Transcribe a video to text — without uploading it
Open the video in the editor, run captions, and the words come back as text you can edit and export. The audio never leaves your device: the speech model runs inside your browser.
Short answer: open the ShrinkVideo editor, drop your video in, and choose Generate captions. A speech recognition model downloads once (about 45 MB), then runs on your own machine. You get a timed transcript you can correct, and you can export it as an SRT subtitle file or burn it into the picture.
Why this one doesn't upload your video
Every other free transcription service works the same way: you upload your file, their server transcribes it, you download the result. That means handing a recording — an interview, a lecture, a client call, a private video — to a company you have no relationship with, and trusting a retention policy you did not read.
There is no server here to hand it to. The speech model is Whisper, compiled to run in the browser through WebAssembly. Your browser downloads the model once, the way it downloads an image, and from then on the transcription happens on your CPU. You can watch the network tab while it runs and see that your audio never goes anywhere.
How it works, step by step
- Open the editor and drop in your video. Nothing uploads at this point either — the file is read straight from disk into the page.
- Choose Generate captions. The first run downloads the model: roughly 45 MB for the fast one, 90 MB for the more accurate one. That download is cached, so it only happens once per device.
- The audio is pulled out of the video and resampled to the 16 kHz mono the model expects, then transcribed in a background worker so the page stays responsive.
- You get a list of timed lines. Click any of them to jump the playhead there, and edit the text in place — automatic transcription is never perfect, and proper nouns are usually the first thing to fix.
- Export the result as SRT, or burn the captions into the video and export an MP4.
The honest limits
Worth knowing before you start, because none of this is hidden once you are halfway through a long file:
- English audio. The models in use are the English-only Whisper builds (
whisper-tiny.enandwhisper-base.en). Other languages will produce nonsense rather than a translation. - The first run downloads a model. 45 MB or 90 MB, once per device. On a metered connection that is worth knowing about in advance.
- It is not instant. Transcription runs on your own CPU rather than a datacentre GPU, so a long recording takes a while, and a phone takes considerably longer than a laptop.
- Clean audio transcribes far better. Crosstalk, heavy accents, music under speech and poor microphones all cost accuracy — the same as any speech recogniser.
- Expect to edit it. This is a first draft, not a finished transcript.
What you get in exchange is that the recording stays on your machine, there is no account, no queue, no per-minute quota and no watermark.
Transcript, captions or subtitles?
They are the same words in different wrappers, and it is worth being clear which one you need:
- A transcript is the plain text, usually without timings — for notes, search, or a blog post.
- Captions or subtitles are the same text with a start and end time on each line. That is what an SRT file holds, and what a player shows along the bottom.
- Burned-in captions are drawn into the picture itself, so they survive re-uploading anywhere. Burn subtitles into a video does that, and it is what you want for social platforms that ignore subtitle tracks.
FAQ
Is my video really not uploaded?
No, it is not. There is no upload endpoint on this site — it is static files plus WebAssembly. Open your browser's network tab before you start and you will see the model being fetched, and your own file going nowhere.
What languages does it handle?
English only. The models used are whisper-tiny.en and whisper-base.en, which are the English-only builds. Audio in another language will produce nonsense rather than a translation.
How accurate is it?
Good enough to be a first draft and not good enough to publish unedited. Clear speech from a decent microphone transcribes well; crosstalk, accents, background music and phone-quality audio all cost accuracy. Names and technical terms are usually the first thing you will correct.
How long does it take?
Longer than a cloud service, because it runs on your CPU rather than a datacentre GPU. The smaller model is noticeably faster and slightly less accurate; the larger one is the reverse. Phones are much slower than laptops.
Do I have to download the model every time?
No. It is cached after the first run, so later transcriptions on the same device start immediately.
Can I get a plain text transcript rather than subtitles?
Yes. The transcript panel holds the text and you can copy it out. If you want the timed version, export an SRT file instead.
Is there a length limit?
No fixed one. The practical limit is your browser's memory, which caps out somewhere between 2 and 4 GB for the video itself. Long files simply take longer.
Does it cost anything?
No. There is no account, no trial, no per-minute quota and no watermark. The site is supported by ads.