We shipped voice-driven prescription in seven Indian languages a year ago. Here's what worked, what didn't, and the unexpected places it transformed the OPD.
Why we bet on voice
Indian senior consultants dictate. Junior doctors type. Software that demands keyboard speed in English filters out the consultants with the most experience and the most patients per day — the exact people you want on the system. We built voice prescription not as a futuristic add-on but as the primary input method for the senior OPD. After one year and roughly 380,000 voice-driven prescriptions across our pilot hospitals, here's what the data says.
Seven languages, not one
- Hindi — the largest single-language volume, ~42% of voice prescriptions.
- English — ~24%, mostly metro consultants and certain specialities (cardiology, oncology).
- Marathi, Tamil, Telugu, Bengali, Gujarati — the remaining ~34%, with strong regional concentration.
- Code-mixing (Hindi-English in the same sentence) is the default in 60%+ of consultations. The model has to handle 'patient ko fever hai, give paracetamol 500 mg BD' as a single utterance, not two.
What we measured
Median prescription time fell from 87 seconds (typed) to 24 seconds (voice + one-tap confirm). Senior consultants who previously dictated to an assistant and waited for transcription cut their per-patient time by 90 seconds, which translates to 8-10 more patients per OPD session. Drug-name accuracy on the first pass is 96.4% — the 3.6% errors are almost all on lookalike-soundalike Indian brand names, which the in-built drug formulary auto-corrects with confirmation.
The unexpected wins
- Junior doctor coverage — a junior can now scribe for a consultant who dictates the prescription directly, with the EMR's voice layer being the typist. The junior moves from typing to clinical observation.
- Documentation completeness — voice users include 30% more clinical context in the note than typing users, because talking is faster than typing.
- Reduced consultant frustration — the largest qualitative theme in our post-deployment interviews. 'I'm a doctor, not a typist' came up word-for-word in 11 separate interviews.
- Better discharge summaries — voice-recorded ward rounds, transcribed in real-time, became raw material for the discharge summary AI feature.
The failures
- Noisy clinics — front desks with phones ringing and patients arguing degrade transcription accuracy by 15-20%. Mitigated with a directional mic and a 'quiet mode' that re-asks ambiguous words.
- Strong regional accents in non-native languages — a Tamil consultant speaking Hindi takes a few weeks of speaker adaptation before accuracy matches a native speaker.
- Numbers and dosages — 'fifty milligrams' vs '15 mg' vs '5 mg' — we built a numeric confirmation step for any dose or duration to eliminate this class of error.
- Stop words — 'cut that', 'ignore that', 'start over' — we had to teach the model these editorial commands in every language.
What V2 does differently
V2 ships with speaker-adapted models (after 50 utterances per doctor, accuracy climbs another 2-3%), a quiet-mode that boosts the directional mic when SNR drops, and a clinical-context window that uses the patient's last three visits as bias for ambiguous drug names. We're also experimenting with a 'silent scribe' for ward rounds — the system listens to the bedside conversation, drafts the progress note, and the doctor signs at the end. Early data is encouraging but we're not shipping it broadly until we close the consent and audit story.