I can interrupt my AI now. Here's how talk-over works.

The most annoying thing about talking to a computer is waiting for it to stop talking. You ask a simple question, it starts a long answer, and you already know you only wanted the price. With a person you'd just say "sorry, just the price". So I taught Edison, my voice chief of staff, to handle exactly that.

On October 26 the first real cut-in worked: I started speaking, and once he was sure it was me, he stopped 71 milliseconds later.

What talk-over feels like

Edison is answering. I start talking. He goes quiet almost at once, hears what I said, and answers that instead. No button, no wake word, no "stop, stop, STOP".

Try it yourself in the demo on the home page: ask him to walk you through the launch plan, then tap "Talk over him" partway through.

How it works, in plain words

There are three parts.

1. Listening while he talks. A small speech detector (Silero VAD, which runs on my own computer) keeps listening to the microphone even while Edison is speaking. It doesn't turn speech into words. It only answers one question, about every 32 milliseconds: is someone talking right now?

2. Being sure it's really me. This is the hard part. A TV, a video or Edison's own voice coming back through the speakers can all sound like talking. So he waits for about 300 milliseconds of steady speech that is loud enough to be me, not the TV across the room. Then he stops, within about 71 milliseconds.

3. Doing the right thing after he stops. Stopping the sound is easy. Stopping the meaning is what matters:

The rest of the answer he didn't say is dropped and never acted on. If he was halfway through "I'll book that and send the email", nothing gets booked or sent.
His brain is told what I actually heard, so his next answer follows on naturally.
If he was reading an approval back to me ("That spends forty dollars. Shall I approve it?") and I talk over him, the approval stays waiting. It is never approved by an interruption.

The numbers

Moment	Time
The speech detector checks the mic	every 32 ms
Steady speech needed before he stops	about 300 ms
From "sure it's you" to silence	71 ms (first real cut-in)
Quiet that ends your turn	0.8 s (longer if you trail off)

So from the moment you start talking to the moment he goes quiet is roughly a third of a second.

The tricky parts

The TV. A speech detector hears a TV show as "someone talking". So the speech also has to be loud enough to be a person near the microphone before he stops.
Pauses. People pause mid-sentence. If he decided you were done after any short silence, he'd jump in while you were still thinking. So your turn ends after 0.8 seconds of quiet, and he waits longer when it sounds like you're trailing off rather than finished.
A switch to turn it off. On a screen where talk-over gets in the way (a noisy room, a demo on a big speaker), it can be turned off for that screen alone.

Why it matters for an AI that can act

Talk-over isn't just politeness. Edison can spend money, post and send messages, always after an approval. An assistant that keeps talking, or worse, keeps doing, after you've tried to stop it is not one you can trust. Being able to cut him off mid-sentence, and knowing the half-said part is thrown away, is part of what makes him safe to give real work to.

Want one?

You can talk to Edison in the demo on the home page, with his real voice and sounds. If you'd like your own, sign up to get Edison: you get him first and help decide what he learns next.