Main reason is that audio is just tremendously inefficient compared to other signals. It's large and expensive to store and process and doesn't really give you that many bytes of information you can't get elsewhere for how expensive it is to handle.
That's not a very compelling experiment, compared to e.g. logging the actual traffic, core dumping the app, and decompiling the APK.
Could have a delay. Could only work when physically moving, indicating activity. Could only be activated for some user profiles based on usage patterns. Could only activate for device owner's voice, like the voice assistants.
I used to think audio would be prohibitively expensive for Facebook to eavesdrop. But they could easily sample at random, compute it on-device, then only send keyword hashes. I think it's much more technically feasible than you're giving it credit.
I agree they probably aren't. Most likely, they predict using their other spying, then people notice frequency illusion/coincidence. But I find it odd nobody's checked.
They're sending pretty much everything BUT audio.
Main reason is that audio is just tremendously inefficient compared to other signals. It's large and expensive to store and process and doesn't really give you that many bytes of information you can't get elsewhere for how expensive it is to handle.