[muzani]

Speech to text isn't difficult to do and within the the capabilities of Meta. I even wrote a basic proof of concept on what it might look like: https://github.com/smuzani/soundrecorder

It's not too hard to transcribe text into a few kb, or summarize them further on the client, and then upload that together with the rest of the data.

I'd be very interested in seeing those studies and whether they match up.