A voice helper that doesn't just talk — it sets the timer while your hands are in the dough

Here's the idea in one line: you talk out loud — "set a timer for ten minutes", "how much is 200 grams of flour in cups", "read the next step" — and the helper doesn't just answer, it actually does it: starts the timer, converts, moves through the recipe. Hands in the dough, you never touch the screen. It listens and presses the buttons for you.

And this isn't the out-loud practice partner that can only chat. This one works with its hands while yours are busy.

Why this just became possible

A voice helper used to be simple: listen → answer in voice. It could say anything — but do nothing. It would say "set a timer" back to you, but you'd set the timer yourself, wiping your hands on your apron first.

On July 6, 2026 OpenAI shipped gpt-realtime-2.1-mini — a speech-to-speech model that calls your functions on the fly. In its own words, the model "plans a step, calls your function, then answers." So mid-conversation it decides "a timer is needed here", fires your set_timer(10) code, and says out loud "done, ten minutes." And it answers about a quarter faster than before — the pause is nearly gone. That's what turns the voice from a chatterbox into a helper.

What you'll learn

• Tools for voice (function calling). You describe 2–3 small functions — "start a timer", "convert grams to cups", "add to the shopping list" — and the model decides when to call each. That's tool use, only not in text — mid-conversation, in a live voice stream.
• Real-time speech-to-speech. Not "record → wait → hear the answer", but a constant voice stream both ways. A different kind of app from the usual request-response.
• Functions are the helper's "hands". You'll feel it plainly: its character and words come from the prompt, but what it can do is the set of functions you gave it. Add add_to_list and it can build a list; leave it out and it can only talk.

A ready starter prompt

Don't ask for "make a voice helper for cooking" — you'll get a talker with no hands. Name the model, the functions, and the behavior:

The strong prompt leaves no guesswork: you can see the model, the three "hands" functions, and that it should answer short and confirm what it did. The helper comes out working on the first try, not chatty.

What you get

Hands deep in the dough. You say "set a timer for twelve minutes" — snap, the timer runs, the voice says "done, twelve." "How much is 200 grams of flour in cups?" — "about one and a half." Finished a step — "next", and it reads the following one. At the end: "add eggs and milk to the list" — noted. You never wiped your hands, never tapped the screen.

Weekend plan

1. Saturday. Get a voice connection to the model up: you talk, it answers in voice with no pause. No functions yet, just a live conversation.
2. Sunday. Add one function — set_timer. Teach the model to call it and say the result out loud. Working? Add the second and third.
3. Point it at a real recipe of yours and cook something without touching the phone once.

Start with one function. "A helper that does everything" can wait — first let it reliably set one timer by voice.

Источник: OpenAI gpt-realtime-2.1 and mini: speech-to-speech with reasoning and tool use (MarkTechPost)