One Moment: the voice agent that waits for you

One Moment is a voice agent whose job is to make the other person wait, so a person with aphasia can make their own phone call. Robert had a stroke. He knows what he means, but finding the word takes six seconds. Most voice agents treat that silence as the end of his turn: on one recording the AssemblyAI defaults ended his turn 0.37s after he stopped, and the word arrived 6.08s later. Measured on speech we did not record. On 120 sentences from 8 speakers with dysarthria (TORGO), streamed live, an ordinary agent put words in the speaker's mouth in 65 of 120. One Moment did it in 11 of the 57 it chose to speak, asking on the rest. Ordinary settings split 48 sentences mid-way; our patient ear split 0. Through a real 8kHz telephone encoder: 46 against 0. Checks on ourselves. On 60 sentences of fluent speech (LibriSpeech) it speaks on 49, waits a 1.7s median, and splits none, so the patience does not get in the way of people who do not need it. Pricing the turn setting itself: at the vendor default the turn closed while the speaker was still talking on 4 of 12 sentences. We publish the row that argues against us, that 2400ms would buy the same for 3.7s less waiting. How it uses AssemblyAI. Two Universal-Streaming sessions listen to one microphone, configured against each other, so their disagreement is the uncertainty signal. keyterms_prompt biases one ear only, so the unbiased ear can falsify a boosted word. ForceEndpoint ends a finished sentence early. The Voice Agent API points its LLM at our endpoint, which is not a language model: it returns only a line deterministic rules approved, or nothing, so the voice cannot invent a word. The pre-recorded API grades every call afterwards. Speech-to-Speech Relay is federally mandated at $8.4822 a minute and the FCC calls it under-used. Not tested with people who have aphasia. Not a medical device. Every number is re-derived by npm run verify. The engine installs on its own: npm install @one-moment/core.