Researchers led by Freeman Jiang, alongside 18 co-authors including Shinji Watanabe and Hung-yi Lee, have released TurnBench, a benchmark for evaluating how well spoken-dialogue systems handle turn-taking: the real-time decision to hold the floor, yield it, or jump in. The benchmark centers on a 30-hour, triple-annotated corpus of two-person human conversations across six distinct interaction styles, paired with a standardized protocol for detecting end-of-turn moments and interruptions.
Testing 14 different turn-taking systems, the researchers found that end-of-turn recall held steady across conversation types, but false positives on interruption detection varied sharply by style, clustering in conversations dense with backchannels (the “mm-hmms” and short acknowledgments that keep a dialogue flowing). Human listeners, by contrast, start speaking a median of 151 milliseconds before a turn actually ends. No tested system matched that timing without triggering excessive false alarms.
Alongside the corpus, the team is putting out a leaderboard anyone can view, a browsable viewer for the data, and a separate 104-hour training set. The paper has been submitted to IEEE SLT 2026.
For teams building voice agents, TurnBench is a concrete way to stop guessing whether a system’s interruption handling actually holds up outside a demo.
Per the paper’s abstract on arXiv, posted August 25, 2026.
