Upload a video and get a timed transcript back in minutes, ready to style as burned-in captions or keep as text. No listening line by line, no manually dragging timestamps, just a script you proofread once before it ships with the video.
Nobody watches video with the sound up anymore, not on a phone anyway. Social feeds autoplay muted, coworkers skim meeting recordings at their desk, and viewers who are deaf or hard of hearing depend on the words on screen entirely. Typing and timing every line by hand takes nearly as long as the video itself runs. Automatic transcription turns that hour of clicking into a few minutes of review.
Most video today is watched somewhere sound is inconvenient: a bus, an open-plan office, a queue at the pharmacy. Feeds default to muted playback, so the first three seconds either earn attention with what is on screen or lose the viewer before the audio gets a chance. A caption stopped being a nice-to-have a while ago; on a muted feed it is the primary way most people receive the first line of dialogue.
Accessibility is the other half of the case, and it is not optional for anyone building an audience past a handful of friends. Viewers who are deaf or hard of hearing cannot follow a video without captions, and plenty of hearing viewers rely on them too, on a noisy commute or reading a second language faster than they parse it by ear. A video without captions excludes part of its own audience before it has said a word.
Retention tends to follow the same logic. A viewer who can read along stays through a cut they might have scrolled past in silence, and platforms that measure watch time notice patterns like that. None of this needs a claim about a specific percentage lift, only the plain fact that captions give a muted or hard-of-hearing viewer a reason to keep watching instead of a reason to leave.
Typing captions by hand means listening to a clip, writing the line, listening again to time it against the audio, and repeating that for every sentence. A ten-minute video can easily cost an hour this way, most of it spent scrubbing back and forth over the same three seconds. Automatic transcription listens once and returns a timed transcript in roughly the time it takes to make coffee.
Speed is the obvious win; accuracy is the honest caveat. A clean recording of one person speaking clearly comes back close to perfect. Add a strong accent, a technical term, a brand name spelled in an unusual way, or a joke that hinges on a homophone, and the transcript will need a human eye before it goes out. Automatic transcription replaces the typing, not the proofreading, and treating the output as a finished script rather than a very good first draft is the most common mistake people make with it.
Punctuation is where the gap shows up most. A model guesses where a sentence ends from pacing alone, and a fast talker who runs clauses together can produce a transcript that is technically accurate, word for word, and still reads like one long run-on paragraph. Reading the captions back at normal speed, not just checking words against the audio, catches most of what a skim misses. On Picasso IA the transcript comes back inside the Toolkit, where the same pass also lets you edit and style it.
Pro tip: run transcription on the raw recording before you edit the video, not after. A rough cut with jump cuts and trimmed pauses confuses the timing, while untouched audio gives the model one continuous voice to follow, and you can trim the caption track along with the footage afterward.
Naming a caption style upfront saves a round of restyling later, because the same transcript can be dressed five different ways and each one reads differently depending on where the video will actually be watched.
| Style | What it looks like | Fits best |
|---|---|---|
| Minimal single-line | one short line, small type, bottom third | talking-head interviews and podcasts |
| Bold word-highlight | large text, one word emphasized as it is spoken | short-form social clips |
| Classic subtitle bar | two lines, dark background strip, centered | tutorials and screen recordings |
| Large mobile-first | oversized centered text, high contrast | vertical video watched muted |
| Lower-third caption | compact box in a corner, does not cover the frame | webinars and talking-head with slides on screen |
For foreign-language dub prep, keep an unstyled export around even after picking a look for the published version. A translator needs the plain timed transcript, not a stylized overlay, and that clean text can seed an AI voiceover once translated.
Transcription follows whichever voice is speaking at each moment, but it does not automatically label who is talking. A two-person interview comes back as one continuous transcript that reads correctly line by line; if you want the final captions to show who said what, that speaker tagging is a manual pass during review, not something the transcript adds on its own.
