Then check the corners. Put your thumb where the platform stacks its buttons and confirm nothing important is hiding underneath. Doing this once on your first export usually settles the question for everything you publish afterwards.
Contrast first. Text with no stroke or shadow disappears the moment the footage behind it turns light, and it will do it on exactly the frame that mattered. A heavy weight with an outline is unglamorous and it works everywhere.
Then line length. Word-by-word styles keep only a few words on screen at once, which suits a phone held at arm's length. A full sentence shrunk to fit the width of a vertical frame is technically present and practically unreadable.
Then timing. A caption that lags the audio by even a fraction of a second is worse than no caption, because the viewer notices the seam instead of the content. Word-level timestamps rather than sentence blocks are what keeps the highlight sitting on the syllable being spoken.
Observations from operating the pipeline in production — not general advice.
A brand kit in a design tool takes a typeface upload, and this one deliberately does not, which is worth knowing before you plan a house style around it. What gets saved is a named caption style out of a fixed set, your two brand colours as hex values, and a logo with a corner and a size. The reason is legibility rather than laziness: every preset has had its weight, stroke and highlight behaviour tuned to survive a 9:16 frame viewed on a phone, and an arbitrary uploaded face would arrive with none of that tuning and fall apart on the first bright background it met.
Bottom-centre is where a television subtitle belongs and it is close to the worst position available on a vertical clip. TikTok stacks a description and a column of buttons over the lower right, Reels puts its furniture somewhere slightly different, and Shorts differs again. Because the same file usually goes to all three, the usable band is the intersection of what every platform leaves alone rather than what any one of them permits, which is why captions here sit noticeably higher than broadcast convention would put them. Looking a little conservative beats losing a word behind a follow button.
The expensive stage of captioning is not drawing the text, it is transcribing the audio and pinning a timestamp to every individual word. That result is cached against the project, which is why choosing a different style and rendering again never sends the audio back through the speech model. In practice that should change how you pick one: instead of judging from a preview sitting on a clean background, export the same clip twice against your own worst frame — a blown-out window, a busy game HUD — and compare those, because the only footage the decision applies to is yours.
That path means a transcription tool, a correction pass, an export, and then a separate upload on every destination that accepts sidecar files — several of which do not. It is more steps for a weaker result on short-form, where nobody is toggling subtitles anyway.
They are free and they are quick, and they tie you to one editor. Text added inside a platform app tends not to survive being exported for use elsewhere, and the auto-generated timing is usually per-line rather than per-word. Fine for one post, painful as a workflow.
A professional editor gives you total control over every word, and it costs a real chunk of an afternoon per finished minute once you include correction and styling. That trade makes sense for a flagship piece. It does not make sense fifteen times a week.
If you already have one finished thirty-second file and you only want text stamped onto it, a dedicated burner is the simpler tool and you should use it. This is a clipping engine: it takes long video, finds the moments, cuts them, reframes them and captions them. The captions are the last step of a bigger job — see what comes out before deciding which shape of tool you need.
Run one video through and check the captions on a phone. Legibility and timing are the two things you can only judge by looking.
⚡ Get A.I Clips — $1 trial
