What can you not rely on yet?

01Interjections

We know the interjection only from the lab

If the moderator calls its name into a running answer, the AI avatar says a short line such as “Moment — den Gedanken führe ich noch kurz zu Ende” (one moment, let me finish this thought) and resumes at the start of the interrupted clause. In the lab, the “Moment” line was heard 1.3 s after the interjection. We triggered the interjection there by command, with nobody calling out, so the time for speech recognition is not in that figure.

Live, at the laptop with people in the room, the “Moment” line has so far come only by accident, for instance on a name added late or on the AI avatar’s own echo, which the built-in microphone picked up. No run has yet caught an interjection as the moderator means it. That is why this capability is listed as “measured · lab”.

02Language

The characters do not speak English yet

For now, all four characters understand and speak standard German and nothing else. Speech recognition and voice are set up for it, and we have not tested any of their capabilities in English.

The words they listen for are German too: “Danke” (thank you) with the name ends the answer, “kurz” (briefly) or “erklär uns” (explain to us) sets its length. An English “thank you” does not stop them.

03Summaries

Long discussions: summarized only in the lab

Whatever was said in the last few minutes the AI avatar keeps word for word; everything earlier it summarizes as it goes. We measured the summarizing in the lab, with prerecorded discussions that ran well beyond what it keeps word for word.

Our runs at the laptop with people in the room were shorter. So live, it has not yet had to summarize anything.

04Until then

What to do until it is proven

• measured · lab

  When someone cuts in

  Expect an objection from the panel to come up only after its answer; a shout without its name goes right past it. If the objection should be heard right away, the moderator taps STOP. The button worked live in under half a second.

• standard German · not yet English

  When your event is bilingual

  Plan the round in which a character takes part in standard German. If a question comes in English, the moderator puts it to the character in German.

• measured · lab

  When the discussion runs long

  If it should respond to a contribution from the start, the moderator briefly sums up that contribution in the question. The question itself it has word for word.

We want to prove this at pilot events, which we offer on pilot terms. What we measure there, we will publish with setup and values.