Most people meet your video with the volume already down, which means the text on screen is doing the work the audio was supposed to do. Captions here are generated from the speech, timed word by word, styled eleven different ways, and rendered into the video itself rather than attached beside it.
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
Captioning is not a separate task here. It happens as part of producing the clip.
Paste a link or upload the file. The audio is transcribed as part of processing, and there is no separate captioning step to start, no transcript to approve first and no timing file to import. If you only want to sample the output, the free demo runs on anything under half an hour.
The transcript is aligned to the audio at the level of individual words, not blocks of a few seconds. That granularity is what allows a word to light up exactly as it is spoken, and it also means the text never drifts behind a fast talker or hangs on screen after a sentence has ended.
Eleven styles are available, from clean sans-serif blocks to the bold high-contrast look that dominates short-form feeds. Switch between them and re-render to compare. The captions are drawn into the frames of the exported file, so what you download is what every viewer sees on every platform.
Captions look simple until you have shipped a hundred of them. These are the details that decide whether people keep watching.
Traditional captions push two lines at a time and swap them every few seconds, which is fine for a film and dull in a feed. Aligning each word individually lets the caption track the rhythm of the speech, so the emphasis on screen matches the emphasis in the voice. The active word is highlighted as it lands.
A caption competing with footage behind it needs weight, contrast and an outline or shadow to stay legible. Every style here is designed against that constraint rather than being a font choice with a colour picker attached, which is why they hold up on a phone in daylight.
The text becomes part of the image. Nothing to upload alongside the video, nothing that a platform can strip, nothing that breaks when the clip is downloaded and reposted by somebody else. Wherever the file travels, the captions travel inside it.
Verbatim transcription is a terrible caption. Real speech is full of "um", "you know", restarted sentences and half-finished clauses, and putting all of that on screen makes an articulate person look scattered. Those are removed from the audio and never reach the caption track.
The bottom of a vertical video is covered by the caption box, the handle, the music strip and the button column, and text placed there is simply lost. The styles position the text in the band of the frame that stays visible, which is a boring detail that decides whether the captions were worth generating at all.
There is no checkbox to remember. Any clip the engine produces comes back captioned, including clips cut from a live YouTube, Twitch or Kick broadcast in real time. Consistency here is worth more than configurability — the clip you forgot to caption is the one that underperforms.
Reframing widescreen to 9:16 while burning in captions is where a lot of tools produce something with the words half off the edge. Because the crop and the caption layout are decided together, the text fits the final frame rather than the original one.
The Brand Kit keeps your fonts, colours and logo applied across exports so a run of clips looks deliberate. A viewer scrolling past their third clip of yours in a week should recognise the look before they read the name on the account.
Nothing but speech carries the clip, so the caption is effectively the entire visual argument for staying.
Chaotic audio, overlapping voices and game sound make text the only reliable channel. Picture-in-picture keeps the face and the captions both readable.
A product point delivered on mute has to land in text or it does not land. Captions are the difference between a scroll and a stop.
Technical terms are easier to absorb read than heard, particularly for anyone watching in a second language.
On-screen text is the baseline for viewers who are deaf or hard of hearing. The video subtitle generator page covers what burned-in text does and does not satisfy.
Commutes, offices, waiting rooms. Your audience is frequently somewhere they cannot turn the sound on even if they want to.
Short-form feeds autoplay silently by default, and plenty of viewers never unmute anything. The practical result is that your opening line is read before it is heard, if it is heard at all. A clip with no text on screen is asking a stranger to make an extra decision in the first second, and most of them decline.
This changes what a caption is for. It is not a transcript service or an accessibility afterthought bolted onto the export — it is the primary interface of the video. Once you accept that, the styling decisions stop being cosmetic. Size, weight, contrast and position are the difference between someone reading your first sentence and someone scrolling past it.
It also changes what you write and how you speak. Creators who caption everything tend to front-load the interesting clause, because they can see it appearing on screen at second zero rather than arriving in the middle of a sentence nobody heard.
Contrast comes first. Text over video is text over an unpredictable background, so the styles rely on heavy weights with outlines or plates behind them rather than thin type that vanishes the moment the shot brightens. A caption that is legible in a dark studio and invisible against a window is not legible.
Then pace. Words appearing faster than they can be read produce a subtly stressful clip that people leave without knowing why. Because the timing here is tied to the actual speech rather than to fixed intervals, the caption naturally slows when the speaker does, and the reading load stays proportional to the delivery.
