First author of the linked paper here. It's an interesting question. Actually the voice activation (on/off periods of sending vs silence) was the first thing

First author of the linked paper here. It's an interesting question. Actually the voice activation (on/off periods of sending vs silence) was the first thing we looked at in that project. There's definitely some information leakage there, but it was really hard to learn anything meaningful from it. The problem seems to be that long strings of words all get lumped into a single activation. It's really hard to discern anything about what those words are, or even what language they come from. We got some very weak results on language detection from VAD. You can learn something about the language, but it's not very precise. For example, maybe you could tell that a given conversation is definitely not language A, B, or C, but it might still be language X, Y, or Z.