Three new transcription models on medical audio
Azure, OpenAI and Google joined the same sealed 1,513-clip clinical board.
See what changed →Medical voice infrastructure
Voice AI built for healthcare.
Add medical speech to the tools you are building—from a weekend prototype to a production healthcare AI product. Use the hosted API or deploy our open model yourself.
For clinicians who code · Healthtech startups · Healthcare AI companies
25 free audio-hours monthlySelf-serve BAANever used for model training
Consultation audio
“Start amoxicillin 500 mg and repeat HbA1c at review.”
✓ Drug and dosage preserved
#1 / 30medical-term accuracy
0.00%drug-name error
97.7%dosage F1 · 86/89 events
Built for the workflow
Turn clinical speech into product data.
Use one medical voice layer for the moments your users already speak through.
Consultations
Ambient documentation
Capture speakers, timing and the medical language needed for notes and downstream agents.
Dictation
Clinical text entry
Give clinicians a faster input method inside EHR, dental, therapy and specialty workflows.
Patient calls
Voice workflows
Feed confirmed medical text into triage, follow-up and care-navigation logic.
Local AI
Private on-device speech
Run the open model in your own application when audio cannot leave the environment.
Hosted API plans
Build for free. Add a card for production.
Every plan gets the complete medical API. Payment protects continuity and raises capacity.
Builder
Free
25 audio-hours includedPooled across batch and live. Resets every month.
For clinicians who code, independent builders and small healthcare AI projects.
- Complete medical batch and live API
- Medical vocabulary, speaker labels and timestamps included
- No card required · no rollover
- Self-serve BAA and DPA
Pay as you go
$0 base fee
$0.29 batch · $0.45 livePer audio-hour after your 25 included hours.
For startups and production products that need continuity without a contract.
- Your first 25 audio-hours each month remain free
- Live draft, confirmed text and a batch-quality final
- Higher production capacity
- No subscription or minimum
Enterprise
Custom
Contract pricing and capacitySized around your production workload.
For healthcare AI companies that need contracted capacity, guarantees or private deployment.
- Capacity sized for your workload
- Private cloud and supported SDK
- Service-level agreement
- Security and integration support
No transcription add-onsNever used for model training
Open-source medical speech
Trust it because you can run it.
omi-medical-edge-1 has open weights under CC-BY-4.0 and runs on Mac, NVIDIA CUDA or CPU. Inspect the model, test it on your own clinical audio, and keep every byte inside your environment.
Open weightsBenchmark, inspect and adapt the model yourself.
Runs locallyMac, CUDA and CPU runtimes.
0.6B parametersSmall enough to embed in real products.
Hosted when readyMove to the flagship API without rebuilding your product story.
Benchmark results
Every layer of the medical record, measured.
Overall transcription, medical terminology and dosage accuracy are scored independently on the same clinical audio.
Overall word error · lower is better
Azure5.97
Omi5.99
AWS6.12
ElevenLabs6.17
Medical-term error · lower is better
Omi0.94
ElevenLabs0.97
Google1.11
AssemblyAI1.43
Dosage F1 · higher is better
Omi97.7
Deepgram86.8
ElevenLabs85.4
Azure83.3
Fresh from Omi
Build from evidence, not marketing claims.
Practical guides, reproducible comparisons and open model work for people putting medical speech into products. Follow the RSS feed →
How to choose a medical speech-to-text API
Evaluate clinical accuracy, dosage safety, BAA access, speakers, languages and real price.
Read the guide →HIPAA-compliant speech-to-text: a practical checklist
What a healthcare builder should verify before sending protected health information.
Use the checklist →Trust center
Built for clinical data.
The hosted API keeps the operational choices visible, while the open model gives you a path to keep audio entirely inside your environment.
01
EU processing
Hosted API audio and transcripts are processed in the European Union.
02
Self-serve agreements
Sign the BAA and DPA in the console on Builder, before sending PHI.
03
1–72 hour retention
Choose how long stored job results remain available for retrieval.
04
Never used for training
Customer audio and transcripts are not used to train Omi models.
Start with your own audio.
Use the hosted API or download the open model.