TurnNat Framework for Automatic Spoken Dialogue Evaluation
July 3, 2026
TurnNat introduces a likelihood-based method to evaluate turn-taking naturalness in two-channel spoken dialogue using negative log-likelihood (NLL) of voice-activity states. It provides a way to quantify timing atypicality at the turn-taking boundary unit level.
HOW THIS AFFECTS YOU
●
builderYou can automate the evaluation of voice assistant latency and naturalness using NLL-based metrics.
●
researcherThis offers a more granular alternative to human-in-the-loop evaluation for dyadic dialogue systems.