These are deals done straight with the phone makers. Full audio is never shared with advertised, just the filtered data, in text, ex: "I'm moving to Canada next week" or "My girlfriend just got pregnant"
Full transcription is really difficult to do at the edge on battery power. It's either bandwith heavy or CPU heavy. Both of which nail the battery. The only really practical way I can see it being done is with keyword tracking, and then recording snippets either side.
But that still takes a reasonable amount of power to do.