LiveKit has released Turn Detector v1, a voice AI model that uses combined acoustic and semantic processing to predict speaker end-of-turn events directly from audio streams. Designed to optimize conversational flow for streaming agents, the model reduces false cut-offs to 9.9% within a 300 ms latency budget and is accompanied by an open-source benchmark suite called eot-bench.
Conversational latency is the primary barrier to human-like AI interactions, where typical silence-based detection forces a choice between awkward pauses and frequent interruptions. By fusing prosody signals with semantic intent, LiveKit reduces the 'waiting tax' of transcription-dependent models. This move positions the agent framework as a critical infrastructure layer that decouples conversational logic from specific STT or LLM vendors. For the broader ecosystem, the simultaneous release of eot-bench attempts to standardize performance metrics in a market where proprietary 'black box' models often lack transparent latency data. Success here would force competitors like Deepgram and AssemblyAI to accelerate their own integrated endpointing features. Watch for whether eot-bench is adopted by rival voice framework developers like Vapi or Pipecat.
LiveKit has positioned Turn Detector v1 as a framework-level solution that decouples conversational timing from specific STT vendors. The company released eot-bench as an open benchmark suite with real human-to-agent conversation datasets spanning 14 languages, establishing what it calls "common ground" for measuring end-of-turn models under production conditions rather than isolated clips. The benchmark evaluates models at real pauses under latency and interruption budgets, mirroring how live voice agents actually operate. This release strategy mirrors LiveKit's broader approach of building open infrastructure layers that work across model providers.
The competitive landscape for end-of-turn detection includes STT providers building endpointing directly into their models. LiveKit's blog post notes that built-in detectors from speech-to-text vendors tend to have incomplete language coverage and unpredictable performance variation, creating vendor lock-in for conversational feel. The company's benchmark results show Turn Detector v1 achieving 9.9% false-cutoff rate at 300ms latency budget, compared to 12.9% for Deepgram Flux and 27.7% for ultraVAD. At 600ms, LiveKit reaches 4.5% versus 5.5% for Soniox and 9.9% for Deepgram Flux. These results position LiveKit's approach as superior to both dedicated endpointing models and STT-integrated solutions across multiple operating points.
LiveKit's distribution strategy ties Turn Detector v1 directly to its cloud platform. The full v1 model runs at no cost for agents deployed to LiveKit Cloud and is the default turn detector, while v1-mini is bundled in Agents SDKs under Apache-2.0 licensing for local CPU inference. The model supports 14 languages including English, Arabic, German, Spanish, French, Hindi, Indonesian, Italian, Japanese, Korean, Dutch, Portuguese, Turkish, and Chinese. This freemium approach uses turn detection as a retention mechanism for LiveKit Cloud, similar to how the company has historically used open-source SDKs to drive cloud adoption.
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source