WebRTC Weekly Issue #652
Full duplex by default, gestures somebody asked for, and an SBC in disguise
Two speeds this week. OpenAI made full duplex the baseline for Voice AI and took WebRTC call setup down to a single round trip, which is the kind of news that makes everything else look incremental. Then there is a press release that spends 600 words describing a session border controller while doing its best not to say so.
Siperb’s Go-Based Media Bridge Lets WebRTC Softphones Work With an On-Premises PBX (Siperb)
Read it twice and the news gets smaller each time. Siperb built a Go media bridge that runs inside its own desktop client: it terminates the WebRTC session locally. That is a WebRTC-aware SBC or a variation of it, and the release says as much one paragraph after explaining that the whole point was to avoid shipping one. The SBC is still in the call path, it just ships inside the softphone now: no appliance in the building, no cloud hop, and one more piece of media plumbing to get right on every customer’s machine. Obviously, they didn’t forget to mention it was written in two days with an AI coding assistant, against a spec modeled on RTP proxy design. But what is it they are really trying to announce here?
📰 This Week in WebRTC
From Bounding Boxes to Motion (Loic Denoue) - tracking mouse cursor in a browser screen sharing session
Shipping Video Calling on Facebook: the agony and ecstasy of building the first version of Facebook video calling (Philip Su) - How time flies when you’re having fun. The days before WebRTC seem so... old
How to Build an AI Voice Agent with Python and WebRTC (Stream) - a cookbook for STT-LLM-TTS loop, turn detection, barge-in. All the things OpenAI is now changing
Control your video calls with gestures using free AI from Google (OpenVidu) - gestures are so 2024. I wonder who asked for it from OpenVidu
📣 Sponsors
Meet Sipfront live on 26th August
Discover the all-in-one voice and WebRTC testing platform trusted by NFON, Parloa, and sipgate. Join Sipfront for a live demo on 26 August.
vCon: The PDF for Your Conversations
An open standard for transcripts, metadata, and consent. Learn vCon.
✍️ This Week on BlogGeek.me
OpenAI just changed how Voice AI is built - OpenAI deleted the turn detector rather than improving it. GPT-Live-1 splits voice from thinking into two tracks with a swappable frontier model, and takes WebRTC call setup from six round trips down to one. A brave new architecture