Captions that never see a server

How to caption a video without uploading it

Four steps, and the video never leaves your machine at any of them. Speech-to-text runs as WebAssembly in the browser tab, so the only things that cross the network are the page itself and the model file — never your footage.

1. 1

   Open the editor in any modern browser. Go to aicaptioncut.com in Chrome, Edge, Safari or Firefox, on a laptop or a phone. There is nothing to install and no account to create — the page is the app.

2. 2

   Add your video from disk. Drop the file onto the page or press "Choose a video". The browser reads it straight off your disk into memory. Watch the network tab while you do it if you want: there is no upload request, which is also why a 4 GB file starts as fast as a 40 MB one.

3. 3

   Generate the captions on your own CPU. Pick the spoken language and press Generate captions. A Whisper model (about 80 MB) downloads once from the CDN and is cached by the browser — that is the only fetch. From then on transcription runs locally, and it works with the network disconnected.

4. 4

   Export, still locally. The MP4 is rendered in the tab and saved straight to your downloads folder. Nothing is stored on a server, so there is nothing to delete afterwards and no retention policy to read.

How it actually works

When you open a video, the browser reads it from disk. It is never copied to a server, and on Chromium browsers it is not copied into browser storage either — the app keeps a handle to the file where it already lives.

The audio is decoded locally and passed to Whisper, OpenAI’s open-source speech-to-text model, running as WebAssembly in a worker thread. The model weights download once from a public CDN and are cached; your audio is never part of that request.

Export encodes frames the browser has drawn itself. The finished MP4 is assembled in memory and handed to your downloads folder. At no point does a byte of your footage cross the network.

You do not have to take our word for it

Open your browser’s developer tools, switch to the Network tab, and generate captions. You will see the model weights download on the first run, and nothing at all after that. No upload, no background sync, no telemetry carrying your video.

That is the difference between a privacy policy and an architecture. A policy is a promise about what a company chooses to do with your files. This is a statement about where the files went — and they did not go anywhere.

Why this matters more than it used to

Uploading a video to a cloud editor means agreeing to that company’s terms, which increasingly cover how your content may be used to train models. Several tools now sell training-data exclusion as an enterprise feature — meaning that keeping your own footage out of a training set requires a custom contract and a large invoice.

Unpublished footage is the case that really matters. A client’s rough cut, an unreleased product, a face that has not consented to being in a dataset. If the file never leaves the machine, none of that is a question you have to ask.

What it costs you

The honest trade: running the model locally is slower than a data centre with a GPU. The first run downloads about 80 MB of model weights, and a two-minute clip takes roughly two minutes to transcribe on an average laptop. After that download it works offline.

A cloud service will be faster. It will also have a copy of your video.

Questions

Related free tools