How does the moderator direct the AI avatar?

01Preview

That second’s head start belongs to the moderator

The moderator has the tablet in front of her. She calls the AI avatar by name, and its answer is there as text, usually just over a second before the audience hears what it says. Only the short opening word, such as a German “Also.” (“Well.”) or “Hm.”, comes without text. She does not have to approve anything. She reads along while it speaks and stays that second ahead of the room.

A mistake from a live test on a laptop shows what a stop is for. The avatar pushed back on a remark and pinned it on the wrong person. The moderator was the one who had said “We’re sticking with the weather.” At an event, a sentence like that would be on the tablet before it reached the room. If the moderator taps STOP then, the avatar goes quiet in under half a second.

So she stops the avatar over a sentence she has read, without guessing where its answer is heading. The tablet is the only device she operates. Someone from our team on site looks after the laptop and the audio interface.

02Stepping in

When the question comes from someone else

Usually the moderator calls the AI avatar by name and asks the question herself. Sometimes, though, a guest asks it, or it comes from the audience.

• One tap

  The ANSWER button

  One tap on ANSWER passes the last sentence someone said to the AI avatar as a question. Nobody has to say its name for that.

• Typed

  A question of her own

  By default it only hears the stage microphones. So when someone in the rows asks a question without a microphone, the moderator types it in, and the avatar answers the typed text exactly as she wrote it.

• Green

  In conversation

  After an answer it takes follow-up questions for a short while, even without its name. As long as that window is open, the tablet shows a green “In conversation” label.

03Transcript

If the transcript gets a word wrong, so does the avatar

Throughout the event, a live transcript runs on the tablet, with the names of the speakers. The AI avatar builds its answers from that text. If the speech recognition heard a word differently from what was meant, the avatar works with the wrong word.

The moderator sees that word in the transcript before the answer to it arrives. If a question came through garbled, she can stop the avatar and type the question in instead of calling on it again. Next to the transcript, the tablet shows a latency readout and a connection indicator.

04Decision

It could speak unprompted. We don’t let it.

Technically, it could chime in on its own. We set it up so that it doesn’t. Even the conversation after an answer stays in the moderator’s hands.

If she calls on a guest by name, the AI avatar goes quiet, even if its 12 s conversation window has not yet run out. A stop ends the conversation as well. After that, it only speaks again when someone says its name, taps the button or types a question.

So when nobody has called on it, its silence is not a malfunction. Who speaks when on stage is decided by a person, and that person has the tablet.

Try every move once before the event: a stop in the middle of an answer, a tap on ANSWER after a guest has spoken, a typed question. Then call on a guest by name and watch the green label disappear. On stage there is no time to hunt for a button.