LiveKit ships Turn Detector v1 with 14-language endpointing

• LiveKit says Turn Detector v1 predicts end-of-turn directly from speech, not transcripts, by combining semantic and acoustic cues in one model, according to LiveKit's launch post.
• The company claims the model delivered its best tested accuracy and latency across 14 languages, per LiveKit's launch post.
• LiveKit also shipped a smaller Turn Detector v1-mini for fast CPU inference, and LiveKit's v1-mini post says it is bundled with the Agents SDKs.
• Benchmarking ran on LiveKit's eot-bench post, which points to an open end-of-turn benchmark suite and datasets.

You can jump from the launch post to the eot-bench repository, and LiveKit also slipped in exact SDK bundle versions in the v1-mini announcement. The short version is simple: endpointing moved from transcript heuristics to a speech-native model, and the smaller model is already packaged for agent builders.

LiveKit framed the main change as modality, not packaging. Turn Detector v1 listens to raw speech and predicts when a speaker is done, instead of waiting for transcript text and then deciding from words alone.

That matters because the model is using two signal types at once:

• semantic cues from what was said
• acoustic cues from how it was said
• one end-of-turn prediction that tries to balance latency and interruption behavior

LiveKit says that setup produced its best tested accuracy and latency across 14 languages on Cloud in the announcement.

LiveKit tied the release to eot-bench, which it describes as an open benchmark suite and datasets for end-of-turn detection.

eot-bench

The interesting detail here is the benchmark choice. LiveKit is not only claiming a model win, it is pointing readers at the evaluation harness it used, which gives voice engineers a concrete place to inspect how endpointing is being measured via the benchmark post.

The second ship was Turn Detector v1-mini. LiveKit says it keeps the same architecture as v1, but shrinks the footprint for fast CPU inference.

According to LiveKit's v1-mini post, the bundled SDK versions are Python 1.6.1 and TypeScript 1.4.7. That is the most practical rollout detail in the thread, because it pins the smaller model to specific Agents SDK releases instead of leaving it as a vague future integration.