RTC.ON NEWSLETTER – July 2026

Early Bird tickets for RTC.ON 2026 are on sale until Friday, 31 July. After that the price goes up and stays up, so if Kraków in September is already on your calendar, this is the cheapest the ticket gets.

On stage, a few of the names already confirmed: Jean-Baptiste Kempf (VLC, VideoLAN) on using QUIC for machine control, Will Law from Akamai on a new streaming format, Gwendal Simon from Synamedia on track switching in MoQ, Daniil Popov from Cyanview on getting phone cameras into a live broadcast chain, and Dan Jenkins from Nimble Ape on running a gallery with an LLM instead of a director. More names are landing over the coming weeks.

WEBRTCHACKS | Link

Vector Vibing to speed up Opus encode by 20%

In 2018 someone left a note in Chromium's libopus build config: if libopus ever gets AVX, wire it up. libopus brought the AVX2 kernels along with the 1.5.2 roll in April 2024, and there they sat for two years, present in the tree but compiled out, guarded by a comment that was no longer true. Fippo finally flipped the switch. Thirty lines of BUILD.gn, nothing changed upstream, and speech encode drops by 17 to 23% at the complexity WebRTC uses by default. Music barely moves, since two of the three kernels live in SILK. If you encode Opus on the server side, go and check what your own build actually enables.

CHROME FOR DEVELOPERS | Link

Chrome 151 Beta

Two things stand out for us. The first is an origin trial for SCTP Negotiation Acceleration, which carries the SCTP init parameters inside the SDP offer/answer and saves up to two round trips when opening a data channel. The second is the <usermedia> element, a declarative way of starting a capture that hangs the permission prompt off a real user gesture instead of a getUserMedia call. There is also unspokenPunctuation for SpeechRecognition, which infers punctuation from pauses and prosody instead of making people say "comma" out loud.

ELLO | Link

Teaching a child in <1000 ms: the architecture behind a real-time tutor

A writeup on building a voice agent against a hard sub-second budget, here an AI tutor for kids aged 4 to 9, where two seconds of silence is enough to lose the child completely. The usual tool loop burns 3 to 4 seconds of dead air per turn, so they dropped it: the model streams several actions in one generation, and an interpreter executes them while it keeps writing. A second agent replans asynchronously while the child is thinking or talking, and the safety classifier gates execution rather than generation, so the check costs nothing on the happy path. Plenty of this carries over to any real-time voice stack.

HYDROGENAUDIO | Link

FFmpeg 9.1's new AAC encoder

FFmpeg's native AAC encoder has been rewritten and now scores above both fdk-aac and qAAC on metrics and in listening tests, which is quite a turnaround for something people spent a decade telling each other to avoid. It matters most for live, where RTMP with H.264 and AAC is still the ingest path everywhere, and OBS will not even offer you anything else in streaming mode. Two caveats: it is tuned for 48 kHz, and there is still no true VBR. It missed the 9.0 branch, so it lands in 9.1.

Last month we told you MoQ streaming had landed in Fishjam. We're also part of the group building the software underneath it: Fishjam is a member of the OpenMOQ Software Consortium, alongside charter members Akamai, Cisco, YouTube, Synamedia, CDN77, Red5 and Oracle.

OpenMOQ exists to turn the IETF's MOQ and QUIC specs into open-source, application-level implementations that actually interoperate: ingest, relay and playback, built once and shared, so the whole ecosystem competes on products instead of re-implementing the foundation ten times over. That's the same bet we made when we picked Luke Curley's open-source relay to run Fishjam's hosted relay, and it's the reason MoQKit, our native iOS and Android SDK, is open source too.

If you want to see where this goes next, a lot of the people doing the work will be on stage in Kraków in September.