Update: OneDrop now transcribes on free servers by default; the video file still never leaves your device; only compressed audio is sent to the provider and deleted right after. On-device processing remains available as an opt-in on capable devices.
Every cloud captioning service has the same first step: you upload your video. That step is easy to ignore for a vlog and much harder to ignore for a client's unreleased product, a patient interview, an internal training video or footage from a story that has not run yet. On-device captioning removes that step entirely: the video never leaves your machine. Here is what that means technically, what it protects (and what it does not), and how to verify the claim yourself.
What "on-device" actually means
When you drop a video into the OneDrop studio, the browser reads the file straight from your disk into the tab. From that point:
- Decoding happens locally. The video and audio are unpacked by your browser using your device's own media capabilities.
- Speech detection runs locally. A small voice-activity model decides which parts of the audio contain speech.
- Transcription runs locally. Whisper, a speech recognition model, is downloaded once into your browser's cache and then executed on your processor. Your audio is fed to that local model, not to a server.
- Rendering happens locally. Burning captions into the MP4 is your browser drawing frames, not a render farm.
At no point is a copy of the video or the transcript transmitted anywhere: the video file is never transmitted; the compressed audio is, to the transcription provider, and it is deleted right after processing. There is no upload endpoint for the video itself.
What this protects you from
- Server-side copies and retention. Cloud services process your video on machines you do not control. Their retention policies, backups and subprocessors become your problem. With the video staying on your device, there is no copy of the footage on anyone else's disk to retain, subpoena or breach; the compressed audio sent to the provider is deleted right after processing.
- The accidentally-public upload. Anyone who has worked with video long enough has a story about a link shared one permission level too wide. A workflow where the video file never leaves your device cannot spring that leak.
- Content that legally cannot be uploaded. Client NDAs, medical footage, unreleased products, material involving minors: categories where "we delete it after processing" is not an acceptable answer because the answer to "did the video leave my device?" has to be no.
- Metered and slow connections. A side benefit: a one-gigabyte video that never leaves your device costs zero upload bandwidth and does not wait on transfer time.
What it does not protect you from
Honesty requires the other side of the ledger. On-device processing does not make your device secure. If your laptop is compromised, nothing about local processing matters: a compromised device can read everything you do, in any tool. It also does not protect a video after export: once you post the MP4 publicly, it is public. And the opt-in on-device mode still needs an internet connection on first run to download the transcription model into your browser cache; that download is code coming to your device, never a copy of your data going out.
Where your project lives instead
The studio keeps your projects, edits and style choices in your browser's local storage so you can close the tab and come back later. That storage sits on your machine, inside your browser profile, tied to the site.
The privacy consequence: clearing your browser data, or using "Clear saved data" in the footer, deletes your saved projects; there is no cloud copy to fall back on. That is the trade in one sentence: nothing on a server means nothing recoverable from a server either.
How to verify the "no upload" claim
Trust, but verify: this claim is checkable in about a minute:
- Open your browser's network tools (F12 → Network) before you caption.
- Drop in a video and run a transcription.
- Watch the requests. In the default mode you will see the compressed audio go to the transcription provider (deleted right after processing) and you will not see your video: no large POST request ever carries the video file itself.
You can go further: switch on the opt-in on-device mode, disconnect your Wi-Fi entirely after the model is cached, and caption something. It works. A cloud tool cannot pass that test by definition.
When to choose local processing, in plain terms
Choose on-device captioning when the video contains anything you would hesitate to email to a stranger: client work under NDA, unreleased material, sensitive interviews, anything with legal or contractual exposure. Choose it for the practical reasons too: no upload wait, no account, no file size gatekeeping beyond your own device's limits.
The trade-offs (local storage instead of cloud projects, your processor doing the work instead of a server fleet) are the cost of that posture. The FAQ covers what "Clear saved data" removes in detail, pricing explains what the paid plans will and will not change about this, and if your work is interview-heavy, the guide to captions for interviews covers the workflow end to end.