AI Medical Scribe India: Testing Indian Languages

On 11 November 2026, the National Health Authority (NHA) selected EkaCare, a Bengaluru health-AI company, as a technology partner for an ambient voice scribe for public health facilities. Coverage of the announcement appeared the same day. Judging an AI medical scribe India-wide is different from judging one built for English-speaking clinics. Doctors switch languages mid-sentence, family members often speak on the patient’s behalf, and consulting rooms are noisy. This post looks at what recent research says about those conditions and how a clinic can test a tool before trusting it.

What the NHA Announcement Covers

Analytics India Magazine and Physical AI News both reported the selection on 11 November 2026. EkaCare was chosen through an NHA Expression of Interest process that included a demonstration before an evaluation committee. According to those reports, the Voice Scribe listens to a consultation and drafts structured notes covering complaints, history, medicines, vitals, lab findings, diagnosis and plan. A doctor must review and approve each note before it is saved.

EkaCare says the tool supports more than 15 Indian languages, including conversations that switch between languages. Its speech model has been benchmarked on Indian English, Hindi, Marathi, Kannada and Telugu. The company also says its model had the lowest error rate in a 6,550-consultation test. That is EkaCare’s own claim, and the coverage does not verify it independently. Neither report gives a rollout timeline, the number of facilities involved or pricing. Coverage also notes that accuracy in noisy clinics and patient consent will decide how well the tool is adopted.

Why Indian Consultations Are Hard for Speech AI

A 2026 paper on arXiv, “Evaluating Ambient Clinical Scribes in India”, argues that the country still lacks a public, real-world benchmark for these tools. The arXiv listing for version 1 is dated 15 September 2026 (arXiv:2609.17355). Its authors describe Indian consultations as brief, often involving family members who speak for the patient, frequently code-mixed and multilingual, and recorded in noisy settings. They note that Indian outpatient clinicians often see more than 200 patients a day, with consultations averaging about two minutes, compared with roughly 18 minutes in the Global North.

The same paper found that widely used clinical dialogue datasets, where the clinician does most of the talking, do not reflect Indian interaction patterns. It also found that each organisation it interviewed built its own proprietary evaluation pipeline, so results across vendors could not be compared. Some interviewees said that LLM-based judging missed dosages and drug names, which are the errors that matter most in a prescription.

A separate study shows why speaker and accent matter. ASR Under the Stethoscope, posted to arXiv in late 2025, audited eight speech recognition systems on 162 real psychiatric interview recordings in Kannada, Hindi and Indian English. The best-performing system had a word error rate of about 14% in English, about 19% in Hindi and about 35% in Kannada. In the detailed analysis of one system, patients had higher error rates than doctors, and patients with less than a graduate education had higher error rates than graduates. The authors also say the dataset is small, comes from a single hospital, and includes per-attribute results from only one system. Their recommendation is still clear: fairness testing across speaker role, gender, education and region should be a standard step before clinical deployment.

Taken together, these papers suggest a practical rule. A scribe that looks accurate on clean, scripted audio can still fail on the conversations a clinic has every day.

Testing an AI Medical Scribe for India: A Checklist

Vendor demos usually use clear audio and scripted speech. A useful evaluation uses your own clinic’s conversations. Start with these checks:

• Record real, consented consultations in every language your doctors and patients use, including code-mixed speech and family members who answer for the patient.
• Score omissions and invented content, not only word error rate. A transcript can have a low error rate and still drop a dictated dose.
• Check drug names, doses and numbers separately. These are the lines where a small error does the most harm.
• Track clinician edits per note, broken down by language and by speaker type. If patient turns are edited far more often than doctor turns, the tool is not ready for those patients.
• Trace every drafted line back to the transcript. A line the transcript does not support should be flagged for review, not accepted by default.
• Ask where audio and text are processed, who can access them, and what happens to them afterwards.

The researchers’ broader call is for shared, public benchmarks. Until those exist, a clinic can at least use the same test set across every vendor it considers.

How InstaPraxis Handles Scribe Output

InstaPraxis is designed so that the doctor checks the note, not the other way round. Here is what the platform does today:

• Live captions run in the browser using an on-device speech model, so that audio stays on the device.
• Server-side recognition uses IndicConformer-600M with IndicTrans2-200M translation to cover 22 Indic languages. English is handled by Whisper-medium. The language is detected for each utterance, and the output is native-script text with an English translation.
• Hindi and Marathi clause splitting was added on 28 October 2026. Unpunctuated speech is split into clauses before translation, so a dictated medicine is less likely to be dropped.
• Note evidence matches each drafted line to the transcript. An examination finding the transcript does not support must be confirmed by the doctor or removed before the note is finalised.
• Honest scope. Publicly, we state only Hindi, English and Hinglish as tested. The other Indic languages are supported by the engine, but a clinic should test them on its own consultations before relying on them.
• Data handling. Patient records are stored on InstaDataHelp servers. Text sent to external generative AI services is masked first.
• ABDM. InstaPraxis does not yet connect to ABDM. That integration is on our roadmap, and we will not describe it as live until it is.

If your clinic is comparing options, run the checklist above on your own recordings, whichever tool you choose. The results will tell you more than any benchmark quoted in a brochure.

Watch: Check If a Prescription QR Code Is Genuine in India: What to Verify

Watch on YouTube · More tutorials on our channel

Frequently Asked Questions

Is word error rate enough to judge a medical scribe?

No. Word error rate measures how many words are wrong, but it does not show whether a dropped medicine, an invented symptom or a wrong dose reached the note. Clinical teams should also measure omissions, hallucinations, speaker attribution and the accuracy of drug names and numbers.

Does an AI scribe write the final note without the doctor?

Not in InstaPraxis. The doctor reviews and finalises the note. Examination findings that the transcript does not support must be confirmed by the doctor or removed before the note can be finalised.

Which Indian languages does the InstaPraxis scribe support?

The speech engine covers 22 Indic languages plus English, with automatic detection for each utterance. Publicly, we state only Hindi, English and Hinglish as tested, so please test any other language on your own consultations first.

Is the NHA voice scribe connected to ABDM?

The NHA’s Voice Scribe is described in coverage as linked to the patient’s Ayushman Bharat Health Account. InstaPraxis does not yet have ABDM integration. It is on our roadmap.

Sources

• Analytics India Magazine: NHA Chooses EkaCare for AI Voice Scribe Solution, 11 November 2026
• Physical AI News: NHA Selects EkaCare for AI Voice Scribe in Public Health Facilities, 11 November 2026
• Jaiswal et al., Evaluating Ambient Clinical Scribes in India: The Need for Multilingual Real-World Clinical Conversation Data (arXiv:2609.17355), September 2026
• Kumar et al., ASR Under the Stethoscope: Evaluating Biases in Clinical Speech Recognition across Indian Languages (arXiv:2512.10967), November 2025

Ready to see how a scribe handles your clinic’s languages? Explore InstaPraxis, or read more of our healthtech guides on the InstaDataHelp blog.