- Speech and drafting use India-hosted models from Sarvam AI; verification is 1Heath's own rules engine.
- 1Heath has trained no model and used no real patient data.
- All results come from synthetic consultations, so they compare designs; they do not predict real-world accuracy.
- Next: consented real consultations, with every metric published.
The Models and What Each One Does
| Job | Model | Why this one |
|---|---|---|
| Speech to text | Sarvam speech recognition, in two modes: translate (to English) and code-mixed (original language) | Supports English and 22 Indian languages, including several languages mixed in one sentence[1]; India-hosted. In an earlier comparison it kept dosing phrases far better than the free models we tried |
| Drafting the note | Sarvam-105B, an India-trained large language model | Works in English and 10 Indian languages[1]; it reads both the English translation and the original-language transcript, so speech in other Indian languages still reaches it in English. India-hosted |
| Verification | 1Heath's own rules engine and name matcher (no machine learning) | Deterministic, explainable and testable; it can say exactly why a line is flagged |
Speech recognition is treated as a replaceable part, not the core of the product. It sits behind one gateway so it can be swapped if a better or cheaper model appears.
What Is Known About These Models
| Model | Published facts | What it means for 1Heath |
|---|---|---|
| Saaras (speech) | Supports 22 Indian languages plus English, detects the language automatically, and handles code-mixed audio; a code-mixed mode keeps English words in English and Indian words in their own script[2] | Telugu, Hindi and English are tested today; the other languages are available but untested in 1Heath |
| Sarvam-105B (drafting) | About 106 billion parameters with about 10 billion active per word (a mixture-of-experts design), trained from scratch in India, open weights under the Apache 2.0 licence[3] | Can be self-hosted later, which helps with data control |
| Sarvam-105B, independent test | An independent benchmark reported a high hallucination rate on a knowledge test: the model attempted far more answers than it declined[3] | This is exactly why nothing the model writes reaches the record without verification |
What Data Exists Today
| Data | What it is | Used for |
|---|---|---|
| Synthetic consultations | 18 scripted consultations in 13 specialties, in Telugu, Hindi and English, voiced by text-to-speech, with simulated rooms: fan noise, corridor, phone rings, relatives interrupting, and noisy ward versions at 10, 5 and 0–3 dB | Speech and end-to-end tests |
| Native-language set | 30 sentences in natural word order, six each in Telugu, Hindi, Tamil, Kannada and Malayalam | Checking languages beyond the main three |
| Frozen clinical test set | 131 cases in eight languages covering allopathy, paediatrics, gynaecology, Ayurveda, homeopathy, look-alike names, banned combinations and "current versus new" medicines | Acceptance testing; frozen by checksum so it cannot drift |
| Drug catalogue | Demo data: a public dataset of Indian medicine products, 2,111 Jan Aushadhi generics, 1,027 Ayurvedic formulations, 1,117 homeopathic remedies, 711 banned combinations, about 500 common brands with spoken forms in English, Hindi and Telugu and spellings in five Indian scripts | Name matching and safety checks |
| Real consultations | None yet | Next step, only with written consent and the safeguards on the privacy page |
The public medicine dataset in the demo has uncertain provenance and is used only for the demo. Before any pilot it is replaced by the government's National Drug Registry, which lists more than 1.23 lakh branded and 10,000 generic medicines[4].
Training: What We Do and Do Not Do
- 1
No training on patient data without consent. Any future 1Heath model will learn only from consultations recorded with written consent, reviewed by a doctor and de-identified.
- 2
No training on other AI models' output. Text produced by any AI model is used for testing only, never as training data, because a model trained on another model's mistakes repeats them confidently.
- 3
Any new model is checked by the same rules. A model can make the draft better; it never replaces verification.
- 4
Speech providers must not train on our audio. Retention set to zero and training switched off, confirmed in writing, before any real patient audio is sent.
How We Test
What We Measure, in Order of Importance
- Dosing phrases kept exactly (strength, frequency, duration)
- Critical errors, including negation flips ("not Dolo" heard as "Dolo")
- Medicine names correct, and zero invented medicines
- Numbers and units correct
- Blocking lines per consultation (target: three or fewer)
- General word error rate, last, because a transcript can read well and still get the dose wrong
Held-Out Discipline
Once a test set has been used to tune the system, it no longer counts as evidence. Every claim on this site comes from consultations the system had not seen while being tuned.
What We Measured
Model Plus Rules Beats Either Alone
Brand Names Are Where Speech Fails
Translation Changes Numbers
Translating speech to English helps the wording but corrupts doses. "Azithral 500 once a day" became "once a day for 500 days"; "Pan 40, one" became "Pan 41"; "one in the morning" became "1 AM". The fix: audio chunks that contain a medicine name, a number or a dosing word are also transcribed in the original language, and the model reads both. On recorded consultations this second pass ran on about 62% of audio and raised speech cost from about ₹0.50 to about ₹0.81 a minute[5].
What Published Studies Say About AI-Written Notes
A rate of 1.47% sounds small, but a note has dozens of sentences, so many notes will contain at least one invented fact. That study used English primary-care conversations; Indian code-mixed consultations are harder. The lesson is the same as ours: an AI draft needs a check that does not depend on the AI.
How Close Is This to the Real World?
Not close enough to make accuracy claims. The synthetic voices come from the same vendor whose speech recognition we use, which flatters the results. Real doctors speak faster, patients interrupt, and clinics are noisy.
This is a problem for the whole field in India. A September 2026 study found that most clinical conversation datasets are simulated or Western, that interruptions were about 12% of turns in Indian simulated data against roughly none in Western datasets, and that Indian scribe vendors each evaluate privately with no shared benchmark[7]. Even in the United States, a 2025 randomised trial found that ambient scribes made clinically significant mistakes occasionally and required ongoing vigilance[8].
Next: record real consultations with written consent from doctors who volunteer, measure every metric above on them, and publish the method and results here, including the failures.
Sources
- Sarvam AI. Models overview (Saaras v4 speech: 22 Indian languages plus English with a code-mix mode; Sarvam-105B: 10 Indian languages plus English), API documentation, opened 6 November 2026. docs.sarvam.ai
- Sarvam AI documentation: Saaras speech-to-text models and supported languages. docs.sarvam.ai
- Artificial Analysis. Sarvam 105B and Sarvam 30B: everything you need to know, 2 April 2026. artificialanalysis.ai
- Digital Health News. India launches National Drug Registry under ABDM, July 2026. digitalhealthnews.com
- 1Heath internal benchmarks, October to November 2026. Synthetic consultations; method described on the Models and data page. pawankalyan.dev
- A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation. medRxiv preprint, 2024. medrxiv.org
- Jaiswal SD, Krithi S, Makani A, Datta S, Sitaram S, Jain M. Evaluating ambient clinical scribes in India: the need for multilingual real-world clinical conversation data. arXiv:2609.17355, September 2026. arxiv.org
- Ambient AI scribes in clinical practice: a randomized trial. NEJM AI, December 2025 (238 physicians, 14 specialties). pubmed.ncbi.nlm.nih.gov