At @hellomundoai, we recently worked with @sesame and others on TurnBench, an open benchmark for measuring turn-taking in spoken dialogue. One of the hardest challenges we faced: how do you collect specific conversational behaviours without making them feel staged? Ask someone to “interrupt more” or “sound angry,” and you often end up measuring their performance of the instruction rather than the behaviour itself. We wrote about what we learned building TurnBench: mundoai.world/research/the-n…