what "private" actually means here
private-transcription mode runs the speech-recognition model inside your browser, on your device, using WebGPU. your audio never makes a network request. there is no server we could be compelled to hand over, no cache, no analytics ping, no third-party processor in the chain. when you close the tab the transcript is yours and ours is gone.
this is different from "encrypted in transit" or "we don't train on your data" — those promises rely on a vendor's word. on-device transcription doesn't need a promise: there's no opportunity for the audio to be misused because it never leaves the device that recorded it.
who this is for
two audiences for whom upload-based transcription is a working compromise rather than a working solution: journalists protecting sources, and writers protecting unpublished work.
journalists and reporters
source protection isn't theoretical. subpoenas reach third-party processors more easily than they reach reporters. if a court orders a transcription vendor to hand over a file, the vendor will hand it over. if there's no vendor in the chain, there's nothing to hand over. for journalists
podcasters
unpublished material. embargoed announcements. off-the-record passages a guest asked you to trim. an interview with a source who agreed to talk on the condition that the audio doesn't pass through additional third parties. for any of those, on-device transcription is the workflow that matches the promise. for podcasters
what gets traded off
on-device transcription has limits and we don't pretend otherwise.
- your machine matters. a 30-minute file transcribes in roughly real-time on a modern laptop with WebGPU support, and slower than that on anything older. we publish the hardware floor on the browser tool page.
- english first. on-device models are smaller than cloud models. they handle english near-fluently and degrade on most other languages. for non-english audio that doesn't need privacy, the cloud-mode tool is more accurate.
- you can't share a link. the transcript lives on your device. you can export it (.docx, .srt, .vtt, plain text, json) and share that. the click-word-to-replay-audio editor works locally; sharing it with a colleague means sharing the audio file itself.
how this fits the rest of the product
private mode is a setting, not a separate product. you can run any file in private mode, you can run any file in cloud mode, the editor and exports are identical, and switching between them is one click. the default for an account is chosen at signup based on use-case.