Four days in Hannover
We took an early build to a trade show floor. Almost everything we assumed about noise, turn-taking and etiquette turned out to be wrong.
By BabelGO Team
Our lab tests were clean, quiet and completely unrepresentative. So we took an early build to a industrial trade fair in Hannover and used it at a real booth for four days.
Here is what broke.
Noise was not the problem we prepared for
We had tuned for ambient noise — the steady roar of a hall. What actually degraded recognition was competing speech from the neighbouring booth two metres away. Ambient noise is easy to subtract. Another human speaking clearly in a language the model also supports is not.
Beamforming helped more than any amount of denoising. This is a large part of why we are building hardware.
People will not take turns
We designed a full-duplex mode where both speakers can talk freely. On the floor, it produced constant overlap and a translated audio stream that interrupted itself.
What worked was the opposite: an explicit half-duplex mode where one person holds the floor, speaks, and releases it. It is less elegant and it is what everyone actually used. Half-duplex is now the default for loud environments, and full-duplex is the option rather than the other way around.
The screen ruins the conversation
The single most consistent observation: the moment someone held a phone between themselves and a visitor, the interaction changed character. People addressed the device instead of each other. Eye contact collapsed.
The best sessions were the ones where the phone sat face-down on the counter and audio came from a small speaker. Nobody looked at the translation. They looked at each other, which is the entire point.
Jargon was decisive
Generic translation handled small talk beautifully and product conversation badly. Our part numbers came out as words. Industry terms were translated literally into nonsense.
Uploading a glossary of roughly 200 terms the night after day one changed the tool from a novelty into something the sales team asked to keep. That feature moved from "nice to have" to a launch requirement on the strength of a single evening.
What we changed
- Half-duplex became the default in noisy contexts.
- Speaker-side audio output is now a first-class mode, not a setting.
- Workspace glossaries moved into the private beta scope.
- We stopped believing lab benchmarks that we had not reproduced in a hall.
Four days cost us a quarter of replanning. It was the cheapest research we have done.