Nova-3 Medical Streaming: Pushing Real-Time Medical Transcription to New Heights

Earlier this year, we introduced Nova-3 Medical, Deepgram's most advanced speech-to-text model designed specifically for clinical environments. With industry-leading accuracy, support for highly specialized medical terminology, and security designed to meet HIPAA and healthcare regulatory requirements, Nova-3 Medical set a new standard for AI-powered transcription in healthcare. But the work did not stop there.

Recently, we raised the bar again with Nova-3 Medical Streaming, a major upgrade that brings clinical-grade accuracy to real-time transcription without sacrificing the ultra-low latency required by live clinical workflows.

The Evolution: Extending Nova-3 Medical’s Accuracy Into Streaming

When we launched Nova-3 Medical, we introduced significant accuracy improvements across medical transcription, setting new benchmarks for batch evaluation. Now, with the latest streaming upgrade, we are extending those accuracy gains even further into real-time use cases.

The initial Nova-3 Medical release delivered:

• 63.7% WER improvement over the next-best competitor

• 40.35% reduction in Keyterm Error Rate (KER)

• 10.6% improvement in Keyword Recall Rate (KRR)

These metrics translated directly into real-world impact. Fewer medication name errors, sharper symptom capture, and cleaner notes flowed into EHR systems with minimal manual correction. But for live clinical scenarios such as telemedicine consults, bedside note-taking, or ambient scribe solutions, transcription speed matters just as much as accuracy.

Nova-3 Medical Streaming closes that gap.

Why Nova-3 Medical Streaming Matters

In healthcare, even small transcription delays or inaccuracies can create cascading downstream effects, from physician frustration to billing errors and potential patient safety risks. Nova-3 Medical Streaming directly addresses these challenges by delivering:

• 11% WER reduction vs. Nova-3 General streaming

• 30% WER reduction vs. Nova-2 Medical streaming

• 2.7x improvement in Keyword Recall Rate (KRR) vs. Nova-3 General streaming

• Maintained ultra-low latency for real-time interactions

What this means in practice is that drug names are captured correctly in the moment. Symptom details are recorded accurately as they are spoken. The transcription keeps pace with the clinician even during fast, natural speech, minimizing the need for backtracking, clarifications, or corrections.

Built for Real-World Clinical Complexity

Healthcare audio is uniquely challenging. Multiple speakers, far-field microphones, equipment noise, and dense medical terminology all combine to strain transcription systems. Nova-3 Medical Streaming is designed to operate in exactly these environments.

As in batch transcription, the streaming update applies the same set of capabilities to handle real-world clinical audio:

• Robust model training designed to maintain accuracy even in challenging audio conditions, including far-field speech, multiple speakers, and clinical background noise.

• In-context learning with Keyterm Prompting, allowing developers to inject up to 100 custom terms to handle emerging drug names, procedures, or specialty vocabularies.

• HIPAA-compliant security architecture, ensuring patient data remains protected at every stage.

Whether deployed in scribe applications, ambient documentation tools, telehealth platforms, or contact center triage, Nova-3 Medical Streaming ensures that transcription quality remains high without compromising patient privacy or system responsiveness.

Seamless Deployment for Developers

For teams already building with Deepgram, upgrading to Nova-3 Medical Streaming is simple:

wss://api.deepgram.com/v1/listen?model=nova-3-medical

No additional integration work is needed. The model is live across all hosted Deepgram regions, with self-hosted deployment options coming soon.

The Deepgram Approach: Continuous Innovation for Healthcare AI

Nova-3 Medical Streaming reflects our broader commitment to voice AI infrastructure purpose-built for healthcare applications. Rather than forcing general-purpose models into clinical workflows, we continue to refine models tuned for the unique needs of healthcare providers, tech ISVs, and digital health innovators.

From batch to streaming, from far-field dictation to conversational AI agents, our models are optimized for real-world medical scenarios where accuracy, latency, security, and cost-efficiency all matter.

With Nova-3 Medical Streaming, we enable developers to deliver clinical-grade transcription at the speed of conversation.