Speech during a consultation: preserve meaning, not just words

A consultation is not a dictated report. People interrupt themselves, change a number, return to an earlier event and speak over one another. A patient may quote another person, while a specialist may offer a hypothesis as a question. Converting this exchange into a fluent paragraph can erase distinctions that are essential to understanding the encounter.

The central research problem is incremental interpretation. An early fragment may be incomplete: “I take fifty…” can be corrected or qualified in the next sentence. A near-real-time transcript must distinguish provisional text from a stable segment. The interface should allow the specialist to revisit the original utterance before treating a proposed fact as confirmed.

Speaker separation is another independent task. Identifying two acoustic voices does not by itself establish which one belongs to the patient or the clinician. Role assignment needs confirmation. Otherwise a clinician’s question can be recorded as a patient’s symptom, or an accompanying person’s account can be attributed to the patient without qualification.

An evaluation should include spontaneous pauses, overlapping speech, self-corrections, medication names and statements containing negation. Overall word error rate is useful, but it does not express the full significance of an error. A single omitted “not” can change the interpretation of a sentence more than several mistakes in ordinary connecting words.

We distinguish a prompt for attention from a clinical finding. An assistant may highlight an ambiguous dose, a contradiction between two statements or a phrase that requires clarification. Such a signal should identify its source and remain open to dismissal or correction. Vocal style or word choice alone should not be presented as proof of a psychiatric condition.

The intended benefit is continuity of attention: the specialist can return to a precise part of the conversation, and the patient can correct a misunderstanding before it enters the record. The research challenge is to achieve this without creating a distracting stream of unreliable alerts. Latency, correction burden and faithful attribution need to be assessed together.