There is a room at the centre of every ambient AI scribe on the market. It has a door that closes. One patient sits in it. The consultation runs fifteen or twenty minutes, in one language, and a microphone on the desk hears both voices cleanly. That room is real — it sits inside well-resourced health systems, mostly in North America and Northern Europe, and it is where nearly all published scribe evidence comes from.
Now walk into a high-volume outpatient department in Delhi at eleven in the morning. The door is open because there is a queue. A ceiling fan runs directly above the desk. A relative answers questions on the patient's behalf. The consultation lasts four minutes, and a single sentence may switch between Hindi and English twice.
Physics, not software
Sound intensity falls with the square of distance. A phone on the desk sits perhaps thirty centimetres from the doctor and a metre and a half from a soft-spoken patient — meaning the patient's voice arrives at a small fraction of the power. No model upgrade recovers information the microphone never captured. Published comparisons of speech models on Hindi-English code-switched audio show word error rates ranging from roughly 27% to 70% — and that is before the fan and the corridor are added.
What actually works
The clinically necessary information in a short consultation flows overwhelmingly through one voice: the doctor's. The doctor repeats the history back, states the assessment aloud, and dictates the plan. A scribe designed around near-field capture of the clinician — rather than eavesdropping on the whole room — sidesteps the physics instead of fighting it. It is a different product from the quiet-room scribe, built for a different room. Most of the world's consultations happen in that second room.