LiveKit

We learn to speak before we learn to read. Voice is the most natural interface we have. We just raised a $100M to make building voice AI as easy as a web app.

Full guide: livekit.com/blog/async-too…

Async tools are a new way to handle long-running work in a LiveKit voice agent. Instead of going silent while a slow tool runs, the agent keeps talking: it acknowledges the request, streams progress as each phase finishes, lets the caller cancel mid-task, and won't fire the same action twice. How are you handling tool calls in your voice agents today?

Learn more & test it out here: livekit.com/products/infer…

Frontier models keep getting smarter…and slower. This is bad for voice agents, where every millisecond before the first word counts. So we optimized for speed. Introducing Gemma 4 31B on LiveKit Inference: 🗣️ 381ms to first sentence — 2x+ faster than the next model ✅ 88% task completion (see the criteria in reply) 💰 $1.20 / 1M output tokens Fast enough for real conversation, smart enough for effective agents.

The fight over whether AI creates jobs or destroys jobs misses the point. AI does not take your job from the top. It takes it from the bottom. That is the read from @dsa , founder and CEO of @livekit , the infrastructure powering voice AI for Tesla and xAI. He has raised over $180M from Index, Altimeter, and Redpoint. His point: AI eats the entry level first. The undifferentiated, repeatable work. The exact rung most people use to climb into a career. He is watching it at home. His wife's younger siblings just graduated into one of the hardest entry level markets in years. So what actually protects you? Not your resume. When Russ hires, he ignores the Y intercept and bets on the slope. One of his best engineers joined LiveKit at 17, knew almost nothing, and had relentless agency. What AI does not commoditize is judgment, taste, reputation, and relationships. LLMs have no emotion. They can do the work. They cannot be the person other people trust. So the move is not to outrun AI. It is to climb above the rung it is automating before it gets there. Get your foot in the door. Build the judgment and the relationships now. Stay more useful than the version of you that just got automated. We’d like to thank @AlphaSenseInc for sponsoring this episode! Full episode below 👇 youtube.com/watch?v=HyVD4R…

Join us at the official AI Engineer World's Fair Hackathon in San Francisco, June 27 to 28. 48 hours to bring an idea, build a voice agent on LiveKit, and ship it. Our team will be there to help. Come build something cool. Register here: cerebralvalley.ai/e/aiewf-hackat…

The second official Engineer World’s Fair Hackathon with @aiDotEngineer Join us for 48 hours of building at the frontier of recursive self-improvement (RSI): systems that learn from their own outputs and become more capable through continuous iteration. Excited to be partnering with @GoogleDeepMind, @MongoDB, @MiniMax_AI, @LiveKit, and @digitalocean. 🏆 $10,000+ in prizes and partner credits June 27–28 hosted at @SHACK15sf Apply below👇

Build a voice AI agent with memory: livekit.com/blog/supabase-…

Most voice AI agents forget you the second the call ends. We built a starter kit that fixes it. LiveKit for the voice, @supabase for memory, RAG, and per-user auth, all in one Postgres database. TypeScript and Python included. Give it a try and tell us what you're building.

GitHub + harness: github.com/livekit/eot-be… Leaderboard: livekit.com/benchmarks/eot… If you build EoT models, run your numbers and submit them.

Voice AI has a benchmarking problem. Everyone claims their end-of-turn model is the best, but you couldn't actually compare them. Datasets are proprietary, methods are opaque, and there is no shared ground truth. That changes today. We hit this while developing Turn Detector v1, so we open-sourced eot-bench. 5,000+ real user conversation turns across 14 languages, an evaluation harness that measures the real production tradeoff between latency and false cutoffs, and a live public leaderboard. This should become the default way we evaluate turn detection models.

We evaluated performance using eot-bench, an open benchmark suite and datasets for end-of-turn detection: github.com/livekit/eot-be…

LiveKit Turn Detector v1-mini is also available today. It has the same architecture but a smaller footprint, optimized for fast CPU inference and bundled with our Agents SDKs (Python 1.6.1, TypeScript 1.4.7).

We shipped LiveKit Turn Detector v1. Instead of reading transcripts, it listens to speech directly, combining semantic and acoustic cues into one end-of-turn prediction. The result: high accuracy, low latency—the best model we tested across 14 languages. Available on LiveKit Cloud.

We released Sonic-3.5 and Ink-2, the #1 streaming models for text to speech and speech to text you can use in your voice agents today. New architectures enable new frontiers for speed and quality. We're now the only provider to have #1 models for both speaking and listening.

Read more: livekit.com/blog/teleopera…

Operating a robot over the internet means camera frames and joint state arrive at different times, so your observations drift and training data gets misaligned. LiveKit Portal fuses them back together with the same code, whether the robot's in the next room or another continent.

We built a live multilingual, multi-person video call with Gemini 3.5 Live Translate on LiveKit. Everyone picks their language, speaks naturally, and hears each other in real time in their language of choice. Watch the demo and check out the open source repo: github.com/livekit-exampl…

NVIDIA's Nemotron 3.5 ASR transcribes 40 language-locales from a single 600-parameter streaming model with ~100ms latency. It’s small enough to run on your laptop and drops straight into LiveKit Agents. We built a multilingual teleprompter to show it off. See the full breakdown: livekit.com/blog/nemotron-… @NVIDIAAIDev #VoiceAI #NemotronSpeech