If you are building a system that mimics human speech you need to teach it to be imperfect and use common parlance. Otherwise you will fall into the uncanny valley. If you listen to the conversation again there are several points where they lose immersion. For example no one would say 12 pm, they would say 'noon' instead. Google has clearly done some impressive work here, and I'm now a bit more confident/scared that they will be able to successfully fool me in the next few years.