LiveKit EoT Benchmark

This dataset contains a minimal public subset of the LiveKit end-of-turn benchmark. It is sampled from the full release-ready dataset preserved on the full revision.

Links

• Blog post: Solving end-of-turn detection
• Benchmark code: github.com/livekit/eot-bench

The sample is deterministic: up to 400 turns per language were sampled without replacement using seed 20260603. Languages with fewer than the requested sample size include all available rows: ar (381 available), id (396 available), ja (356 available), ko (377 available), zh (344 available).

Configs

• all: all available sampled languages.
• Language configs such as en, de, and ja: one language only.

All configs use the same schema. English rows include language = "en".

Schema

• id: globally unique turn id.
• audio: 16 kHz mono WAV audio.
• language: normalized language code.
• duration: audio duration in seconds.
• silence_spans: candidate silence spans with start and end.
• words: word-level transcript timings.
• messages: prior conversation messages.

Rows whose final end-of-turn silence span is shorter than 0.2s are excluded from this release split, so every turn can be evaluated at a 0.2s score point. Rows with any silence span longer than 5.0s are also excluded so the benchmark focuses on ordinary endpointing pauses.

Languages