By Juste Thales Gnimavo (CEO, ZeroSuite) and Claude Opus 5.5 (Claude Code)
On November 2, 2026, Juste tested the "I'm presenting" mode of Déblo Duo. One speaker talks in French; the room hears a live English translation, each listener on their own phone or laptop. He said hello, asked how everyone was doing, and stopped talking. The translation did not stop. "Are you okay?" came back. And again. And again.
The easy explanation was the one everyone reaches for: the AI hallucinated. We did not write that down, because it explains nothing and fixes nothing. We read the session instead.
Déblo Duo is our real-time interpreter. (Why it is a separate app is in Why Déblo Duo exists.) The speaker's audio goes to Google's real-time speech translation model through a server-side worker on our real-time audio layer, and the translated audio goes back out to every listener.
What the transcript said
The session ran from 15:02:25 to 15:03:29 UTC: 63 seconds, 26 turns, French to English. We pulled the stored transcript and the logs of the worker container that had handled it. Here is the source side, with the worker's timestamps in seconds:
1753.87 « Bonjour. » → "Hello."
1759.93 « Bonjour. » → "Hello." (+6.1 s)
1765.98 « Comment est-ce que vous allez ? » → "How are you?"
1773.95 « Tu vas bien ? » → "Are you okay?"
1777.99, 1781.93, 1785.96, 1790.00, 1793.94, 1797.97, 1801.90, 1805.94:
« Tu vas bien ? » → "Are you okay?"
1809.36 speaker leavesThe first thing to notice is which side repeats. It is not the translation that loops on a fixed input. It is the source. The model was being fed "Tu vas bien ?" in French, over and over, and translating it faithfully each time. After the first one, it was fed that sentence eight more times, roughly every 4.0 seconds.
The second thing is that Juste never said "Tu vas bien ?". He said "Comment est-ce que vous allez ?". The source drifted, once, and then stopped drifting.
Reading the pattern
Three features of that transcript pointed at one cause before we looked at anything else.
A fixed period. Eight repetitions spaced by 4.0 seconds, to within a tenth of a second. Hallucinations on noise are not periodic. A physical loop is: its period is the sum of the delays around it. Here, translation latency, plus playback, plus the end of the turn.
A drift, then a fixed point. "Comment est-ce que vous allez ?" became "How are you?". "How are you?", played back into the speaker's microphone, came out of the source transcription as "Tu vas bien ?". "Tu vas bien ?" became "Are you okay?", which came back as "Tu vas bien ?" again. That is the signature of a lossy round trip: each pass loses a little until the sentence lands on a form that survives the trip unchanged. After that, the loop is stable.
The same shape elsewhere. The second "Bonjour." 6.1 seconds after the first was the same thing, one round trip of "Hello." that happened to survive.
The cause: a listener on the web was in the same room, playing the English translation on its loudspeaker. The speaker's microphone picked it up. The model translated it back into French, as if the speaker had said it, and the listener's loudspeaker played the new translation. Juste had become a relay in his own feedback loop.
We ruled out the two other candidates explicitly. It was not a loop inside our bridge: every repetition was a distinct source input with its own timestamps, coming in through the microphone. It was not ambient noise: noise does not produce a coherent French sentence on a 4.0-second clock.
Browser echo cancellation was on. It does not help here. It removes what the device plays itself, and the speaker's device was playing nothing. The sound came from another machine across the room. The presentation screen already advises listeners to wear headphones. A warning is not a guarantee in a room full of people who have just scanned a QR code.
Why the earlier echo fix did not cover it
We had met a cousin of this bug five days earlier. On October 28, in two-phone calls on speakerphone, a phone could hear its own translation through its own speaker and send it back to the model.
That one we fixed on the server, not in the app. While the track a phone is listening to is playing, the worker replaces that phone's microphone with silence toward the model, and tells the phone so it can show "Listen to the translation" instead of an active microphone. We chose the server over muting the microphone in the app because the app's speech detection reacts too late, and it works per participant, not per audio track. The server knows exactly which track each phone is hearing and when it is playing.
That fix is precise because it pairs a microphone with a known speaker: the same phone's. In a presentation, the loudspeaker belongs to a listener the worker treats as audience, and the microphone belongs to the speaker. Nothing pairs them. The worker cannot know that a listener across the room is physically near the speaker's phone.
Where the guard had to sit
The brief for the fix proposed the obvious place: compare each new translation to the last few translations and drop repeats.
We did not do that, and said so before writing code. In our bridge, the translated audio is sent to the listeners' track the moment it arrives from the model, at the same time as its text. By the time a translated sentence is complete enough to compare, its audio has already left. A guard on the translation would have removed the subtitle and left the listener hearing the burst. The model also gives us no reliable end-of-turn signal for a translation, so there is nothing to wait for.
So the guard sits on the source segment, which we see earlier. In presentation mode only, when a new source segment matches one of the most recent ones (after normalizing case, punctuation and spacing) and starts soon after it, the worker treats it as an echo. The translated audio for that direction is dropped, and neither the source nor the translation is published to subtitles or to the end-of-call transcript. The guard releases on a short silence, or as soon as the speaker says something that is not a repeat of a recent echo.
The one-phone and remote modes have no code path through it.
What we verified, and what it costs
We replayed the twelve source segments measured in that session through the guard. It kept the first "Bonjour.", "Comment est-ce que vous allez ?" and the first "Tu vas bien ?". It suppressed the eight repetitions after that, and the second "Bonjour.".
The first "Tu vas bien ?" is itself an echo, the drifted one. The guard cannot catch it, and we do not pretend it can: it does not repeat anything yet, and from the text alone it is indistinguishable from a speaker asking a new question. So the listener can still hear one sentence twice, at most, instead of a burst.
Then we ran a simulation inside the production worker container, with the real audio SDK: speech, an echo four seconds later, more speech. The first and last were played in full. The echo played zero bytes, and the worker logged the suppression with the amount of audio that had leaked before detection: none. Only the two real utterances reached the final transcript.
The guard has known costs, written down with it. A speaker who repeats the exact same sentence quickly loses the second one. The first audio packets of an echo can leave before the source transcript is complete enough to match; the worker measures that leak on every suppression so we will see it in production. On the speaker's own screen, a partial line of an echo can stay visible.
The same session also gave the presenter a mute button. "I'm presenting" was the only mode without one.
What it changed, and what is still open
The fix is deployed. It is not proven. The observation we owe is specific: a real speaker and a real listener on a loudspeaker in the same room, the suppression events in the production worker's log, and a transcript with no burst. That needs two devices in one room and a person to speak, and until it happens, the item stays open.
What changed is how we read a misbehaving model. "The AI hallucinated" is a description of a symptom, and it closes the investigation before it starts. The transcript had a fixed period, a drift and a fixed point. Each of those has a physical meaning, and together they named the cause before we had touched the code: two devices, one room, one open microphone.
Every investor demo of Duo happens in a single room. That is exactly the condition of this bug, which is why we moved this fix to the front of the queue the day we found it.