Should a translator show when it is guessing?

I’m building an offline speech-to-speech translator for mixed-language meetings, and I’m stuck on how honest the interface should be. Code-switching and technical terms are exactly where a smooth sentence can become a misleading one. I’m leaning toward preserving the original term and showing a small confidence marker instead of silently “fixing” it.

For uncertain segments, the listener could tap to replay the source audio, edit a local technical glossary, or add a correction to a device-only log. I don’t want every sentence to sound hesitant; that would make real-time conversation exhausting. But selective uncertainty around names, jargon, language switches, and low-confidence audio seems more trustworthy than polished nonsense. I haven’t seen evidence that visible hesitation automatically builds trust, so this is a design bet. What examples or interfaces would you suggest—and who thinks fully polished output is still the better choice?

I’d avoid a single confidence number. Users will read 0.82 as “82% correct,” which it isn’t. Mark the span, preserve the audio, and let the operator decide whether to interrupt. Also make the glossary correction explicit: changing a term should not silently rewrite old transcript segments.

The placement matters more than the existence of the marker. A red warning in the middle of a live meeting turns the translator into the main event. I’d use a quiet underline during speech, then a review tray for anything the system flagged. Trust can be supported without demanding attention every second.

“Never leaves the device” is a useful boundary, but not a complete privacy guarantee. Audio replay and correction logs still need retention controls, deletion, and a clear recording indicator. The safest default is probably short-lived buffers plus an explicit save action.

I’d separate uncertainty from disagreement. A translated phrase can be linguistically confident but still wrong for the meeting’s terminology. The editable glossary handles local vocabulary; confidence markers handle the model’s uncertainty. Mixing those signals will confuse users.

There are two users here: the person following the conversation and the person responsible for the record. Give the first a clean stream with lightweight cues, and the second a detailed local review view. One interface trying to serve both will feel either noisy or dishonest.

The cautious claim is the right one: confidence estimation may help identify unreliable segments, but that does not establish that visible hesitation improves trust. I’d test comprehension and correction behavior, not ask users whether the interface “felt transparent.”

Replay is cheap conceptually and awkward in practice. If the system waits for perfect segmentation, the replay target will be wrong; if it buffers too little, context disappears. I’d ship source-audio replay first and postpone fancy hesitation animations.