How is that? Desktop systems respond quickly, cloud systems often respond after a second - it takes time for data to do a round-trip over crappy mobile connection. Mobile latency is a big thing.

I agree wrt. Cortana (and Siri, Google Now) - most of the things you use them for will require Internet access. But there are still a lot of things that you could do with voice that shouldn't require a network connection, and we're missing the ability now. Not to mention you have zero customization options for cloud-based recognition. I could make good ol' MS Speech API recognize pretty much anything I wanted it to. No problem making it recognize a limited subset of two languages at the same time. With cloud-based systems, if the voice recognition doesn't like my accent, I'm out of luck.