Open source

Run medical transcription on your hardware

Try Omi Med STT v1 through the hosted API, or run the same open English medical model on your own hardware. The hosted route is intended for evaluation; use omi-medical-1 for production workflows.

GitHub PyPI Model card Try through the API

Try the hosted model

curl https://api.omi.health/v1/audio/transcriptions \
  -H "Authorization: Bearer $OMI_API_KEY" \
  -F file=@consultation.m4a \
  -F model=omi-medical-edge-1 \
  -F language=en \
  -F diarize=true \
  -F max_speakers=2 \
  -F response_format=verbose_json

omi-medical-edge-1 accepts English short and long audio and returns transcript segments and word timestamps. Add diarize=true for speaker labels; the transcript words and their native timestamps remain unchanged. It is $0.10 per processed audio-hour after the same pooled 25-hour monthly allowance used by the other speech routes. Internet upload time is separate from processing time.

Choose the flagship for production

Use omi-medical-1 for production pipelines, realtime transcription, all supported languages, encounter vocabulary, Patterns, and multilingual or realtime work. Hosted Edge is English batch ASR with optional speaker labels and may queue or return a retryable capacity response while flagship traffic has priority.

Install and transcribe

# Base CLI
pip install -U omi-med-stt

# Apple Silicon / MLX
pip install -U "omi-med-stt[mlx]"

omi-med-stt consultation.m4a --json

# Verify the local runtime and model artifact
omi-med-stt check

The package reads common audio formats, downmixes to mono, and resamples to 16 kHz. Model weights download on first use from the published Hugging Face repositories.

Choose a runtime

Hardware	Runtime	Released artifact
Apple Silicon	`mlx`	MLX q8
CUDA-capable GPU	GPU runtime	Canonical published weights
Linux / Windows CPU	`cpp`	GGUF q8_0

Runtime exports have separate benchmark rows because quantization and conversion can change accuracy. Use the result for the artifact you actually deploy, not a different published artifact's number. See GitHub for the current platform-specific GPU installation command.

Cloud and local are different products

Capability	Hosted Edge	Open-source runtime
Processing	Managed in the EU	Inside your process
Languages	English	English
Long jobs	Yes	Your application owns the workflow
Word timestamps	Yes	Available from JSON output
Speaker labels	Optional, up to four speakers	Not part of the released CLI contract
Realtime and vocabulary	Use omi-medical-1	Not part of the released CLI contract
License	Service terms	MIT runtime code; CC-BY-4.0 weights

Review transcripts before clinical use

The local runtime is speech-to-text, not a diagnostic, triage, prescribing, or clinical decision system. It is not clinically validated. Keep human review in the workflow.

Continue

Hosted API quickstartUse Edge for evaluation or the flagship for production workflows.Production checklistOperate the hosted API safely and predictably.