There are smarter ways to do this without constantly sending all the audio. Wakewords + surrounding snippets for example. In fact if the audio recognition is do

There are smarter ways to do this without constantly sending all the audio. Wakewords + surrounding snippets for example. In fact if the audio recognition is done on-device then only the text data could be sent which would be indistinguishable from telemetry.