[refulgentis]

My jump to conclusions mat says "The author doesn't understand tech!"

I worked on on-device speech at Google, both frontend and backend. Unless there was a secret cabal mirror team that did all the evil stuff and made us think we were writing the code that was shipping, it doesn't come close to describing how things actually were worked through internally.

Most trivially, and narrowly, Most Evil Amazon would be just as happy with a transcript of what you said that was recognized locally. There's 0 reason they need the audio. I'd also hazard a guess they're not streaming audio to the server, no one would create a hotword detector that would rely on that, due to the latency and bandwidth demand alone.