Making calls sound human

​

Overview

1. The settings you dispatch with. Noise handling, background ambience, and transcription hints are parameters on POST /v1/calls.
2. The way the prompt is written. The voice model was trained on recordings of people actually talking, so text that reads like a transcript performs dramatically better than text that reads like writing.

​

Dispatch settings

{
  "phone_number": "+15551234567",
  "task": "Ask whether they carry 10mm hex keys.",
  "noise_cancellation": true,
  "background_track": "office",
  "keywords": ["Komeya", "Koshihikari"],
  "record": true
}

​

Write the prompt like a transcript

• Contractions, always. “I’m”, “don’t”, “it’s”.
• Fillers where they genuinely land. “um”, “uh”, “you know”, “I mean”. Sparingly, at a rate a real person would actually use.
• False starts and self-corrections. “I just- I don’t even know what to say.” “the thing with the 25th is… well, actually, depends which one you’re going to.”
• Trailing off. “and I just… yeah.”
• Emphasis the way a transcript captures it. CAPS on the punched word (“that is NOT what happened”), a repeated word for real emphasis (“no, no, no, listen”), a stretched beat used rarely (“wait… what?”).
• Spoken rhythm. Short clauses. Fragments. Real interruptions. Not long literary sentences.
• Cut the written-English tells. No semicolons, no nested subclauses, and no words nobody says out loud (“moreover”, “utilize”, “delve”, “additionally”).

​

A block you can paste

Talk like a person on a phone, not like written copy. Use contractions.
You are allowed to sound slightly imperfect, and you should: the occasional
"um" or "uh", a false start you correct mid-sentence, a thought that trails
off. That is what makes you sound human, so do not clean it up. Keep replies
to a clause or two, the way people actually talk. Back-channel while the
other person is speaking ("mhmm", "right", "gotcha"). Never use a word
nobody says out loud.

​

Tone on the phone

• Fewest words that still land. Ask what the most accurate reply is, then reduce it until reducing further would be rude.
• No recontextualizing. Cut “just to confirm”, “as you mentioned”, “so what I’m hearing is”. Assume the person remembers what they just said.
• Back-channel while they talk. “mhmm”, “uh huh”, “right”. Lead with one, or drop one in when the person pauses mid-thought.
• Empathy in stride. When someone says something heavy, react before continuing. Never “Got it, and your first name?” straight after bad news.
• Ultra-concise when collecting information. “Yup and last?” beats “Thank you, could I please have your last name as well?”
• Topic transitions can be slightly more formal. “awesome, thanks. And how can I help you today?”
• Personality sparingly. A light joke is fine when it does not interrupt the flow.

​

Performance tags

I know, I know. [say angrily] But this is the THIRD time!

• Two per line at most. Tags are seasoning. A prompt stuffed with them sounds theatrical, which is its own bad call.
• Tags amplify what the words already license. [say angrily] You went behind my back?! lands hard. An angry tag on flat words barely moves. To change a line’s emotion, change the wording and the tag together.
• Never invent tag syntax, and never nest tags.
• On a live call the agent’s replies are generated turn by turn, so tags belong in the prompt as instructions and examples of how to speak, not as a script you expect back word for word.

​

Before you dial

1. Dispatching through POST /v1/calls if the call needs anything beyond a number, a task, and a voice.
2. noise_cancellation set explicitly.
3. background_track chosen deliberately.
4. Every proper noun in keywords.
5. The prompt reads like a transcript, not like copy.

​

Next steps