The consultation
is not in English.
Most clinical AI is built for one language spoken cleanly. Indian consultations switch language mid-sentence, in a noisy room, in under five minutes. We trained for that case rather than adapting to it.
Why translated models fail here.
The consultation is not in one language
A patient describes symptoms in Marathi, the doctor names the condition in English, the dosage instruction comes back in a mix of both. Systems that expect one language per session mis-handle the sentence where the clinical content actually sits.
Translated models inherit the wrong priors
An English model adapted to Hindi carries English phonetics and English clinical phrasing. It transcribes what it expects rather than what was said, and the errors cluster exactly on drug names and dosages.
The room is loud
Indian OPD rooms have queues outside, families inside, and consultations measured in minutes. Capture quality is a hardware problem before it is a model problem, which is why Guardian is a device rather than an app.
Language coverage.
Trained natively, not transliterated from English. Code-switched speech is handled as the normal case rather than as an error condition.
- Hindi
- Bengali
- Marathi
- Telugu
- Kannada
- Tamil
- Malayalam
- Gujarati
- Punjabi
- Odia
- Assamese
- Urdu
- Sindhi
- Hinglish
Coverage is not uniform. Languages with more available clinical speech perform better than those without, and dialect variation within a language is a live research problem for us rather than a solved one — it is listed among the open problems on our technology page. We would rather you test the languages your hospital actually runs in than trust a list.
Test it in your languages.
The useful evaluation is your own consultations, in your own rooms, in the languages your patients speak.