why this exists
most reporters who do sensitive interviews develop a workflow for handling the audio: a dedicated recorder, an air-gapped drive, a transcription pass done by hand or by a trusted freelancer. AI transcription is the obvious time-saver and most reporters won't use it on the most sensitive files for the obvious reason: the audio gets uploaded to a server they don't control, kept by terms they didn't write, in a jurisdiction they didn't choose.
on-device transcription is the version that fits the workflow that already exists. the audio stays on the same machine you handle the rest of the file on. no vendor. no logs. no upload.
what this is, structurally
private mode runs the speech-recognition model inside your browser, on your laptop, using WebGPU. the audio you drop in never makes a network request. transcription happens locally in roughly real-time on a current machine. the editor where you fix labels and verify quotes also runs locally. when you close the tab, the in-browser session is gone.
there is no vendor in the chain. nothing for a court to subpoena from a transcription company. nothing in a vendor's backups, nothing in a vendor's logs, nothing in a vendor's training data — because there is no vendor. the audio remains on your device, governed by the protections you already use for it.
workflow
- record as you already do. encrypted recorder, air-gapped drive, signal-disconnected phone — whatever fits the story.
- move the audio file to your laptop. open audiohighlight in private mode.
- drop the file in. it transcribes locally. typically one minute per ten minutes of audio on a current laptop.
- verify quotes that matter against the audio — the editor links each word to its second of audio so you click rather than scrub.
- export the parts you'll publish. the rest stays where it already was.
what we don't promise
shield laws and source protection are jurisdiction-specific and fact-specific. removing a third-party vendor closes one predictable route a court might take to your audio. it does not change what's already on your laptop, what your phone's backups contain, or what your news organization's IT policy requires. those are still your decisions, made with people better suited than us to advise on them.
we also don't claim AI transcription is perfect. it isn't. for quotes that will run in copy, verify against the audio. the editor is built for this — click a word, hear the second.
what we publish about ourselves
our private mode runs locally. we know this because we wrote the code. you can know it because the code that loads in your browser is shipped from a domain whose network requests you can audit. when private mode is on, the only network requests the page makes are the initial page load and the model file download — neither contains your audio. we publish a how-to on auditing this, and the model files are loaded from a well-known content-distribution endpoint with published checksums.
pricing
$0.25 per minute. $7.50 for a 30-minute interview, $12 for an hour. pay per file. private mode and cloud mode are the same price. no subscription, no minimum. for newsrooms with batch needs, group pricing arrives after launch — see about.