I have always been dubious of this because:

Someone would have noticed if all the phones on their network started streaming audio whenever a conversation happened.

It would be really expensive to send, transcribe and then analyze every single human on earth. Even if you were able to do it for insanely cheap ($0.02/hr) every device is gonna be sending hours of talking per day. Then you have to somehow identify "who" is talking because TV and strangers and everything else is getting sent, so you would need specific transcribers trained for each human that can identify not just that the word "coca-cola" was said, but that it was said by a specific person.

So yeah if you managed to train specific transcribers that can identify their unique users output and then you were willing to spend the ~0.10 per person to transcribe all the audio they produce for the day you could potentially listen to and then run some kind of processing over what they say. I suppose it is possible but I don't think it would be worth it.