In my application I would like an arbitrary browser peer (client) to connect to a public host (server) through WebRTC. I am trying to understand if I need any STUN/TURN, or if I can do without since the server is public.

My understanding is that, without STUN, peers on both sides will initially only generate host ICE candidates. While the server's candidates will be viable (as its IP is public and ports accessible), the client's might not if it is behind NAT. As far as I understand, at this point the client can generate prflx candidates, that could be used by the server as viable endpoints (no matter the NAT) and so a connection would be established. I am not sure though what the conditions are for generating prflx candidates, and if the availability of a host candidate for the server is sufficient.

Is this more or less accurate? Could this get the peers connected reliably, or does it depend on the specific ICE implementations on both ends/network conditions?