Transcribe a recording, on your device SCRIBE

The model comes to the recording, not the other way round

Every transcription service works by uploading your audio. That is not a criticism of any of them; it is how the architecture works. The model is large and lives on their machines, so the recording has to travel to it, and once it has travelled it exists on somebody else's disk under somebody else's retention policy, whatever that policy says.

This runs the model here instead. A Whisper model is downloaded once, executes in this tab, and the recording is read from your disk into memory and never sent anywhere. What crosses the network is program code and model weights, moving one way, towards you.

The weights come through obscuraos.com rather than straight from the model host, which is a small thing with a specific consequence: no third party learns that this browser is about to transcribe something. Once the model is cached on the device it works with no network at all, which is also the easiest way to satisfy yourself that the claim on this page is true. Turn the network off and transcribe something.

What the recordings people transcribe actually are

This is why it matters more here than for most tools. The recordings that need transcribing are therapy sessions, medical consultations, disciplinary hearings, research interviews with vulnerable people, solicitor and client conversations, calls with a source. Uploading one of those can be a breach of confidence, a breach of contract, or a notifiable data incident, and in several professions it is a thing you would have to declare.

A journalist protecting a source, a clinician bound by confidentiality, a lawyer holding privileged material, and a researcher under an ethics approval all have the same problem and it is not a preference. On-device transcription is the only version of this that does not require anybody to trust a third party's word about deletion.

What to expect from it

The model is a small one, chosen so it runs in a browser tab on an ordinary machine rather than on a graphics card. It handles clear speech well and degrades where any speech recognition degrades: heavy background noise, several people talking over one another, strong accents against a model trained mostly on other accents, specialist vocabulary, and poor microphones. Expect to correct it. A first pass that is ninety per cent right and took no upload is worth more than a perfect one that cost you the confidentiality.

The first run downloads the model, which takes a moment and happens once. Around thirty minutes of audio at a time is the comfortable working limit; longer recordings are better cut into pieces, which SNIP does. Afterwards, HUSH bleeps named words out of the audio itself, SUBS turns a transcript into subtitles and retimes them, and MEDISCRUB takes the identifying details out of the text before it goes into a record.