Here’s what I have ascertained:
The voice commands work by processing your input locally, at first, and then by sending it to the cloud to a more capable engine. You can see how it first transcribes what you said, often times not quite correct, and then it fixes it after half a second or so. If the cloud connection fails, instead of just rolling with what it has, it does nothing (a rather puzzling implementation).
This should be an easy fix by their software team, however they may be tied by agreements with their voice recognition provider who, I bet, would very much like to collect all your voice input data for their own legit and totally-not-nefarious purposes.