Most online caption tools work by uploading your video to a server, processing it there, and sending the result back. That is simple to build, but it means your footage leaves your device, long uploads come before anything happens, and the provider pays for every minute of processing, which is why so many of those tools end up with watermarks, paywalls, or tight limits. AutoCap is built the other way around: the heavy work happens in your browser. This guide walks through each stage, says exactly what runs where, and ends with a way to verify the claims yourself.
The short version
- Speech recognition runs on your device. A Whisper speech model is downloaded into the browser and transcribes the audio locally.
- Captions are drawn onto the video frames in your browser, and the exported video or subtitle file is created in your browser.
- Your video and its audio are not uploaded to AutoCap servers in the live product.
- A small amount of metadata does reach the servers, listed in full in the table further down.
Step 1: loading the video
When you choose a file, the browser reads it from disk and plays it through an ordinary video element. Nothing is sent anywhere at this point. The studio reads the duration and dimensions, builds the preview, and keeps a recovery copy of your last video in the storage that belongs to your browser (IndexedDB), up to a size limit. That copy exists so a refresh, a crashed tab, or an accidental close does not cost you your work, and it stays on your device.
Step 2: turning speech into text
The audio track is decoded in the browser and converted to the 16 kHz mono format that Whisper models expect. A Whisper model, packaged for the browser in ONNX format and compressed to 4-bit weights, then runs in a background worker so the page stays responsive while it works. If your browser can use WebGPU, the model runs on your graphics processor. If not, it falls back to WebAssembly on the CPU, which works everywhere but is slower.
The model files are downloaded from Hugging Face the first time you need them: about 300 MB for the standard small model and about 760 MB for the Accurate mode. Browsers normally keep these files, so later visits should not need to download them again. The size is the price of running recognition locally. In exchange, the transcription cost is paid by your device instead of by a server, which is what lets the tool stay free and keep your audio on your machine.
Honest trade-off
Local processing is slower on old phones and low-memory devices than a data-centre GPU would be, and the very first run includes the model download. After that, speed depends mostly on your hardware and on which model you choose. The companion guide on Fast versus Accurate models explains the choice.
Step 3: editing and styling
The transcript and the timing data live in the page while you work. Fixing a word, regrouping captions, nudging the sync offset, or switching presets updates the preview immediately, because nothing has to be sent to a server and processed again. The preview is drawn on a canvas by the same code that produces the export, so what you see while editing is what you get in the file.
Step 4: exporting
To export, the video is played through that canvas renderer and the result is recorded with the browser MediaRecorder. When the browser can record MP4 directly, it does. Otherwise it records WebM and, if you asked for MP4, converts it with ffmpeg compiled to WebAssembly, which also runs in your tab. Very large recordings skip that conversion because converting them inside a browser tab can run out of memory. SRT and VTT downloads are generated directly from the caption list and involve no video processing at all.
What does reach the servers
AutoCap does run a small backend, used to keep usage limits fair and to store caption jobs. This is everything it receives and why.
| Data | Why it is needed | Notes |
|---|---|---|
| A random device ID kept in your browser storage | To count usage per device without accounts | It is not linked to a name or email address. AutoCap has no sign-in. |
| Your IP address and browser user-agent, as salted hashes | To recognise repeat abuse | The database stores the hashes, not the raw values. |
| Caption jobs: file name, video duration, caption style, caption text | To count jobs and apply fair-use limits | The video and audio themselves are never part of a job. |
| Usage events: durations and timestamps | To apply daily limits | Counts only; no media. |
Some requests go to other companies rather than to AutoCap. The speech models come from Hugging Face. The ffmpeg engine used for MP4 conversion is loaded from the jsDelivr content network. Fonts are loaded from Google Fonts. If you accept advertising cookies, Google AdSense may serve ads. The privacy policy lists these in the same way.
How to check this yourself
- 1Open your browser developer tools and switch to the Network tab.
- 2Load a video in the studio and generate captions.
- 3Look through the requests. You should see model files coming from huggingface.co and small requests to the AutoCap API.
- 4Check the size of the API requests. None of them carries your video file; they are small text requests, and a request that contained a video would be many megabytes.
Limits worth knowing
- Accuracy depends on the audio. Clear speech works well; heavy accents, crosstalk, and background music make errors more likely, so read the transcript once before exporting.
- Browsers differ. WebGPU availability varies by browser and device, so speed varies too.
- Memory is finite. Long, high-resolution videos can strain a browser tab, and very large exports may be saved as WebM instead of MP4.
- Local processing means closing the tab stops the work. Nothing keeps running on a server for you.