Speech-to-text, tuned for medicine
Build clinical voice applications with Deepgram Nova-3 Medical, optimized for clinical terminology and for multilingual transcription in both batch and real time.
Leading voice teams build with Deepgram
Clean transcripts, wherever care happens
Build medical transcription into your EHR platform, ambient scribe, or healthcare app, then customize it and deploy it your way.
AI scribes
Ambient notes
Deliver accurate, multilingual medical transcription for ambient AI scribes across healthcare settings. Capture clinical terminology with precision, speaker labels, and word-level timestamps, with sub-250ms streaming latency to support documentation as visits happen.
Instant adaptation to clinical terminology
Speaks your specialty's vocabulary
Use Keyterm Prompting to improve recognition of up to 100 custom terms, from specialty-specific terminology in cardiology, neurology, and oncology to new medications, procedures, and diagnoses. No model retraining required.
Enterprise-ready deployment
Deployment options for compliance
Choose managed cloud, on-premises, or your own VPC, and encrypt patient data at rest and in transit to meet your customers' security and regulatory requirements.
Automate healthcare workflows
Ready-to-use summaries
Deepgram Nova-3 Medical delivers accurate transcription in real time and batch, with summaries, topic detection, and sentiment analysis to power healthcare analytics and automation.
The most accurate speech-to-text for clinical audio
Benchmark-backed accuracy that holds up even in the toughest audio.
- 38%lower WER than the next-best competitor in English streaming
- 10 of 10languages outperforming evaluated streaming competitors in WER
- 9.7%lower batch WER than ElevenLabs Scribe v2 Medical
Made for medicine and made for your stack
Accurate across devices, specialty vocabulary, and speech patterns. Flexible enough to fit your pipeline and adapt to specialty-specific terminology.
Trained on clinical terms
Learns symptoms, diagnoses, treatments, medications, and clinical jargon from real medical audio, not just general conversation.
Accurate across recording conditions
Delivers reliable transcription whether the audio comes from a headset, a phone, or another recording device.
Gets real speech right
Handles disfluencies, cross-talk, and accents that show up in everyday clinical settings.
APIs for your existing pipeline
Get up and running in minutes with SDKs, clear docs, and a REST and WebSocket API.
Keyterm Prompting supports up to 100 terms
Send clinical vocabulary as query parameters without model retraining.
FAQ
Deepgram Nova-3 Medical is trained on clinical audio rather than adapted from a general model, so it transcribes symptoms, diagnoses, medications, and clinical jargon. It deploys in managed cloud, on-premises, or VPC with HIPAA-eligible infrastructure and a BAA available.
General models were trained on everyday speech, so an unfamiliar clinical term gets approximated to the nearest common word. Deepgram Nova-3 Medical was trained on medical audio, which is why it recognizes terms a general model has never encountered.
Deepgram Nova-3 Medical captures clinical terminology in real time, so an ambient note reaches the EHR needing review rather than rewriting. It returns speaker labels, word timings, and smart formatting in the same response.
Deepgram offers HIPAA-eligible infrastructure with a BAA available, plus Zero Data Retention for enterprise deployments. Deepgram Nova-3 Medical runs in managed cloud, on-premises, or VPC, so clinical audio can stay inside an environment your security team already approved.
Keyterm Prompting boosts up to 100 terms per request without retraining, allowing the same Deepgram Nova-3 Medical model to support multiple specialties or use cases.
Yes. Deepgram Nova-3 Medical uses speaker diarization to distinguish speakers in clinical conversations, with word-level timestamps, punctuation, paragraph formatting, and proper-noun handling.
Deepgram Nova-3 Medical transcribes pre-recorded clinical audio up to 40x faster than alternatives, and real-time streaming delivers transcripts in under 300ms.
Deepgram supports both use cases. Streaming handles live visits, dictation, and voice agents, and batch handles archives, quality review, and compliance work.