Accents are handled well across a broad range of clear, well-recorded speech, but a heavy regional accent, mid-sentence code-switching, or a very fast dialect will lower accuracy in specific stretches. A quick read-through, not a full re-listen, catches most of what those sections miss, since errors tend to cluster in short bursts rather than spread evenly through the transcript.
Background noise and overlapping speech are the two conditions that pull accuracy down fastest. A quiet room with one voice at a time gives near-perfect results; a busy cafe, music under the dialogue, or two people talking over each other force the model to separate voice from noise before it can start guessing words. This is less a defect to route around than a reminder a decent microphone still matters more than any software downstream of it.
The whole pass from upload to export usually takes ten to fifteen minutes for a five-minute video, and most of that time goes to proofreading rather than waiting on the model.
Credits are only spent on the generation step, and you can check what a run costs on the pricing page before committing a longer video to it.
An honest tool page says where the tool breaks. Automatic transcription removes the typing, not every source of error, and five limits show up often enough to plan around.
Tools built specifically around transcription, covered on the Descript comparison page, go deeper into audio editing than a general catalog does; the tradeoff is that the same Picasso IA account also covers restyling, upscaling and the rest of the model catalog once captions are done.
For a single speaker recorded clearly, close to word-perfect, with the occasional miss on a proper noun or an unusual term. Accuracy drops as conditions get harder: background noise, overlapping voices, heavy accents and fast pacing all introduce more errors, but they cluster in a handful of lines rather than spread evenly through the transcript. A read-through before publishing catches nearly everything worth catching.
Yes, for anything you plan to publish under your name. Automatic transcription is a fast first draft, not a final script, and a wrong word in a caption is more visible than the same slip would be in speech, since a viewer can pause on it. Budget a few minutes to read the transcript at normal speed against the audio, which is where most missed errors get caught.
Not automatically. It follows whichever voice is speaking at each moment and transcribes it correctly, but it does not label who is talking on its own. If you need the final captions to show speaker names, that tagging happens during your review pass, after the base transcript comes back, rather than as part of transcription itself.
Reasonably well across a wide range of accents in clear, well-recorded speech, though accuracy dips for very heavy regional accents, mid-sentence code-switching between languages, or unusually fast dialects. Those errors cluster in short stretches rather than spreading through the whole transcript, so a targeted read-through of the harder sections usually cleans it up.
A timed transcript matched line by line to your audio, ready to review as text and then style as burned-in captions, or keep as plain text to reuse in a description, a blog post or a translation. You are not choosing a file format up front; you get text you can edit, restyle and export in whatever shape the next step needs.
Yes, but expect more corrections than a clean voice-only recording. Music under dialogue makes the model's job harder, since it has to separate voice from the bed before transcribing the words, and quieter lines are the ones most likely to need a fix. Lowering the music under spoken sections, if that is an option, meaningfully improves the first pass.
For a five-minute video, roughly ten to fifteen minutes end to end, most of it spent reading the transcript rather than waiting for it. The generation step itself takes a small fraction of the video's own runtime; the review and styling afterward is where the real time goes, and that time scales with how clean the original audio was.
Both. Once the transcript is reviewed and styled, you can export a video with the captions baked into the frame, ready to post as is, or keep the timed text separately to reuse elsewhere, like a translated version or a written recap of the video.
New accounts receive free credits, enough to transcribe and style a video or two and judge whether the workflow saves real time before spending anything further. After that, generation costs credits, and current plans are listed on the pricing page. There is no separate captioning tier; the same credits also cover image, video, audio and 3D generation across the catalog.
Speech-to-text models built for timed transcription are the right family, listed alongside the rest of the catalog's 488 models rather than tucked into a separate captioning product. If you are new to the workflow, start with the default transcription model in the Toolkit and only compare alternatives once you have a baseline.
Your next video already has audio worth captioning. Run it through the Toolkit and see how much of the transcript survives without a single fix.
