# LiveKit says its voice interruption rate is just 4.5% _LiveKit's Jesse Hall told AI Engineer his turn detector cuts interruptions to 4.5% versus 9.9%, within a 500-600 ms human-like window._ **Published:** 2026-10-26 **Source:** https://www.startuphub.ai/ai-news/artificial-intelligence/2026/livekit-says-its-voice-interruption-rate-is-just-4-5 --- Jesse Hall told [AI Engineer](https://www.youtube.com/watch?v=OmyqelTwpOo) that LiveKit cuts off callers 4.5% of the time versus 9.9% for the leading alternative he benchmarked, and he tied that difference to a tight latency budget where 1.5 seconds is tolerable but 500 to 600 milliseconds is where a finance call starts to feel human. That budget is the core claim. Hall argued voice agents fail not because models are slow on leaderboards, but because the conversation is measured wrong. He said text-to-speech vendors report time to first byte around 210 milliseconds, according to Hall, yet his tests found some add up to three quarters of a second of silence before audible speech, so time to first audio is the real metric. For language models the same distortion applies, he said, because time to first token does not equal time to first sentence, and the text-to-speech engine cannot speak until a full sentence is ready. A second number anchors the finance risk he demoed, a caller saying "I want to pay on the 3rd" heard as the 30th, which then flows downhill through the LLM unchecked. Hall, who leads work on LiveKit Agents, the open-source real-time framework that [OpenAI](https://www.startuphub.ai/ai-news/investors-news/2026/softbank-s-65b-openai-bet-sparks-investor-concern) chose for [ChatGPT](https://www.startuphub.ai/ai-news/artificial-intelligence/2026/openai-s-codex-in-chatgpt-ai-for-engineering-teams) advanced voice mode according to Hall, demoed a fix for the other failure mode, agents that go silent during tool calls. He showed a hotel booking agent that keeps asking questions while writes to a backend run async, with one await handing the mic back so conversation and tool execution no longer share a thread. LiveKit’s turn detector, he said, listens to audio rhythm and intonation, not just transcript pauses, which is why "$500 and $55.23" does not get truncated at $500. The SAP Architecture Center notes a related infrastructure detail not in the talk, that the voice agent runs inside LiveKit Cloud, a WebRTC media layer that handles transport, jitter and packet loss before audio reaches the speech pipeline. The evaluation method is where the skepticism lands. Hall said LiveKit does not pick models off leaderboards. It runs simulated callers, goals, mind changes, interruptions, mumbling and background noise, against the same hotel agent with a real backend, scoring whether the correct reservation hit the database and whether steps were unnecessary or hallucinated, then swaps the stack byte for byte. On that hotel benchmark, Gemma 4 31B on LiveKit inference ranked highest, according to Hall, and the same model on another provider was slower. The data Hall showed does not include a finance benchmark with noisy speakerphone audio, authenticated payment amounts, or dispute flows, and it does not show error recovery rates or handoff-to-human rates when tools fail. His checklist is precise, measure time to first audio and time to first sentence, confirm critical dates and amounts, verify identity before account access, keep listening while slow tools run, and turn failures into regression scenarios. What would change the read is an independent run of that personalized benchmark on a regulated finance workload with real call audio and a disclosure of false confirmation rates. Until then the open question Hall leaves is whether 4.5% interruptions and a 600 millisecond gap is good enough when a single misheard payment date is not a dropped turn but a wrong charge. --- Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory. © StartupHub.ai. All rights reserved. You may not republish this article in full without a license. Search engines and AI research tools may crawl and summarize for reference. Bulk reproduction or model training on this content requires a license. See https://www.startuphub.ai/terms.