Hi! I'm Freeman, a research engineer at Sesame. I'm the developer of the turnbench.sesame.com website and also a co-author. For the past bit, I've been working on models for real-time conversation at Sesame. We wanted an evaluation for turn-taking, the understanding of when to speak and when to yield in a natural dialogue.

I wanted to make sure the benchmark felt very accessible and understandable to people, so I built this in-browser dataset viewer at https://turnbench.sesame.com/conversations - it lets you inspect the audio directly with overlaid annotations. I thought HN might find it cool. As far as I'm aware, such an interactive web platform is a first for the academic speech community!

With TurnBench, we're also releasing a hand-annotated benchmark corpus, training set, public leaderboard, and our full eval protocol - which is based on human annotation.

Happy to answer questions!

Paper: https://arxiv.org/abs/2608.25218

Code: https://github.com/SesameAILabs/turnbench