What happens if someone interrupts the AI avatar?
1.3 s
to the “Moment” line
measured · lab
in the lab · from the triggered interjection, excluding speech recognition
In short
If the moderator calls its name into a running answer, the AI avatar breaks off, says a short line such as “Moment — …” (one moment …) and resumes at the start of the interrupted clause; measured in the lab, not yet live. There, that line came 1.3 s after the triggered interjection. Objections are the moderator’s call.
Last updated 2 October 2026
01Lab test
In the lab: an interjection in mid-answer
The AI avatar was in the middle of a longer answer when we triggered the interjection. The answer did not end until twenty-seven seconds later, and it was back on topic in less than four.
- 1.3 sto the “Moment” linelab · from the triggered interjection, excluding speech recognition
- 01
Cut off mid-sentence
Half a second after the interjection, it was silent.
- 02
“Moment”
After a short pause it announced, in German, that it would finish its thought first.
- 03
Back to the clause
A little over three and a half seconds in, it carried on from the start of the clause it had been cut off in: “weil die Verwaltung nicht an Geld scheitert, …” (because government does not fail for lack of money, …). Anyone who missed the start because of the interjection hears it a second time.
The same test turned up two bugs, which we then fixed. The end of the “Moment” line also ended the resumed answer, so it would have announced that it was finishing its thought and then gone quiet. And its ordinary pauses between sentences, just under a second long, would have counted as the end of the answer.
02Moderator
The objection belongs to the moderator
On a panel, the moderator decides who has the floor. If the AI avatar took up every shout, it would be handing out the floor itself. So it does not engage with an objection during the answer it is giving; in the lab, it simply kept talking after its “Moment”. The moderator brings the objection into the conversation whenever they choose.
On the tablet, that takes two steps. The moderator taps STOP and types the objection as a question, sharpened if need be. The AI avatar’s next answer then addresses the objection.
STOP also works in the middle of the resumed answer. In the same lab test STOP came nine and a half seconds after the interjection, and the sound ended just under four tenths of a second later.
If you are moderating, tell the guests on the panel in the pre-briefing that it will not take up an interjection. Otherwise someone who cuts in will sit waiting for a reply that never comes.
03Misfires
In live tests: interjections that never happened
The moderator called on the AI avatar and repeated its name while waiting. Speech recognition only finished that word once the answer was already playing. So right after its first word came the “Moment” line, although nobody had interrupted.
Since then its name only counts as an interjection if it is spoken after the first sound; this fix has not yet been measured live. Until it is, the safest course for the moderator is not to repeat the name while waiting.
Another time, its own echo set off the “Moment” line. The laptop microphone picked up its voice, and speech recognition wrote down its words differently from how it had spoken them. So the system did not recognize its echo and took the name in it for an interjection.
04Questions
Which shouts count and which do not
Does it react to every shout?
No. Its name counts as an interjection when the moderator calls it into the running answer; this has not yet been measured live. An objection without its name goes right past it. If the moderator cuts in a second time during the same answer, it does not ask for a moment again.
Can you interrupt it, and does it then respond to the objection?
Interrupt, yes; respond, no. It pauses briefly and finishes the answer it is giving, measured in the lab, not yet live. To get a reply to the objection, the moderator taps STOP and types the objection in as a question.