There is no one good answer for the debate between on premise vs. managed and each of these options have sub-options to consider and decide on. The devil is in the details and in many cases a hybrid solution is required, using some functions (such as NAT traversal) from 3rd party hosted services while licensing (open source or commercial) other components of the solution.
Technically WebRTC and SyntaxNet have nothing in common. What they do have in common is that Google again makes it easy to add capabilities, once in the hands of the rich and experts, as a feature into applications. Once this is possible, Natural Language Understanding (NLU) can be added also into WebRTC applications. How can WebRTC applications benefit from it?
Even though the known implementation of WebRTC is the open source coming from Google, it is not the only implementation. This interview is with Stefan Alund from Ericsson who is heading the team behind OpenWebRTC