> Device-local speech recognition is fundamental to AR — we can’t tap our feet & twiddle our thumbs while we wait for a round-trip to the cloud to conclude how

> Device-local speech recognition is fundamental to AR — we can’t tap our feet & twiddle our thumbs while we wait for a round-trip to the cloud to conclude how to handle “Group these five items, and remember them for next week when I’m at the office” I'm not convinced that this is true. Round-trip latency has gotten a lot lower, especially if you take edge computing into account. 20-30ms round-trip is not unusual. If your mic feed is being streamed in real-time, and the program is able to achieve a high-confidence prediction at (or before) the moment you finish speaking, I think it should be possible to deliver an experience that feels instantaneous, at least for users with cutting-edge connectivity. For visual interactions, 100ms feels instantaneous; I bet tolerance for spoken interactions is even higher.