How speech-to-text runs inside a browser tab, with no upload
· 7 min read
Most caption tools upload your video to a server, transcribe it on a rented GPU and send the text back. AI Caption Cut does the transcription inside the browser tab, on your own processor. People reasonably ask how that is possible, and whether it is true. This is the full answer, with the numbers from our own code — and how to check it yourself in a minute.
The model: Whisper, small enough to download
The speech-to-text model is OpenAI’s Whisper, the open model most caption tools use on their servers too. The difference is size: we run the "base" checkpoint, which is small enough to download once into the browser and fast enough to run on a phone. English uses the English-only base model, which is measurably better at English than the multilingual one of the same size. Every other language uses multilingual base. Hinglish uses a base model fine-tuned on about 550 hours of Indian-accented speech, roughly an 80 MB download.
The model is downloaded the first time you generate captions and kept in the browser’s cache, so the second video starts immediately and works offline.
What happens when you press Generate captions
Nothing below leaves the tab. The steps run in a background worker so the page stays responsive.
The audio is pulled out of the video. The browser decodes the video file from your disk and keeps only the sound, resampled to 16,000 samples per second, mono — the input Whisper was trained on.
It is cut into 30-second windows. Whisper reads 30 seconds at a time. Windows overlap by 5 seconds so a word on a boundary is not cut in half, and the overlaps are stitched back together.
The model runs in WebAssembly. The network runs through ONNX Runtime compiled to WebAssembly, using a quantised (8-bit) copy of the weights that is a fraction of the full size and runs at near-native speed on an ordinary CPU.
Words get times. The English-only model returns a time for every word. The multilingual model’s word times are unreliable, so it returns phrase times and the word times are worked out inside each phrase — accurate enough for karaoke highlighting.
The bug every Whisper app has to fix
Left alone, Whisper sometimes gets stuck repeating one token until the window ends — "the company’s company’s company’s…" two hundred times. The real speech in that window is never transcribed, so the captions are not just wrong, they are missing.
Measured on 75 seconds of a real Hinglish podcast that triggered it: 814 words at 97% repetition came down to 102 words at 6%, and transcription took 35 seconds instead of 154, by penalising repetition in the multilingual models. The English-only model does not get the penalty — on clean English it suppressed words that legitimately repeat, turning "gone in a day" into "gone in day".
That is the kind of trade-off nobody sees in a cloud tool, and the reason to measure on real speech rather than a demo clip.
Why bother doing it on the device
Privacy is the obvious reason: a video that is never uploaded cannot be kept, leaked or used for training. The less obvious one is cost. A cloud tool pays for GPU time by the minute, which is why their plans limit video length or sell credits. Here the processor is yours, so a two-hour podcast costs us the same as a ten-second Reel — nothing — and there is no length limit. See caption long videos.
The trade-off is honest too: base is a smaller model than the largest cloud ones, and it is slower on an old phone than on a laptop. For clear speech it is accurate, and every word stays editable.
How to check that nothing is uploaded
Do not take our word for it. Open your browser’s developer tools (F12 on a computer), go to the Network tab, and generate captions for a video: your file appears in no request. Or load the caption generator once, switch your connection off, and caption a video anyway. A tool that uploads your footage cannot do that. More on the privacy side: caption a video without uploading it.
Questions
Which speech-to-text model does AI Caption Cut use?
OpenAI Whisper base: the English-only checkpoint for English, multilingual base for other languages, and a base model fine-tuned on Indian-accented speech for Hinglish. All run in the browser through WebAssembly.
Does browser speech-to-text work offline?
Yes, once the model has been downloaded the first time. It is cached by the browser, so later videos can be captioned with no connection at all.
Is on-device transcription as accurate as cloud transcription?
For clear speech it is close; the largest cloud models are better on heavy accents and noisy audio. The difference is that every word is editable, the video is never uploaded, and there is no limit on length.
Try it on your own video — nothing is uploaded: Free caption generator
Keep reading
All articles →Can ChatGPT or Claude edit videos?
On their own, ChatGPT and Claude cannot cut, caption or export a video file — they write text and code.
10 November 2026 · 5 min readHow to remove “um”, “uh” and long pauses from a video automatically
Auto-cut in AI Caption Cut removes long pauses and filler words (um, uh, erm) in one undoable edit.
10 November 2026 · 4 min readWhy your exported video plays as a black screen (and the sound still works)
A black picture with working sound almost always means the player refused the file before the first frame — not that the file is broken.
10 November 2026 · 6 min read
Related free tools
Try it on your own video
Nothing to install, no account, and your video never leaves your device. Unlimited free exports, and a one-off payment when you want the watermark gone.
Caption my video — freeUpdated