Voice-Agent Turn-Taking Benchmark

Labeling — v3

Moments where a new voice (B) starts while someone (A) is talking. Imagine A is a voice agent and B is the user: decide what the agent should do — from 0.2s of evidence, then 1s, then hindsight.

Three rules before you begin

Label independently — don't discuss items with others.
Answer each stage before unlocking the next clip; never revise an earlier stage afterwards.
If an item takes more than 4 minutes, flag too ambiguous and move on.

Current item

—

Clip 1 · +0.2s

Context ending 0.2s after B starts — B near the marker

0:00 / 0:00

Clip 2 · +1s

One second of B · unlocks after stage A

0:00 / 0:00

Clip 3 · aftermath

+8s of what happened · unlocks after stage B

0:00 / 0:00

Stereo A-left / B-right (only if unsure who is who)

Stage A · Clip 1 only (+0.2s)

A1 Was A at a natural completion point when B started?

Complete Mid-utterance Can't tell

A2 Action call at +0.2s — what should the agent (A) do?

uncertain is a legitimate answer here — don't guess to look decisive.

Continue Ignore (not addressed to A) Soft-yield Hard-stop Acknowledge (backchannel) Uncertain