Every word is a wake word, but most of the time it just wakes up to record the word in a database with a location and timestamp and then goes back to sleep.

Then it sends that information off in the form of a “software update check” or a “user feedback telemetry ping”.

There’s no need to stream the audio. We’ve had computers capable of turning audio into text since the 90s. Classic MacOS 8 could do it.