Given your framing as "massive scale audio processing" to pick signals out of audio streams, I think you may have misread (or not read...) the paper. It's not

Given your framing as "massive scale audio processing" to pick signals out of audio streams, I think you may have misread (or not read...) the paper. It's not claiming Alexa is listening in while inactive. It's claiming that it's inferring user characteristics (age, health, etc) from the audio of voice commands, instead of just the post-transcription text.