omi-medical-1: The Most Accurate Medical Speech-to-Text We've Measured

General speech models guess when they mishear. In an ordinary transcript that produces a typo; in a clinical note it can invent a drug that was never said. omi-medical-1 is our flagship model built for the other outcome: the most accurate medical speech-to-text we have measured across 28 systems — every major cloud API, the dedicated medical variants, and the leading open models — with zero fabricated drug names on every test set we have ever run.

On OmiMedSTT-Bench, our sealed clinical benchmark, it ranks #1 of 28 on Medical WER — the error rate on the words that carry clinical meaning — while matching or beating the best general models on overall word error rate. It runs in our EU cloud (Azure Sweden Central) today, and can be self-hosted in your own environment. The model we benchmark is the model you call.

This page is the model write-up: what it is, how it scores, and how to run it. The complete, current evaluation — leaderboard, dataset card, methodology — lives on our standing medical STT benchmark.

How it was measured

On OmiMedSTT-Bench, our sealed clinical test set: 1,513 audio segments, 7.2 hours of English medical speech across five scenario types — from short clinical utterances to complete 14-minute consultations — scored by the same open pipeline for every system and ranked medical-first: Medical WER, then drug-name errors, then overall WER. Cloud systems are called through their public production APIs; open models run self-hosted at full precision. Zero file-level overlap between our training audio and the test set. Full dataset card, scoring convention and the complete 30-system leaderboard: the benchmark page.

Results — July 2026 board

Current standing on the benchmark. omi-medical-1 is the top of the board on medical-term accuracy, and its overall word error rate sits inside the noise band of the best general model — the lead we claim is on the clinical words and on safety, not a wide raw-WER margin.

One system — Azure's MAI-Transcribe-1.5 — edges our raw WER by two-hundredths of a point (5.97% vs 5.99%, a difference of roughly a dozen words across the set, i.e. noise). It is not medically tuned: its Medical WER is 1.43% against our 0.94%, and it fabricates drug names where we fabricate none. That gap is the whole point of a medical model.

• #1 of 28 on Medical WER (0.94%) — ahead of every cloud API, every dedicated medical variant, and every open model on the board.
• Zero fabricated drug names across all 442 drug occurrences, and zero runaway files — no looping or invented passages requiring exclusion. Only omi-medical-1 and ElevenLabs needed zero exclusions.
• Matches-or-beats the best general models on overall WER (5.99%, within noise of the top raw-WER system), while leading decisively on the clinical terms.

Full leaderboard and methodology →

Where the gap widens: real consultations

Short utterances flatter every system; long, messy, real consultations separate them. On a separate long-form companion set — 22 complete real consultations (4.1 hours), every provider on identical audio, references human-audited — omi-medical-1 leads the next best system by 1.4 percentage points of word accuracy, a bootstrap-certain margin.

The pattern is consistent with the design of the model: the systems that lead on clean, short audio degrade faster on the long, overlapping, accented speech of a genuine clinic. omi-medical-1 is built to hold up there.

Medical branding is not medical accuracy

Several vendors ship a "medical" speech-to-text variant. On our benchmark, being labelled medical did not predict medical accuracy: the dedicated medical products generally did not lead on clinical terms — the strongest general models did, and omi-medical-1 leads both. Amazon Transcribe Medical, a Google medical-conversation model and Speechmatics' medical mode all scored worse on medical vocabulary than the best general engines on identical audio. It is a reminder that "medical" is a badge until a benchmark makes it a number — which is why we publish the scorer and the full board.

How to run it

omi-medical-1 is available today as a batch API in our EU cloud. The endpoint is OpenAI-compatible, so most existing speech-to-text integrations work by changing the base URL and the model name:

$ curl https://api.omi.health/v1/audio/transcriptions \
    -H "Authorization: Bearer $OMI_API_KEY" \
    -F model="omi-medical-1" \
    -F [email protected]

Three ways to deploy, one model:

• Batch API — live. EU cloud (Azure Sweden Central), OpenAI-compatible endpoint. Long recordings transcribe via the async route at roughly 4× realtime.
• Realtime — preview. Streaming partials for live, in-consultation transcription, with the corrected record produced at the end.
• Custom vocabulary — preview. Your formulary, device names and local terms recognized in about a minute, with a live safety monitor.
• Self-host / your cloud. Run omi-medical-1 inside your own VPC with SSO for data-residency or air-gapped requirements.

See the models and quickstart →

Data and privacy

omi-medical-1 runs in the EU (Azure Sweden Central). We never train on customer data, there is no shared data plane between customers, and audio submitted to the playground is deleted after transcription. Self-hosting is available where data must not leave your environment at all. Sub-processors and the full data-handling terms are published in our legal pages.

Limitations

• Raw WER is a tie at the top. The best general model matches our overall word error rate within noise; our advantage is on medical terms and on safety (zero fabricated drugs), not a large raw-WER gap. We say "matches-or-beats on WER, ahead on medical accuracy" for exactly this reason.
• The benchmark is our own sealed set. We built and hold OmiMedSTT-Bench. The mitigations are an open scorer, published methodology, a human-audit protocol, and full-set re-runs under NDA. A fully provider-disjoint third-party holdout is planned; until it is published we do not make an unqualified "beats system X" claim — we report the numbers and let the method be checked.
• English first. Dutch, German, French and Spanish are in development, matching the markets our Speech-to-Text models serve.
• omi-medical-1 is speech-to-text only. It is not a diagnostic, triage, prescribing or clinical decision model, and it is not clinically validated. Transcripts must be reviewed before any clinical use.

What's next

Three things are in active development: the realtime streaming model moving from preview to general availability; custom vocabulary for per-customer terminology; and multilingual support — Dutch, German, French and Spanish first. On the evaluation side, a provider-disjoint third-party holdout is the next step in making the #1 claim independently checkable.

Where to find it

• Models & quickstart: omi.health/speech
• Standing benchmark: Medical Speech-to-Text Benchmark (28 systems)
• Open scorer & evaluation code: github.com/Omi-Health/medical-STT-eval
• API key / self-host: [email protected]
• Pairs with: omi-medical-edge-1 — the open, on-device model for when audio cannot leave the machine.

Cite this model

APA — Omi Health. (2026). omi-medical-1: The Most Accurate Medical Speech-to-Text We've Measured. https://omi.health/research/omi-medical-1

@misc{omi_medical_1_2026,
  title   = {omi-medical-1: The Most Accurate Medical Speech-to-Text We've Measured},
  author  = {{Omi Health}},
  year    = {2026},
  url     = {https://omi.health/research/omi-medical-1},
  note    = {Flagship medical ASR, #1 of 28 on OmiMedSTT-Bench v5 (sealed), EU cloud / self-host}
}

Related research

• Medical Speech-to-Text Benchmark — 28 systems ranked by Medical WER on a sealed test set
• omi-medical-edge-1 — the open, on-device 0.6B medical speech-to-text model
• Clinical SOAP Note Safety Evaluation — 6 models, 300 dialogues, safety-first scoring