Skip to content
subtl

Recognition that never leaves your Mac

There are calls you would rather not send anywhere. Subtl can run recognition locally with Whisper, in which case the audio of the call is processed on your machine and no part of it is transmitted.

What “local” actually means here

With local recognition on, audio goes from the call to a model running in Subtl on your Mac, and the text stays on your Mac. Nothing about the audio is sent to us or to any provider.

Translation and the prompter are separate stages and they do use a cloud model — a sentence of text, not the audio. If you want the whole call to stay on the device, switch translation and the prompter off and keep the transcript: that combination sends nothing at all.

Turning it on

Settings → AI → Speech recognition offers three choices: Auto, Cloud, On this device. Auto picks for you and says on the start screen which it chose and why.

The first time you choose the local option, Subtl downloads a Whisper model — around 1 GB — inside the app. It is a one-time download; after that the option works with no network at all.

Which Macs are fast enough

Apple silicon with 16 GB of memory or more runs it comfortably. On 8 GB machines and on Intel Macs it works but lags behind a fast speaker, which matters more for live translation than for a transcript you read afterwards.

Accuracy on clear speech is close to the cloud model. On crosstalk, strong accents and technical vocabulary the cloud model is still better, and telling you otherwise would not survive your first call.

The middle option: your own API key

If your objection is commercial rather than technical — you want the cloud quality but on your own account — the Pro plan with your own key is $14 a month instead of $29. Your key is stored in the system keychain on your machine, and our credits are not touched on that plan.

Read next