By the end of a six-minute consultation, the doctor has usually said everything the record needs: the complaint, the relevant history, the diagnosis, the medicines and when to come back. Then it has to be typed. An AI medical scribe is meant to remove that second pass, and whether it does depends on what happens between the microphone and the saved record.
Thravi's voice scribe is now part of the consultation screen in the EMR. This post walks through what it does at each step, in the order a doctor meets it, so you can judge it against your own OPD. For the wider background on why Indian consulting rooms are hard for speech technology, see our earlier piece on AI voice scribes in Indian clinics.
Record the consultation, or upload audio
From the visit's EMR screen, the doctor can record the consultation as it happens or upload audio recorded elsewhere. Either way the path is the same: the audio is transcribed, a draft of the structured fields is prepared from the transcript, and the doctor reviews that draft before any of it reaches the visit.
Recording offers a choice between two modes. An uploaded file always takes the second, because a finished file has nothing to show live.
Live mode: the transcript appears as you speak
In the live mode, words appear on screen while the consultation is still going on. That is also the plainest confirmation that the microphone is hearing the room. By the time the doctor presses stop, the transcript already exists and only the drafting is left to wait for.
The live transcript does not label who is speaking. It does break at the natural pauses where one person stops and the other starts, and the drafting step works out who the clinician is from what was said.
Accurate mode: speakers are separated
In the accurate mode, the recording is held until the doctor stops and is then transcribed in one piece, with the speakers separated in the transcript. There is no live text to watch. The choice is between speed with live feedback, and speaker attribution with a transcription pass built for clinical speech.
Choose the language first: English, Hindi or Telugu
Before recording, the doctor picks the language of the consultation: English, Hindi or Telugu. One is always chosen. Automatic detection is deliberately not offered, so the system is never left to guess which language it is hearing.
The choice describes what the audio is. It is not a translation setting. A consultation that mixes languages, say Telugu sentences carrying English drug names, is transcribed as the declared language, so pick the one most of the conversation will be in. The draft that comes back is in English whichever language was spoken.
Pause when the conversation is not for the record
Both modes can be paused: for a phone call, an examination, or a few sentences the patient would rather keep out of the notes. Paused time is not recorded and does not count towards usage. Whatever the microphone picks up after the pause is pressed is left out, including when the doctor stops the recording straight from a pause.
A pause left running for fifteen minutes ends the recording on its own, and the notes are drafted from what was captured up to then. A countdown is shown while paused, so this never comes as a surprise.
The transcript stays faithful, corrections included
Speech tools often tidy what they hear by removing false starts and filtering quiet sounds as background. In a consultation, both habits are risky. When a doctor says “500, no, 650 milligrams”, the correction is the half that matters. A softly spoken patient is exactly the person who must not be filtered out.
So the transcription is kept faithful: false starts stay in, and the bar for what counts as speech is not raised. Deciding what belongs in the record is left to the drafting step and, after it, to the doctor.
The draft arrives as suggestions you tick and copy
The draft is organised into the same sections as the EMR form: vitals, medical history, symptoms, diagnosis, medications, investigations, advice for the printed prescription, the doctor's own notes and the follow-up interval. Each suggestion has a checkbox. The doctor ticks what is right, or selects all, and copies the selection into the form. It is the same interaction used to copy sections from a patient's previous visit.
Nothing is written to the visit until the doctor copies it and saves the form. A suggestion that is wrong is simply left unticked.
Copying adds to what is already there. If the doctor has typed two symptoms, the copied ones are appended and nothing is replaced. A vital that has already been entered is never overwritten, and a follow-up interval the doctor has set is kept. A temperature spoken in Fahrenheit is placed in the Celsius row after conversion; a reading that makes sense in neither scale is left blank for the doctor.
An empty section is treated as a correct answer. On a short consultation the draft is short, because the scribe is instructed not to fill a field with something that was never said.
Each suggestion is matched to the coded masters
This is the step that separates a voice-to-EMR tool from a dictation tool. “Sore throat”, “pain in throat” and “throat pain” are one SNOMED CT concept with one preferred term. If each phrasing were saved as spoken, the same finding would be spelled three ways across three visits, and the record would stop being searchable by its own vocabulary.
Before a suggestion can be copied, it is matched against the reference data the EMR already uses, in this order:
By code. If the draft proposes a code, that code is checked against the reference tables. Where it exists, the entry is identified outright.
By name. Otherwise the wording is looked up through the same search that powers the symptom, diagnosis, investigation and medicine fields. A copied “sore throat” therefore lands as the same coded entry the doctor would have got by typing it and choosing from the dropdown.
Not at all. If neither works, the wording is copied as plain text with no code, and the doctor can pick a coded match from the field's own search.
The name match is much stricter than the dropdown a doctor types into. In a dropdown, a loose match costs nothing because nobody clicks it. Here nobody is choosing, so a candidate is accepted only when its wording is effectively the same. “Chest pain” is not matched to “chest wall pain”. A vague phrase stays as plain text instead of becoming a guess.
A diagnosis carries both its SNOMED CT concept and an ICD-10 code. Where a proposed code cannot be verified, it is shown with a warning marker and the entry copies as plain text. The review also shows the standard's own term beside the drafted wording. Verification confirms that a code exists; whether it is the right one for this patient remains the doctor's judgement. Our guide to LOINC, ICD-10 and SNOMED CT in clinical records explains what each standard contributes.
Medicines are handled a little differently
The medications section holds what the doctor instructs in this consultation. A medicine the patient mentions already taking is not turned into a prescription line. The condition it treats is kept as medical history, and the drug is left out of the draft. The doctor heard it in the room; the draft is not a substitute for that.
A spoken medicine is matched to the drug master, which lists specific products. The strength that was spoken is used to choose among them, so “650” does not land on a 500 mg entry. Whenever the matched entry differs from what was said, the review prints its full name. That line is worth reading before copying.
What is kept afterwards
The draft is stored with the visit, so the suggestions survive a page reload and there is a trail of what was proposed and what the doctor accepted. The audio and the transcript are not kept. They are used to prepare the draft and then discarded, because a word-for-word record of everything said in the room includes much that never belonged in a medical record.
One practical consequence: a recording lives in the browser tab until it is stopped. If the tab is closed mid-consultation, that recording is lost.
Where the doctor stays in charge
The scribe does not diagnose, prescribe or sign. It drafts from what was said, matches the draft to codes that can be verified, and waits. Usage is metered against the hospital's prepaid credit balance, and the panel shows how many minutes that balance covers.
The follow-up interval is a good example of the boundary. “Review in two weeks” is drafted as fourteen days, and copying it fills the follow-up box. Whether the patient is sent a reminder is still the doctor's own tick, as described in our guide to WhatsApp and SMS follow-up reminders for clinics.
If you are comparing AI scribes, put the same questions to each one. Where does the output land? What happens to a term the system cannot match? What is stored after the visit? Can it be paused? To see these steps on your own consultations, explore Thravi's EMR software for Indian practices and ask for the scribe in your demo.
