Speech-to-Text

Medical speech-to-text
that doesn't guess.

Two models, one benchmark. #1 of 28 systems on medical accuracy — with zero fabricated drug names.

Create a free API key See the benchmark

#1 of 28

medical accuracy, sealed benchmark

0 drugs

fabricated, every test set

3 ways

cloud API · self-host · on-device

Two models

Pick the deployment, not the accuracy.

Flagship · API & self-host

omi-medical-1

The most accurate medical STT we measured. Managed cloud API, or self-hosted in your own tenant.

✓ #1 of 28 · M-WER 0.94% · 0 fabricated drugs
✓ Cloud API (live) or self-host
✓ Custom vocabulary, languages & async long audio
✓ Realtime available on request

Get an API key Model card

Open · on-device

omi-medical-edge-1

Open weights that run fully offline — for when audio can’t leave the machine.

✓ #1 open model that runs on-device
✓ Runs on Mac, CUDA or CPU
✓ Open weights (CC-BY-4.0) · MIT runtime

Model card & download Commercial SDK

API capabilities

One endpoint, shaped around the encounter.

Pick a language, pass the terms you expect, choose your output shape. The transcript stays the source of truth — timestamps and speaker labels attach without rewriting a word.

English + 7 languages

English is the only medically evaluated language. ES, PT, FR, DE, NL, AR and HI are supported — local medical evaluation ongoing.

enpt-BRes-419

Encounter vocabulary

Pass the 5–30 drugs, devices or names you expect this visit. Larger dictionaries live in the console.

vocabulary

Speakers & word timing

Speaker-labelled segments and word timestamps — structure attached to the sealed transcript, never changing its words.

verbose_json

Realtime by request

Live transcription over WebSocket, in preview. Enabled per API key while we expand capacity.

WebSocketpreview

Read the full API guide →

Benchmark

How they rank.

28 systems, identical audio, one open scorer — ranked by medical-term error rate. The flagship leads the cloud APIs; the edge model leads the open ones. Full board & methodology →

Closed-source APIs

System	M-WER	Drug
omi-medical-1 (ours)	0.94%	0.00%
ElevenLabs Scribe v2	0.97%	0.00%
Google Chirp 3	1.11%	1.13%
Gemini 3.1 Pro	1.34%	0.23%
Speechmatics Enhanced Medical	1.36%	1.81%

Open models

System	M-WER	Drug
omi-medical-edge-1 (ours)	2.16%	5.43%
Voxtral Mini Transcribe v2	2.47%	5.66%
Whisper Large v3 Turbo	2.54%	6.11%
Qwen3-ASR 1.7B	2.79%	6.11%
Parakeet TDT 0.6B v2	3.80%	8.60%

Deployable open models. A 9B research model scores higher on medical terms but can’t run at edge latency.

Quickstart

Running in minutes.

omi-medical-1 · Cloud API

curl https://api.omi.health/v1/audio/transcriptions \
  -H "Authorization: Bearer $OMI_API_KEY" \
  -F model="omi-medical-1" \
  -F language="en" \
  --form-string 'vocabulary=["Tinel","Hepcludex"]' \
  -F file=@consultation.wav

OpenAI-compatible — change the base URL and key. Create a free key or read the API guide →

omi-medical-edge-1 · On-device

pip install omi-med-stt

omi-med-stt consultation.wav
# runs offline on Mac, CUDA or CPU

Open weights, MIT runtime — audio never leaves the machine. Model card →

Medical speech-to-textthat doesn't guess.