2026-10-04
First conversation with an AI on a real phone call
- What worked
- The box placed a call, a person answered, and an AI held a three-turn conversation with them: it greeted them, heard what they said, and answered. On that call the AI's words were spoken by the modem's own built-in voice — our fallback — because sending the model's audio into the modem was still failing. Later the same day, after the fix below, the box answered calls coming in and the AI spoke in its own voice. The whole path — driving the modem, handling the audio, and the link out to the model — is our own firmware running against a stock development board on a bench.
- The number
- 2 minutes · 3 turns · on a live mobile network
(fallback voice; the own-voice calls after the fix were shorter, because we hung up early) - What broke
- Plenty. The modem needs a different command to end a call than the standard says. It drops the first bytes of any file sent to it unless you wait after its prompt. One of its two serial channels can dial but cannot play audio. An idle model session timed out while waiting for the phone to ring, so now we connect only after pick-up.
- Honest limits
- This conversation is half-duplex. The box records, thinks, then speaks, so each turn takes several seconds and a caller cannot interrupt it. Real-time two-way audio needs an audio codec wired to the modem. The first one was ordered the same day.
- Hear it
- The AI's side of a later call that day, in its own voice. The caller is not
in these — the box plays a file into the call, so only our half exists as
audio. The voice is the model provider's synthetic voice, not something we
built.
GREETING · 3s
Hello, this is CallBox. How can I help you today?
REPLY · 5s
I'm sorry, I didn't quite catch that. Could you please tell me again what you need?
- Next
- Wire the codec, stream call audio both ways, and hold the same conversation without the pauses. Then put it in a box with a power button and a reset button.