The WebRTC.ventures Blog.

On September 22–23, I was in Chicago for the IIT Real-Time Communications Conference at Illinois Tech, an annual IEEE conference where industry engineers and academic researchers present work on real-time communications. This year's tracks covered programmable networks, WebRTC applications, emergency communications, and AI in real-time communications. I had the privilege of presenting a talk in the WebRTC track titled "eBPF Makes Python a Good Language for a TURN Server." Across the other sessions and tracks,

A production voice agent has a latency budget of roughly 800 milliseconds before a conversation starts to feel slow. That budget is spent across eight sequential stages, typically split across three to five different vendors depending on how the stack is assembled. This post maps that voice AI pipeline stage by stage: mic capture, WebRTC transport in, turn detection, speech-to-text, LLM inference, text-to-speech, WebRTC transport back, and playout. Each adds its own piece to the

LiveKit can run on Kubernetes without hostNetwork, which lets self-hosted LiveKit scale like any other workload and run in clusters that don't allow host networking. Its pods stay private behind an ordinary ClusterIP Service, and STUNner, a Kubernetes-native TURN server, provides the single public address that WebRTC media flows through. LiveKit is open source and distributed by design, which makes it a natural fit for Kubernetes. LiveKit's Kubernetes guide, though, deploys the server with host

Once again, WebRTC Live broadcasted direct from the RTC.ON conference in Kraków, Poland where WebRTC.ventures CTO Alberto Gonzalez interviewed two speakers: Luke Curley,  MoQ co-creator at moq.dev, ex Discord, ex Twitch. Luke’s talk, “On a Boat: Why Pull-Based Media Streaming Beats Push,” argues that pull-based protocols like MoQ handle bandwidth-constrained, high-camera-count streaming better than push-based options like WebRTC, SRT, and RTMP, covering multipath QUIC, AI-driven rendition switching, and lossy live plus pristine VOD from a

The most frustrating thing about speaking with a Voice AI agent is when it interrupts you, or it doesn’t understand the normal flow of human conversation. There are multiple ways to solve this depending on your use case, and coming up with the best architecture for your needs is crucial to creating positive customer experiences. For over a decade, our team has built voice and video applications for our clients using WebRTC. This has involved

Self-hosting a voice AI pipeline gives you control that managed APIs can't offer: you choose the models, you decide where they run relative to each other, and you keep audio data within your own infrastructure. Getting there means managing GPU workloads, model placement, and service networking yourself. Kubernetes provides the underlying mechanisms for GPU scheduling, autoscaling, and service discovery. Amazon EKS packages them into a managed platform, using Karpenter for GPU provisioning, so you get

Now that everyone's back from summer break here in Madrid, it feels like the right time to recap AI Tinkerers Madrid: AI in Real-Time Communication Systems, which WebRTC.ventures sponsored and hosted this past July alongside BTS, who provided their offices and a great spread of local snacks. We aimed for 30 attendees and comfortably surpassed that number. The feedback was strong, too! People are already asking about the next edition, and I'll have news on

When WebRTC quality degrades on your SFU (jitter spikes, frozen frames, choppy audio), your monitoring tells you something is wrong. It doesn't tell you what is responsible. Is the network dropping packets? Or is your server too loaded to read them from the socket on time? The fix for one is scaling infrastructure. The fix for the other is pushing back on network conditions. eBPF gives you kernel-level observability to find out. eBPF (extended Berkeley

The annual ClueCon conference founded by the creators of FreeSWITCH has always been where the real-time communications community gathers to share what's working, what's breaking, and what's next. This year's edition, August 10-14 at The Midland Hotel in Chicago, made one thing unmistakably clear: AI is no longer a side experiment in the telecom world. It's embedded in production systems, and the conversation has shifted from ‘look what AI can do’ to ‘how do we

SIP and RTP feed a call and its media into a voice AI pipeline. What happens next runs through the control plane: what the voice agent is told, and what it decides to do about it.  In my talk at ClueCon this year, From SIP to Tokens: Deterministic Telephony Meets Real-Time Voice AI, I walked through this idea. The voice AI control plane, not the model, decides whether an agent hears something correctly and acts

