Omi speech-to-text

One medical product, from API to edge.

Use the flagship model through the hosted API or deploy the open edge model on your own hardware.

Choose a model
Flagship medical model

Production accuracy through one API.

Built for consultations, dictation and long clinical recordings, with vocabulary, eight-language transcription and speaker labels on the same endpoint.

0.94%medical-term error
0.00%drug-name error
97.7%dosage F1 · 86/89
5.99%overall word error

Tied for first of 29 systems by M-WER (the top two are statistically inseparable); the dosage result is significant against every competitor tested.

Hosted API features

Everything around the transcript.

One model for clinical workflows across eight supported languages.

01

Medical vocabulary

Send encounter names, products and terminology with each request.

02

Eight languages

English, Spanish, Portuguese, French, German, Dutch, Arabic and Hindi.

03

Speakers and timing

Speaker-labelled segments and word timing where supported.

04

Long audio

Async jobs, polling, webhooks and model-build provenance.

OpenAI-compatible

Call the hosted API.

Short clips return inline. Long audio returns an async job.

curl https://api.eu.omi.health/v1/audio/transcriptions \ -H "Authorization: Bearer $OMI_API_KEY" \ -F "[email protected]" \ -F "model=omi-medical-1"
Open weights · CC-BY-4.0

Medical transcription inside your product.

Run omi-medical-edge-1 locally on Mac, CUDA or CPU. Audio can stay entirely inside your environment.

0.6Bparameters
3Mac · CUDA · CPU
8.61%Mac MLX q8 WER
23 / 442Mac MLX q8 drug errors

Device numbers are kept separate from the faster GPU-served leaderboard row. See the model card for runtime-specific accuracy and memory.

Open model features

Own the complete runtime.

Open weights and a local runtime for products that cannot send audio away.

01

Open weights

Inspect, benchmark and adapt under CC-BY-4.0.

02

Fully local

No audio upload or hosted API dependency.

03

Three runtimes

Run on Mac, NVIDIA CUDA or CPU.

04

CLI and SDK

Start from the CLI or embed the runtime in your product.

Local deployment

Run it offline.

Install the runtime and transcribe without sending audio anywhere.

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

Choose your model.

Start with the API or bring the open model into your stack.