WebRTC vs SIP for Voice AI: what I've learned from production deployments

I manage enterprise Voice AI deployments and recently wrote
a detailed breakdown of this decision:
https://www.voiceaipm.com/2026/04/webrtc-vs-sip-which-protocol-for-your.html

The short version of what I've found in production:

• If users call from a real phone number (mobile/landline): you need SIP. No way around it - the PSTN speaks SIP.

• If the voice interface lives in a browser (click-to-call, web widget): WebRTC. Opus codec, no carrier cost, NAT traversal handled automatically.

• Enterprise deployments almost always end up needing both, bridged via an SBC. This is where most teams underestimate the complexity.

The mistake I made on one project: chose WebRTC because the
demo worked great in a browser, then discovered the client's
contact centre platform only accepted SIP. Spent 6 weeks
building a gateway that wasn't in the scope or budget.

A few things I'm genuinely curious about from people
building in this space:

1. Are you using a managed Voice AI platform (Vapi,
   Retell AI, Bland AI) or building the pipeline yourself?
   How did that change the protocol decision?

2. For WebRTC deployments - what TURN provider are you
   using? Twilio NTS, Coturn, or something else?
   What's your experience with relay latency?

3. Has anyone built a SIP-WebRTC gateway from scratch
   rather than using a managed SBC? What did you use -
   Asterisk, FreeSWITCH, or something more modern?

4. How are you handling VAD (Voice Activity Detection)
   differently across WebRTC vs SIP? In my experience
   the acoustic environment assumptions are quite different.

Happy to share more specifics from the deployments I've
managed if useful.