Human speech is often unstructured and messy. To keep up with real conversation, Voice AI bots must parse partial sentences, filler words, and interruptions, all without the visual cues a human listener relies on. On top of that linguistic challenge, real-world latency budgets and network conditions test every Voice AI pipeline, and users don’t forgive one that stutters. In this episode of WebRTC Live, host Arin Sime talks with Dr. Varun Singh, Chief Product Technology Officer

In the space of thirteen months, Google took real-time speech translation from an English-and-Spanish demo at I/O 2025 to Gemini 3.5 Live Translate, a 70-language streaming model announced in June 2026 and now expanding into Google Meet through private preview, on top of the five-language version already generally available there. That speed says something: real-time translation is fast becoming a baseline expectation, not a differentiator. That shift won't stop at video conferencing. Contact centers, telehealth,

We built a self-hosted video calling system for a healthcare platform client, replacing the usual "meeting link" pattern with a real phone-call flow: clinician dials, patient's phone rings, patient answers. Highlights: Web-to-Android calling with one-tap answer Jitsi Android SDK startup optimization for near-instant connect Server-owned call lifecycle: ring, answer, decline, timeout, stale-call replacement Jitsi on Kubernetes, portable across cloud, hospital, and on-premise Prometheus, Grafana, and OpenTelemetry observability Load-tested and handed off for client-owned operation Read

STUNner is a Kubernetes-native TURN server, TURN being the relay protocol that gives WebRTC clients a public endpoint to send media through. It builds on the Gateway API to terminate TURN at a single load balancer endpoint and relay media to the pod behind it, so media-server pods stay on ordinary ClusterIP Services, cluster-internal only, with no public IPs of their own. Examples in this post use Elastic Kubernetes Service (Amazon EKS), but the resource

Voice AI tooling broadly falls into two categories: managed platforms that get you a production agent fast, and custom-built frameworks that give you full ownership of your voice pipeline. Within each category, the specific tools differ in ways that are worth understanding before you commit engineering time or lock yourself into constraints you'll hit later at scale. This voice AI comparison looks at three representative managed voice AI platforms: Retell AI, Bland AI, and Synthflow,

AI-powered voice agents continue to expand into telephony, healthcare, and enterprise customer service. Teams building these systems face a shared set of problems: keeping latency low, keeping calls reliable, and keeping the underlying infrastructure fast enough to support real-time audio at scale. Two members of the WebRTC.ventures team will address these problems from different angles at ClueCon 2026, taking place August 10–14 at The Midland Hotel in Chicago. Alberto Gonzalez: Deterministic Telephony Meets Real-Time Voice

Telephony agents built on Voice AI platforms like Retell AI make it easy for businesses to run live phone conversations through an AI agent instead of an IVR or human call center staff. Getting one of these agents into production means solving two things: integrating it with existing telephony infrastructure, and adding a compliance layer that preserves the organization's security and governance standards. On the integration side, the agent must connect to existing PBXs, SIP

How can AI scale clinical roleplay training beyond what live sessions with human trainers can deliver? For CETA Global, a nonprofit that trains frontline mental health providers around the world, the answer was an AI “flight simulator”: a real-time platform where psychologists practice difficult client sessions with an AI-generated patient, complete with a lip-synced streaming avatar, live coaching, and replayable feedback. Behind the scenes, a team of six AI agents works in concert to play the

An agentic workflow is the disciplined form of what's often called "vibe coding": giving a coding agent a task and the context it needs, then letting it plan, edit, and test across your codebase while you review the results. When it's done well, the productivity gains can be dramatic. In his vibe-coding WebRTC post, Philipp Hancke describes getting changes merged into libWebRTC, Firefox, and even Pion, work that used to wait on finding a free

Most AI applications operate in turns: a user submits input, the model processes it, and a response is returned. A delay of a second or two goes unnoticed. Live voice AI does not offer that margin. In a roleplay training session where a psychologist practices a difficult client conversation with an AI-simulated patient, every unnatural pause undermines the realism of the exercise. And with it, the training value. This was the central engineering challenge behind