Then position, which is the one most people get wrong. Every platform stacks its own interface over the lower portion of a vertical video, and captions parked in that zone are covered by the description, the audio strip and the row of buttons. Keeping the text above that band is unglamorous and it is the single highest-return caption decision you can make.
Finally, restraint. Every word on screen at once is a wall; one word at a time can feel jittery on a slow talker. The styles here sit between those extremes, showing a readable phrase with the current word emphasised, which reads as movement without becoming noise.
TikTok, Instagram and YouTube all offer to caption your upload, and those captions are better than nothing. They are also small, styled to the platform rather than to you, positioned wherever the platform decides, and frequently switched off by default depending on the viewer's settings. You are handing your most important design decision to somebody else.
They also do not travel. A clip downloaded from one app and re-uploaded to another arrives with no captions at all, which is exactly the workflow most creators actually use. Burned-in text is immune to this: the words are in the frames, so they survive every download, repost and cross-post.
The last problem is timing. Platform captions typically chunk text into blocks that appear and disappear on a fixed cadence. That is serviceable for comprehension and useless for pace, which is what you want captions doing in a feed.
Pick one style and stay with it. Rotating through all eleven because each looks good in isolation costs you the thing captions can quietly buy — recognition. Somebody who has scrolled past your clips twice should register the third as yours before they read the account name, and consistent type does more of that work than a logo does.
Check the first clip from every new recording setup. Not every clip, just the first one after you change microphone, room or guest, because those are the changes that move transcription accuracy. Thirty seconds of checking catches the mistake that would otherwise repeat across fifteen exports.
Turn the platform's own caption feature off when you upload. If TikTok or Instagram lays its automatic text over a clip that already has words rendered into it, the two sets overlap and the result looks careless. This trips up almost everyone the first time.
And read your captions rather than watching them. Mute the clip, look only at the text, and see whether the opening line still makes an argument for staying. If it does not read well as a sentence on a screen, the clip is relying on delivery that a silent viewer will never receive.
Speech recognition is very good and it is not perfect, and it fails in specific, predictable places. Proper nouns are the worst offender — brand names, product names and unusual surnames are exactly the words a model has the least evidence for and exactly the words you least want mangled. If your show is full of them, check the first clip before you build a habit on the output.
Heavy background music, two people talking over each other, and low-quality recordings all reduce accuracy. So does very fast speech with a strong regional accent. None of these are unique to this tool; they are the boundaries of automatic transcription generally, and any product claiming immunity from them is overselling.
One thing worth being explicit about: the captions are in the language being spoken. Speech is transcribed, not translated, so a French video produces French captions. If you need a different language on screen than the one in the audio, that is a translation workflow and this is not the tool for it.
A word runs off the edge of the frame. Caption size is stored per clip, and the layout maths is calibrated around the default value rather than around whatever the slider was last dragged to. Above the default the engine allows fewer characters on each row automatically, so enlarging the type gives you bigger words on more rows instead of a line bleeding off the side. When one clip in a batch overflows and the other fourteen are fine, open that clip and read the size control before you suspect the renderer.
A word appears that nobody said. Speech models label audible events as well as speech, and a strong music bed can arrive in the transcript as a literal word. For example, a scored intro under a cold open can put the word "music" on screen, and the giveaway is the same phantom word surfacing in several unrelated clips cut from one recording. Mixing the bed lower under the voice removes it at the source, which is the only place it can be removed.
The bottom line of text is hidden. You almost certainly checked the clip in a plain video player, which carries none of the description, audio strip and button column that a short-form app stacks over the lower third. Post one clip to a private account and look at it there instead. Doing that once tells you where the real safe band sits on the platform you publish to, and you will never have to think about it again.
One name is wrong in every single clip. That is transcription rather than captioning, and it does not vary between clips from the same recording — the model will mishear a surname identically fifteen times. Run thirty seconds containing the name through the demo, see exactly what comes back, and then decide whether to introduce it differently on camera or accept it. Learning this from one clip costs nothing; learning it from a published batch costs a correction.
A long-form upload that somebody sits down to watch is the clearest case against. Imagine a forty-minute recorded lecture going onto a course platform or a television app: that viewer wants the option to switch text off, and a caption track gives them the toggle while also handing the platform something machine-readable to index. Rendered-in words do neither. For that deliverable a separate caption file is simply the better instrument, and the subtitle generator page maps out exactly where the line between the two falls.
