Consented mic capture → transcript → LLM SOAP extraction → physician sign-off
The visit starts like any other — consent shown, mic listening quietly.
Patients see a clear indicator before any audio is recorded.
Consent is logged alongside the encounter for audit purposes.
Ambient audio includes clinical talk plus ordinary small talk and noise.
Nothing is filtered yet — that happens in later stages.
Passive capture removes the need for manual typing during the visit.
A speech model turns the recorded conversation into timestamped text.
Audio frames convert into text tokens as the conversation unfolds.
Accuracy depends heavily on background noise and speaker overlap.
Noisy rooms and cross-talk lower transcription accuracy noticeably.
Speaker diarization helps separate doctor and patient utterances.
Cleaner audio input directly improves downstream note quality.
A language model reads the transcript and pulls out relevant clinical facts.
Small talk about weather or parking gets dropped automatically.
Symptoms, exam findings, and plans get routed to their sections.
Each clinically relevant sentence is classified into S, O, A, or P.
The model reasons over context, not just keyword matching.
Structuring turns free-flowing conversation into a usable clinical record.
Subjective, Objective, Assessment, and Plan fill with extracted content.
SOAP structure mirrors how clinicians already organize their notes.
Each section fills incrementally as the LLM finishes drafting.
The generated note is a starting point, not a finished record.
It still requires physician review before it becomes official.
A structured draft saves far more time than typing from scratch.
The doctor reviews the draft, corrects details, and signs off.
The physician remains accountable for everything in the final note.
Quick edits catch any misheard or misclassified details.
Less typing during and after visits means more time with patients.
Ambient scribing aims to reduce clinician documentation burden.
Signed notes close the loop from conversation to chart in minutes.