OPD Scribe: an AI scribe for doctor consultations in Hindi, Hinglish and other Indian languages
Engineering · October 18, 2026 · 2 min read
Record the consultation, get back a transcript split by speaker, a clinical summary with red flags, differentials with ICD-10 codes, a draft prescription and a SOAP note. Notes on the parts that were hard: Indian languages, speakers, and not paying twice.
Outpatient clinics in India are loud, fast and multilingual. A doctor sees a patient for a few minutes, speaks a mix of Hindi, English and whatever the patient speaks, and then has to write it all up. OPD Scribe was built so the writing-up part happens on its own, with the doctor checking rather than typing.
From a recording to a note
The doctor records the consultation in the browser (there is a live waveform so they know it is working) or uploads audio of up to thirty minutes. The system then:
- Separates the speakers into doctor, patient and attendant. In an Indian OPD there is almost always a family member in the room, and their words matter.
- Transcribes Hindi, Hinglish and other Indian languages.
- Writes a clinical summary with red flags, a list of differentials with ICD-10 codes, the vitals mentioned, a printable draft prescription, and a SOAP note.
Everything is a draft. The doctor edits, signs and prints.
The transcript stays glued to the audio
Each line of the transcript is linked to its moment in the recording. Click a line, the audio jumps there. When a doctor doubts a word, they listen to the three seconds that matter instead of the whole consultation. This one feature did more for trust than any accuracy number.
Two speech engines, and automatic key rotation
Indian-language speech recognition is good now but not uniform; one engine handles some accents and languages better than the other. OPD Scribe uses two engines (Sarvam AI and ElevenLabs) and rotates API keys automatically when one hits a rate limit, so a busy morning does not stall. Failed runs can be retried from the saved transcript without paying for transcription again, which matters when you are billed per minute of audio.
Stack and shape
A FastAPI and Python backend, running in Docker on a server the hospital controls. The audio and transcripts stay there.
The honest limits
It is a scribe, not a doctor. The red flags and differentials are prompts for the clinician to consider, labelled as such, and nothing goes on a prescription without a human signing it. If you run a clinic or a hospital department and spend your evenings on documentation, AI assistants like this one are built to your workflow, not the other way round.
Tags: AI, speech to text, healthcare, FastAPI, Hindi