Formal accessibility work is the second case. A track that assistive software can read aloud, resize and disable is a different artefact from words painted into a frame, and if you are working to a published standard, a captioning service with human review is what actually satisfies it. Treating on-screen text as a substitute for that is the kind of shortcut that gets discovered during an audit.
The third is any job that needs the words in a language other than the one being spoken. Speech is transcribed and never translated here, so a translation workflow has to sit entirely outside this tool rather than being configured inside it.
And if you honestly publish one video a month, the alternative worth using is the caption feature already sitting in your phone editor. It has become competent. The argument for an engine is volume and consistency across a run of clips, and at one clip a month you have neither of those problems to solve.
Automatic captioning is a transcription problem plus a layout problem, and both are solved well enough now that the output is usually indistinguishable from careful manual work. Where it is not is emphasis that depends on meaning rather than on sound: a caption held back half a beat so a punchline lands after the setup, a word deliberately oversized for a joke, text that moves with a gesture. Those are authored decisions and no timing model is going to make them on your behalf.
The same is true of any typography beyond dialogue — a lower third naming a guest, a figure pulled out and animated because it is the entire point of the clip, a callback appearing in a corner thirty seconds after the line it refers to. That work belongs in an editor and it always will.
The division most people settle on is unglamorous and it holds. Say you cut fifteen clips from one recording and one of them is going behind paid promotion: let the engine caption all fifteen so they exist and match each other, then take that one into an editor and finish it properly. Trying to hand-finish every clip is the habit that dies in week three, and captioning nothing while you wait for time to do it properly is the worst outcome of the three.
Observations from operating the pipeline in production — not general advice.
Close-mic talking-head audio reflects off a desk or a wall and returns to the microphone roughly 50 to 200 milliseconds behind the voice, and a speech model can hear that reflection as a second utterance. Rendered literally it ghosts the caption: THAT THAT, WAIT WAIT. The rule that removes it is deliberately narrow — a repeated word is dropped only when it directly follows the first and lands within four tenths of a second — because natural speech that repeats a word almost always has another word between the two. Widen the window and you start deleting somebody genuinely shouting "go go".
Every style shares one chunking rule rather than each preset carrying its own: never more than two rows on screen, never more than three words on a row, a budget of roughly 16 characters per row, and a word is never broken across rows. The part worth knowing is what happens when you enlarge the type, because the character budget contracts in step with it. A larger font therefore buys you fewer words per row rather than a line running off the frame, which is why pushing caption size up changes the rhythm of the text as much as its scale, and why one clip with the size raised reads noticeably choppier than the rest of a batch.
The moving highlight picks the longest meaning-carrying word in each group and steps over the function words — the, of, and, to, pronouns, auxiliary verbs. When a group happens to contain nothing but those, the highlight is suppressed for that group instead of landing on "OF", because a colour parked arbitrarily on a preposition reads as a glitch to anyone watching closely. In practice this is what makes the emphasis look considered rather than metronomic: it follows the words carrying the sentence and goes quiet across the connective tissue between them.
Manual captioning on a timeline is among the least rewarding tasks in video work. A minute of speech takes many minutes to type, time and position, and it has to be redone every time you change a cut. The result can be excellent. Almost nobody sustains it past the first few videos.
Transcription services will give you accurate text with timestamps, which is genuinely useful if you want a document. It still leaves you to import the file into an editor, style it, position it and render — the part that actually takes the time. Here the styled, positioned, rendered result is the output.
Phone editors have decent auto-caption tools now, and if you are producing one clip that is a fine place to do it. The difference shows at volume: fifteen clips means fifteen sessions of importing, captioning and exporting on a small screen, and by clip nine the styling has quietly drifted.
Premiere and Resolve can caption automatically and give you total control over the result. They also require the project to exist, which means you have already cut the clips. If you are generating the clips anyway, captioning them in the same pass removes an entire round trip — see the clip generator for the rest of what happens in that pass.
Play the result with the volume all the way down, which is how most of the feed will meet it anyway. If the point still lands in silence, the captions are doing their job; if you catch yourself reaching for the volume to follow the sentence, the text on screen is not carrying enough of it yet.
⚡ Get A.I Clips — $1 trial
