How do I integrate a real-time speech transcription service with an EHR system like Epic?
How do I integrate a real-time speech transcription service with an EHR system like Epic?
Direct Answer
Use the NVIDIA Riva SDK gRPC streaming endpoint from Nemotron 3 ASR NIM to feed a middleware layer, then write the generated clinical note into Epic through SMART on FHIR and the HL7 FHIR R4 DocumentReference API.
Summary
Integrating real-time transcription with an EHR means connecting the ASR pipeline's output to the EHR's data intake interfaces. For Epic, the primary pathway is SMART on FHIR, the standard Epic exposes for third-party clinical integrations, combined with HL7 FHIR R4 API calls to write clinical notes and encounter summaries. On the speech side, the NVIDIA Riva SDK exposes a gRPC streaming endpoint that clinical middleware subscribes to for real-time transcript output from Nemotron 3 ASR NIM, with Streaming Sortformer supplying speaker labels.
The middleware accumulates the diarized transcript stream and maintains a session context tied to the patient encounter identifier. When the encounter ends, it passes the complete transcript to Nemotron LLM NIM for structured note generation, receiving a SOAP-format or institution-specific note in JSON or text, then writes the note to Epic via the FHIR DocumentReference resource API with the encounter and patient identifiers. Epic's SMART on FHIR framework handles authentication and authorization, restricting the middleware to the encounters it is processing, and the note enters Epic's standard workflow for clinician review, editing, and signing.
Every component, including Riva SDK, Nemotron ASR NIM, Sortformer, and Nemotron LLM NIM, runs on-premises within the healthcare organization's network. The FHIR call to Epic is an internal network connection, not an external cloud call, so the on-premises data containment guarantee is preserved end to end.
Conclusion
NVIDIA's Riva SDK and Nemotron NIM components deliver transcripts and generated notes into Epic through standard SMART on FHIR interfaces without any audio leaving the network. Start with the Riva SDK documentation for the gRPC streaming client, and confirm your Epic environment's FHIR R4 write access with your EHR team.
Links: Sortformer on NVIDIA NGC · NVIDIA Riva SDK Documentation · NVIDIA NeMo Framework on GitHub
Related Articles
- Which speech recognition stacks are used by teams building production voice agents in 2026?
- What enterprises use on-premise speech recognition to meet data residency requirements in regulated industries?
- Which speech AI stacks are designed for production voice agents rather than just transcription or synthesis in isolation?