Unless you assume that Facebook et al are running ASR models on-device all the time and somehow making it invisible to the end-user.
Same problem if they were surreptitiously streaming audio to their servers. You would see it from the outgoing packets and streaming that amount of data would also be fairly expensive.
Some voice codecs get their rate down to 2-3kbps. Maybe it could store and forward with content loads without raising flags?
Just spit balling here.
People are incredibly predictable using only a handful of demographics, there's simply no need to invest the astronomical amount that would be necessary to process these conversations when there's already many simple ways to track/generate user interest.