# Omi Health

> Medical voice infrastructure for healthcare AI builders and companies.

Omi Health provides a hosted medical speech-to-text API and an open medical ASR model. The hosted lifecycle is Builder, Pay as you go and Enterprise. The open model can be evaluated through the hosted API or self-deployed; it is a model choice, not another pricing plan.

## Start building

- [Product overview](https://omi.health/speech): Hosted batch and live API plus the open edge model.
- [Guides and evidence](https://omi.health/guides): Medical speech evaluation, implementation and compliance guidance.
- [Research](https://omi.health/research): Model cards, evaluation methods and technical findings.
- [Pricing](https://omi.health/pricing): 25 pooled audio-hours included every month; then $0.10 hosted Edge, $0.29 flagship batch and $0.45 live per audio-hour.
- [Playground](https://omi.health/playground): Test medical transcription on your own audio.
- [API documentation](https://docs.omi.health): Hosted API reference.
- [Open model](https://omi.health/research/omi-med-stt): omi-medical-edge-1 hosted evaluation, model card and self-deployment instructions.
- [Runtime engineering deep dive](https://omi.health/research/omi-med-stt-runtime): Same weights, 21–23% lower WER than earlier published runtime rows across NVIDIA CUDA, Apple MLX and CPU.

## Evaluate Omi

- [Medical speech benchmark](https://omi.health/benchmark): 30 systems evaluated on the same 1,513 clinical clips.
- [Compare providers](https://omi.health/compare): Accuracy, product features and published pricing.
- [Medical speech API buyer guide](https://omi.health/guides/best-medical-speech-to-text): Compare 30 APIs and open models on clinical accuracy, BAA access and pricing.
- [HIPAA speech API checklist](https://omi.health/guides/hipaa-compliant-speech-to-text): Contract, controls, retention and workflow questions.
- [Medical word error rate guide](https://omi.health/guides/medical-word-error-rate): WER, medical-term error, drug-name error and dosage F1.
- [Healthcare speaker diarization guide](https://omi.health/guides/medical-speaker-diarization): Speaker attribution, failure modes and a clinical test plan.

## Current benchmark summary

- Overall word error: 5.99%.
- Medical-term error: 0.94%, statistically tied for first of 30 systems.
- Drug-name error: 0.00% on 442 drug mentions.
- Dosage F1: 97.7% on the published 89-event dosage set.
- The benchmark covers prerecorded clinical English; it is not evidence for realtime latency or multilingual quality.

## Trust boundary

- Hosted API processing is in the European Union.
- A self-serve BAA and DPA are available from Builder.
- Stored job-result retention can be configured from 1 to 72 hours.
- Customer audio and transcripts are never used to train Omi models.
- [Trust Center](https://omi.health/trust)
- [Legal documents](https://omi.health/legal/)

## Machine-readable updates

- [RSS feed](https://omi.health/feed.xml)