That’s naive. Really naive. The reason you train them to learn a specific phrase is to avoid false positives that would cause it to wake up when that was not the user’s intention. If the goal is to record all conversations then waking up on any phrase would require no training at all. Just enough factory programming to recognize sounds in the typical human frequency range.
You’ve got the requirements for surreptitious recording completely bass-ackwards.
Furthermore there is no need for real-time transmission of recordings back to the mothership. That’s only a requirement when the user is actively engaging with the device. An eavesdrop-mode could record to local flash memory and then batch up transmission when the user deliberately interacted with it. Just 4GB of flash could easily hold 24 hours of compressed speech, which would probably cover at least a week of eavesdropping since people won’t be talking around it non-stop. And if it runs out of local storage, no big deal, just throw out the old recordings to make room because the goal would not be perfect coverage, just to get whatever is easily gettable.
