Omi Medical v1.1: accuracy, speed and safeguards
General speech models can change clinically important terms when they mishear. In an ordinary transcript that may be a typo; in a clinical record it can alter a medication or dosage. Omi Medical v1.1 is our flagship model built around those errors. On the October 2026 batch board it has the lowest observed Medical WER among 17 tested models, with one drug-term error across 442 references.
On OmiMedSTT-Bench, our medical speech benchmark, it ranks #1 of 20 batch models on Medical WER — the error rate on words that carry clinical meaning — while remaining close to the best tested overall WER. It is available today through the EU or US API and for supported private deployment. The model we benchmark is the model you call.
How it was measured
On OmiMedSTT-Bench, our medical speech development benchmark: 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 by Medical WER. Hosted systems are called through their public production APIs; open-model results are labelled separately. Zero file-level overlap between our training audio and the test set. Full dataset card, scoring convention and the complete batch and realtime leaderboards: the benchmark page.
Results — October 2026 batch board
Omi Medical v1.1 records the lowest observed medical-term error rate. Microsoft MAI-Transcribe-2 records the lowest overall word error rate. We report both measures separately.
| # | System | WER | M-WER | Drug-term errors |
|---|---|---|---|---|
| 1 | Omi Medical v1.1 (ours · batch API) | 5.77% | 0.87% | 1 / 442 |
| 2 | ElevenLabs Scribe v2 Medical | 5.87% | 1.08% | 2 / 442 |
| 3 | Google Chirp 3 | 6.31% | 1.11% | 5 / 442 |
| 4 | ElevenLabs Scribe v2 | 6.49% | 1.15% | 3 / 442 |
| 5 | Microsoft MAI-Transcribe-2 | 5.64% | 1.18% | 9 / 442 |
Microsoft MAI-Transcribe-2 records the lowest raw WER in this draw (5.64% versus Omi's 5.77%). Omi records the lower Medical WER (0.87% versus 1.18%) and fewer drug-term errors (1 versus 9). That difference is why the benchmark reports general and clinical terms separately.
- #1 of 20 batch models on Medical WER (0.87%) on the benchmark board, measured in the US in October 2026.
- One drug-term error across 442 references, compared with two for the next Medical WER row.
- 5.77% overall WER, close to the lowest observed batch WER while leading on the clinical-term measure.
Full leaderboard and methodology →
What changed in v1.1
Realtime medical-term errors fell from 38 to 30 on the same clinical board, a 21% reduction. Drug-name errors fell from 5 to 3. Overall realtime word error fell from 6.55% to 6.17%, and dose / unit matches rose from 82 to 90 out of 94. Batch medical-term and drug-name results are unchanged.
New guards help prevent stray text from digital silence and tiny fragments at turn boundaries. Improved speech detection helps recover speech masked by electrical hum while preserving the original audio for the recognizer. Mumbled or heavily masked speech can still be misheard.
The API model ID stays omi-medical-1. See the release comparison and test conditions →
Accuracy and speed, batch and live
| Omi Medical v1.1 | Medical WER ↓ | Dose / unit matches ↑ | Median / p95 speed ↓ |
|---|---|---|---|
| Batch | 0.87% | 96.81% · 91/94 | 131 / 222 ms response |
| Realtime | 1.04% | 95.74% · 90/94 | 86 / 169 ms final latency |
Omi measured in the US (Ohio), October 2026: batch speed on short clips under 30 seconds; realtime speed is the final transcript latency after the end of speech, 200 short clips; every API timed from the US. Realtime accuracy uses confirmed segments, not drafts. Full realtime comparison →
Earlier study: full consultations
A separate, earlier study covered 22 complete consultations (4.1 hours) with human-audited references. Omi recorded 94.0% word accuracy, compared with 93.2% for Gemini 3.1 Pro and 92.6% for ElevenLabs Scribe v2. These results use a different dataset and are not part of the current 1,513-recording board.
| System | Word accuracy | vs Omi Medical v1 (earlier release) |
|---|---|---|
| Omi Medical v1 (earlier release) | 94.0% | — |
| Gemini 3.1 Pro | 93.2% | −0.8 pt |
| ElevenLabs Scribe v2 | 92.6% | −1.4 pt |
Long consultations remain an important evaluation case alongside short clips and live speech.
Medical branding is not medical accuracy
Model names alone do not predict clinical accuracy. ElevenLabs Scribe v2 Medical improves on its base model in this evaluation, while Deepgram Nova-3 Pharma outperforms the tested Nova-3 Medical configuration. We publish the exact model names and separate word, medical-term, drug and dosage scores so the differences are visible.
How to run it
Omi Medical v1.1 is available today as a managed cloud API. 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 file=@consultation.wav
Three ways to deploy, one model:
- Batch API — live. OpenAI-compatible endpoint. Handles everything from short clips to full-length consultations.
- Realtime — live. Streaming medical transcription through the public API, with confirmed-segment accuracy and response speed on the benchmark.
- Custom vocabulary — live. Pass a focused encounter list of medicines, devices, names and local terms, with an automatic safety fallback.
- More languages — live. Spanish, Portuguese, French, German, Dutch, Arabic and Hindi, including regional BCP-47 variants.
- Speaker decoration — live. Speaker labels and word timestamps attach structure without changing the sealed transcript text.
- Self-host / your cloud. Run Omi Medical v1.1 inside your own VPC with SSO for data-residency or air-gapped requirements.
See the models and quickstart → Read the API guide →
Data and privacy
Omi Medical v1.1 is available as a managed service and as a customer-hosted deployment. 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
- Omi does not have the lowest raw WER in this draw. Microsoft MAI-Transcribe-2 records 5.64% overall WER versus Omi's 5.77%. Omi leads the batch board on Medical WER and records fewer drug-term and dosage errors. Choose the metric weighting that matches your workflow.
- This is our development benchmark, not an untouched holdout. We use OmiMedSTT-Bench to develop and compare models and runtimes. Some reference text also appears in training data. The shared audio and scorer support consistent comparisons, but do not establish performance on unseen clinical recordings. Independent evaluations on customers' own audio remain important.
- English is the only medically evaluated language. Spanish, Portuguese, French, German, Dutch, Arabic and Hindi are supported. Formal medical evaluation in each of those languages is still in progress.
- Omi Medical v1.1 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
Our next focus is multilingual medical evaluation and broader independent testing. Batch and realtime results are available on the current benchmark.
Get a short note when new models and medical-language evaluations ship — no newsletter spam.
Where to find it
- Models & quickstart: omi.health/speech
- Standing benchmark: Medical Speech-to-Text Benchmark (batch and realtime)
- Open scorer & evaluation code: github.com/Omi-Health/medical-STT-eval
- API key / self-host: hello@omi.health
- 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 v1.1: accuracy, speed and safeguards. https://omi.health/research/omi-medical-1
@misc{omi_medical_1_2026,
title = {Omi Medical v1.1: accuracy, speed and safeguards},
author = {{Omi Health}},
year = {2026},
url = {https://omi.health/research/omi-medical-1},
note = {Flagship medical ASR, lowest Medical WER among 20 tested batch models in October 2026, EU-hosted API / private deployment}
}
Related research
- Medical Speech-to-Text Benchmark — separate batch and realtime leaderboards on the same recording 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