RTC.ON NEWSLETTER – May 2026

The RTC.ON lineup is starting to take shape – and there's no time to waste if you want in on the action.

Four things you need to know today:

1. First Speakers announced! 🥁

We're kicking off the RTC.ON 2026 lineup with two names we're especially excited about:

- Luke Curley – co-creator of Media over QUIC, also hosting this year's workshop. Talk title and description coming soon!

- Daniil Popov – Head of Technology at Cyanview (the company behind Super Bowl streaming) will share another super exciting streaming challenge, this time from a desert music festival – I'm sure you know which one 👀

More speakers announced in the coming weeks!

2. CFP closes May 24 🚀

If you've been sitting on that talk idea – this is your final nudge. Share a hard problem you solved, an optimization that made a difference, or something you shipped and are proud of. Voice agents, AI encoding, streaming infrastructure, and beyond – but if yours isn't on the list, submit it anyway. Proposals close May 24.

3. The workshop outline is here!

Luke Curley's MoQ workshop now has a full outline – exactly what you'll build, what you'll walk away knowing, and what to bring. If you've been curious about Media over QUIC and want a hands-on day with one of the people building it, this is the page to bookmark.

4. Blind Bird tickets end May 28 💸

Tickets are still at the lowest price they'll ever be – but not for long. Blind Bird pricing ends May 28, and spaces are limited. Grab yours before prices go up.
Use code news-15 for an additional 15% off – valid until May 29.

OPENAI BLOG | Link

How OpenAI delivers low-latency voice AI at scale

OpenAI walks through how they rebuilt their WebRTC stack to fit Kubernetes. The classic one-port-per-session model doesn't play nice with cloud load balancers, so they split media routing from protocol termination - a thin stateless UDP relay forwards packets based on the ICE ufrag, while a separate transceiver service owns all WebRTC session state (ICE, DTLS, SRTP). Combined with geo-steered signaling through Cloudflare, this gives them a small public UDP footprint and ingress close to users. Worth a read for anyone who has ever struggled with WebRTC and K8s, which, let's be honest, is everyone that tried.

MOQ DEV | Link

OpenAI’s WebRTC Problem

Predictably, Luke Curley couldn't let the OpenAI post pass without comment. Drawing on his experience writing WebRTC SFUs at Twitch and Discord, he argues the whole architecture is solving the wrong problem - WebRTC aggressively drops audio packets (which is awful for prompts), the 8+ RTT setup adds noticeable latency, and routing on the ICE ufrag breaks the IP/port migration that was supposed to be WebRTC's main selling point. His pitch: ditch WebRTC entirely for WebSockets, or better yet QUIC with QUIC-LB and anycast. As always with Luke's posts, it's spicy but technically sharp.

CLOUDFLARE BLOG | Link

When "idle" isn't idle: how a Linux kernel optimization became a QUIC bug

A flaky integration test in cloudflare/quiche turned out to expose a real CUBIC bug: after a heavy loss event drove cwnd to its two-packet minimum, it would stay pinned there forever, oscillating between recovery and congestion avoidance every 14ms. Root cause was a kernel idle-period optimization ported to userspace incorrectly - it measured the "idle" delta from the last sent packet instead of from when bytesinflight actually hit zero, inflating the delta by a full RTT and pushing the recovery boundary into the future on every cycle. The fix is essentially one line. A great read on how subtle CCA bugs hide in corners no one tests.

PICOVOICE BLOG | Link

Noise Suppression Guide 2026: Algorithms, Metrics, and Implementation

A solid reference write-up on noise suppression - what it actually is, how it differs from noise cancellation and echo cancellation, and how the typical STFT-based pipeline works. There's good coverage of objective metrics (PESQ, STOI, SNR improvement) versus subjective ones (MOS), plus a comparison of traditional DSP approaches (spectral subtraction, Wiener filtering) against deep learning methods (RNNs, CNNs, transformers). Closes with a tour of available solutions - RNNoise, WebRTC NS, Krisp, Dolby.io, and Picovoice's own Koala. Vendor content, but the kind of comprehensive overview worth bookmarking for newcomers to the field.

Big news from the Fishjam side of the house: real-time video composition is coming to Fishjam, powered by Smelter.

Combine multiple video streams, overlays, and effects in real-time, control your scenes with code, and skip the infrastructure headache – Fishjam runs the servers, you focus on building. Think live sports overlays, multi-camera productions, dynamic studio scenes, all composed on the fly.

The early access list is open now – sign up to be first in line when it goes live (and play with the Smelter demo while you wait 👀).