[m463]

I would't put it past the cloud providers to ALSO snarf up the transcript (and possibly the raw audio to help with training).

It might even be a cost savings.

[jrpt]

If they're federating the ML, they can also do federated learning. (https://en.wikipedia.org/wiki/Federated_learning) Although in most cases I suspect it's probably not federated learning, instead just a deployed model that they've trained with in house data.