0 %
Lab

Santa picks up, and the elves take notes

How do you really test real-time voice? Nine weeks after OpenAI opened it to developers, we put it in front of the toughest callers there are, children calling Santa. Every wish lands on a list for their parents, and the same voice work now points at support calls.

Live since Dec 2024

Santa picks up, and the elves take notes

It started with a question: how do you really test a real-time voice integration? Not with a demo script, but with the hardest callers there are. Children interrupt, giggle, mumble, change the subject mid-sentence, and know instantly when something sounds fake. If a voice can hold a conversation with a six-year-old on Christmas Eve, it can hold one with anyone.

On October 1, 2024, OpenAI opened its Realtime API: speech in, speech out, fast enough to hold a conversation. On December 2, Call a Claus was on the App Store. A child pressed call, Santa answered, and they talked, live, with no recordings and no script.

Santa became the test. He answers in the moment, warm and in character, and parents tell him a few things beforehand, so he knows about the new puppy and the science project.

The wishlist that listens

From the first version, a call never just ended. What the child told Santa became gift ideas for the parents, the way a real Santa would whisper them on the way out.

In the next version the list gets honest. It holds only what the child actually asked for, dated ("Asked on Oct 3"), with Santa's own ideas kept apart. Ask for the same bike twice and it counts twice instead of appearing twice. Nothing is invented, and only parents see it.

Where it is going

The plan from the start: make the night magical for the child, and the shopping easy for the parent. Every wish links straight to the right product, so "he asked Santa for a red bike" becomes the red bike in a tap. Affiliate links help pay for the magic, so it never needs ads. The shop lives with the parents; children only ever see Santa.

Beyond the North Pole

Everything Santa had to learn is what a good support call needs: answer at once, never talk over the caller, recover from an interruption, remember who is calling and what they said last time, speak their language, and hand off cleanly when a human should take over. That is where this work is heading next: real-time voice for support lines, proven first on the most demanding audience we could find.

In the next version

  • Santa calls you. Parents schedule the call, and the phone rings with a real incoming-call screen, even on the Lock Screen. Miss it, and Santa calls back.
  • Santa speaks their language. 84 of them, chosen once by a parent.
  • He remembers. Santa keeps track of what each child told him, and each character has its own voice.
  • He listens. Santa always hands the turn back, so a child is never talked over.
  • A sleigh tracker. Sleeps until Christmas, a live countdown, and what Santa is up to.
  • "Hey Siri, call a Claus." A call starts without opening the app.
  • The reaction. The front camera can catch the child's face during the call, and the clip is saved or shared only after a parent checks it.

Projects, press or anything else.