AURA51.aiBeta

interactive ai agents for events

A conversation partner.
Not a chat window.

Aura51 puts an AI character visibly on your panel instead of off to the side: it follows the entire discussion, takes a position and speaks when the moderator calls it by name. Four characters — you decide who takes a seat. Albert is proven live; Vera, Tina and Ken are draft roles.

Scroll

01The cast

Four stances. One system.

They all run on the same technology and can do the same things. What sets them apart is face, voice and stance — and you decide those per event. Albert is proven live; Vera, Tina and Ken are draft roles.

02What they all share

The same technology, four characters.

The technology is the same for all four characters. The measurements come from the tests with Albert — so far he is the only character who has stood on a stage.

  • They follow the entire discussion.
  • They speak when the moderator calls them by name — and for twelve seconds afterwards, even to a bare “And why?”.
  • They understand and speak standard German and English — no dialect.
  • They can be stopped in under half a second.
  • They invent no numbers and say when they do not know.

What you decide

  • The name — phonetically distinct, not the name of anyone on the panel.
  • Face and voice.
  • The stance: contrarian, skeptical, explanatory.
  • The briefing: topic, participants, role.

03Where they appear

The role your panel is missing.

Four occasions, the same character — what changes is the stance you give it.

  • Conference panel

    An AI participant who challenges the discussion — visible on the screen, not a chat window off to the side.

  • Customer event

    Albert argues the position of the product. Or the position of its critic — the role is configured per event.

  • Internal town hall

    The devil's advocate no one else dares to play.

  • Education and universities

    AI as a visible conversation partner on the panel, not as a tool behind the stage.

04On stage

Eight things measured live.

Measured means: it occurred in this week’s live tests, with people in the room and natural speech.

  • Listening

    Albert knows every statement of the last few minutes verbatim, and the rest of the discussion as a running summary.

    measured · live
  • Answering when called by name

    “Albert, how do you see this?” — he responds. If someone merely mentions him (“Albert said earlier …”), he stays quiet.

    measured · live
  • Staying in the conversation

    After an answer he listens for twelve seconds for follow-up questions, even without his name: “And why?” is enough. He does not respond to a “thanks” or an “okay”; once someone else is called on, the conversation is over.

    measured · live
  • With reference and a position

    He addresses people by name, disagrees, takes a position — as long as the question requires, not as if he were in a chat window.

    measured · live
  • Choosing his own length

    “Briefly”, “in one sentence” or “yes or no” gets one sentence; “explain that to us” or “give an example” gets a short explanation; everything else two to three sentences. No button needed, the moderator keeps both hands free.

    measured · live
  • Starting naturally

    He opens with a neutral “Well.”, “Hm.” or “So.” and composes the answer knowing that opening. A “Yes.” followed by a “no” no longer happens.

    measured · live
  • Understanding split calls

    “I will now ask Albert. Albert, what do you make of …?” produces an answer to the whole question. If the moderator keeps talking, he waits.

    measured · live
  • Letting himself be stopped

    One tap on STOP and he falls silent in under half a second.

    measured · live

05Not yet on a stage

Three we are not claiming yet.

They are built and verified in the lab or in testing — but they have not stood in front of an audience. That is why they are down here and not further up.

  • Reacting to interjections

    If someone cuts into his answer, he says “One moment — let me finish this thought” and resumes exactly where he was interrupted.

    measured · lab
  • Being stopped by voice

    “Thank you, Albert” ends his answer just like the STOP button. The same “Thank you, Albert” in ordinary conversation does not call him in.

    verified in testing · not yet live
  • Knowing who is in the room

    The names of the moderator and the guests are entered on the tablet before the panel; he addresses them correctly and recognizes them in the transcript.

    verified in testing · not yet live

06The bridge

Everything runs through one point.

Every answer appears as text on the tablet before the room hears it — about 1.0 seconds ahead of the audio. What gets stopped is the content, not a hunch — and without a name call or a button press no character says anything.

  • Answer preview

    Every answer appears as text on the tablet before the room hears it — the text runs about a second ahead of the audio. The moderator can stop it based on the content, not on a hunch.

  • Stop, feed, type

    Cut him off — on the tablet or with “Thank you, Albert” —, hand him the last sentence someone said as a question, or type a question.

