The dialect the model cannot hear

Ask a frontier model a medical question in clean, written Bangla and it answers well. This is the demo that convinces people the language problem is solved. It is not, because the language the demo uses is not the language Glyph's patients speak.

A patient does not describe her chest pain in textbook Bangla. She describes it in Sylheti, or Chittagonian, or the Noakhali her village speaks, in half-sentences, with the word for a symptom that only exists in her district, mediated by a son who is translating as he goes. The distance between that and standard written Bangla is not an accent. It is a different problem, and frontier models trained mostly on English and a thin slice of formal Bangla fall off it completely.

The failure is quiet, which makes it worse. The model does not announce that it misheard. It produces a confident, fluent answer to the question it thought it heard, and a confident wrong answer in a triage is more dangerous than no answer at all.

This is why dialect is not a feature to add later. It is the floor the whole thing stands on. A clinical model for Bangladesh that cannot follow a woman describing her symptoms in the speech she actually uses is not a clinical model for Bangladesh. It is a clinical model for the small, urban, formally literate slice of it, which is the slice that was already best served.

Closing that gap is most of what building KhaM-Med actually is. Not new medicine, the base already knows board-exam medicine. The work is teaching it to hear the country: the dialects, the local words for pain and fever and dizziness, the way an attendant speaks for a patient, the way a real symptom arrives wrapped in worry and idiom. A model can only learn that from the speech itself, and that speech lives here, in these districts, in these patients' own words. It has to be learned where it is spoken. That is the deepest reason the model has to be built where the patients are.