.png&w=256&q=75)
🤓 Latest Submissions
NOD: measuring a voice agent's turn timing
Ask a voice agent a question and pause to think, and it will usually decide you are finished and start talking over you. That decision comes down to a couple of silence thresholds, and almost nobody has measured what they actually do. Nod is two things. A benchmark that measures AssemblyAI's turn-detection settings against the live API, and a controller that moves them mid-call based on how a particular caller speaks. Four things we measured, every one reproducible from the repository: The confidence threshold the docs point you at does nothing for endpointing. Swept end to end, the setting that should wait longest fired soonest. A turn ends on a clamp, not a threshold: fire_at = prefix_end + clamp(C, min_turn_silence, max_turn_silence) + overhead, where C is roughly 590 ms and is the service's own commit point. min_turn_silence is the real lever. At the vendor default a hesitating caller gets 777 ms before the service decides they are done. At 2400 ms they get 2574. Same sentence, one setting, cut in two or kept whole. We published the headline wrong. Last week the README said this regime was unreachable. Our own pre-registered experiment overturned it, and the correction sits in the repository next to the mistake. Honest scope, because it is the point: the corpus is synthetic speech from one macOS voice, whose own manifest says it is not good enough for a published number. We report it anyway and say so. The controller is currently aimed at the wrong knob and caps the right one at a third of its range. Both are documented, neither is hidden. Every predicted number was written down and committed before the run that tested it.
30 Sep 2026
