The trick to mishearing someone is not asking them to repeat it three times.

Speech recognition will mishear a resident. It is not a question of if. Soft voices, dentures, a dry mouth, an oxygen cannula, a regional accent the model saw little of in training — all of it pushes transcription error up exactly for the population we serve. The engineering question is not how do we never mishear. It's what happens on the turn after we do, because that turn is where companions are won or lost.

The death spiral

The default failure mode is brutal and familiar from every phone tree: I'm sorry, I didn't catch that. Can you repeat it? The resident repeats it, the same acoustic conditions produce the same bad transcript, and Companion asks again. Two rounds of this and a person who is already tired or anxious simply gives up. The repair strategy made things worse than no repair at all.

Repair the way people do

Humans almost never say please repeat that. We guess, and we check our guess in a way that lets the other person correct just the broken part. Companion does the same, and the choice of repair depends on how confident we are.

• High confidence — just proceed. Most of the time the transcript is fine and asking would be the rude thing.
• Medium confidence — confirm by using the guess in the reply, not by interrogating. Your knee, you said? lets the resident correct one word instead of re-uttering the whole sentence.
• Low confidence, but we caught a topic — ask a narrow question about the part we got, not a blanket repeat. I caught something about your medication — what about it?
• No usable signal at all — fall back to a warm, single open prompt, never a stacked apology, and never more than once before changing tack.

The other half is never repeating the same failing move. We track recent repair attempts within a conversation; if a clarification just failed, we do not run the identical clarification again. We escalate the kind of repair instead — switch from confirm-the-word to ask-the-topic, or, if it's plainly not working, gently offer to get a person rather than trap the resident in a loop with a machine that can't hear them.

Misrecognition near a clinical topic gets extra caution: we would rather under-confirm and route to a nurse than confidently log the wrong symptom. A mished chest versus chair is not a difference we treat casually, and a low-confidence health signal becomes an escalation, not an entry.

So when Companion mishears the resident in 214B, she mostly doesn't notice — it folds its best guess into the next sentence and she corrects the one word that was off, the way she would with a friend who's a little hard of hearing too. What she never gets is the cold loop of a machine apologizing for itself while learning nothing.