The FnKey README says audio streams to Deepgram while the user speaks, or the whole clip goes to Groq after release, so the page places recognition with those services, not on the Mac.

Checked against its source on October 26, 2026

In the source's own words

“Audio streams over a WebSocket while you speak. Fastest. ... The whole clip is sent after release.”

evoleinik/fnkey: FnKey README (GitHub), evoleinik, GitHub (maker's repository), 2026, How it works table in the README.

What it measures: FnKey needs a Deepgram or Groq API key. Deepgram is the default streaming backend and Groq runs only when Deepgram fails or does not answer within 10s.
What it does not show: The table describes where the audio goes. It does not say how long either service keeps it.

A row of Dictation App Features Compared: Offline, Privacy, AI, as FnKey: speech on device: 0 1 = yes, 0 = no.