  • No speaking unprompted

    Without a name call, a button press or a running conversation, Albert says nothing. The conversation ends twelve seconds after his answer, or sooner if someone else is called on.

  • Live transcript

    With speaker names, a connection indicator and a latency readout.

07Measured

Every number with its conditions.

  • 1.8–1.9sto the first audiowith people in the room, median of live tests 5 and 6 · 1.5–1.9 s in the lab
  • 334msuntil he falls silentSTOP on the tablet · measured at 334 and 374 ms
  • 12sconversation without a name callafter every answer · after that it takes his name again
  • 1.3sto the “one moment” lineafter an interjection · real voice

08The signal path

From the microphone to the screen.

  1. 01

    From the desk into the laptop

    The panel microphones go from the mixing desk through an aux send into a laptop.

  2. 02

    Transcribing

    A speech recognition service transcribes live, speakers included (Deepgram Nova-3, names via keyword boosting).

  3. 03

    Recognizing and composing

    The “brain” recognizes the call, checks whether Albert is really the one being addressed, and composes the answer — while the moderator is still finishing the sentence.

  4. 04

    Speaking

    A photorealistic avatar speaks the answer on screen in a natural voice.

09The flat line

What does not work.

A character whose most important rule is “invent nothing” starts with itself. So here is what it cannot do — before you have to ask.

What it will not do — because we built it that way

  • Speak on its own initiative — without a name call or a button press it says nothing.
  • Invent numbers or studies. If it does not know something, it says so.
  • Record voices from the audience.

What it cannot do — because the technology does not allow it

  • Answer faster than roughly 1.6 seconds — that is how long pause detection and speech synthesis take together.
  • Swiss German. It understands and speaks standard German and English, dialect it does not.
  • Keep talking when no one calls it: twelve seconds after an answer it takes its name again.
  • Work without an internet connection.

10Asked

What organizers want to know first.

  • Does he just talk whenever he feels like it?

    No. Only when called by name or by button — and in the twelve seconds after an answer, for follow-up questions. After that it takes his name again.

  • What if he says something wrong?

    The moderator reads every answer before the room hears it and stops him on the tablet in under half a second. Stopping him with “Thank you, Albert” is verified in testing, not yet on a stage.

  • Does he know who is speaking?

    Yes — verified in testing, not yet on a stage. The moderator has a dedicated microphone, and the remaining speakers are assigned once before the panel.

  • Can you interrupt him?

    Yes — measured in the lab, not yet on a stage. He reacts to it, asks briefly for patience and finishes the thought.

  • How fast is he?

    Around two seconds to the first audio, even with people in the room — comparable to a person taking a moment to think.

  • How long are his answers?

    As long as the question requires. “Briefly” or “yes or no” gets one sentence, “explain that to us” a short explanation, everything else two to three sentences. Nobody has to press a button for it.

  • What languages do the characters speak?

    All four understand and speak standard German and English. Swiss German they do not — neither understand nor speak it.

  • What happens if the internet drops?

    None of the characters work without an internet connection; it says so under “What does not work”. For the worst case the moderator has a rehearsed line, and the discussion carries on without them.

11Arrival

Who do we put on your panel?

You bring the topic, the participants and the stance you want. We bring the character, a voice sample for name recognition and a dress rehearsal.

  • A mixing desk with one aux send for the microphones (pre-fader, without the avatar return) and an audio interface.
  • A laptop (Mac) with internet — the three services run in the cloud, the transcript stays local.
  • A screen or monitor for the avatar, and a speaker path back into the desk.
  • A tablet for the moderator on its own network (hotspot or cable), not on the venue Wi-Fi.
  • Topic, participants and role as a briefing; a short voice sample for name recognition; a dress rehearsal.

Fields marked * are required.

AURA51.AIEbnetstrasse 28810 HorgenSwitzerlandpanel@aura51.ai

Aura51.ai · Beta, as of September 20, 2026. Beta means nothing vague here: the characters are built and measured in live tests, but they have not stood in front of an audience yet. All measurements come from lab and live tests with “Albert”; the evidence level is stated with every claim.

Developed with love in Horgen, Switzerland