digital-health-clinical-asr-setup

This is Stage 1 of NVIDIA's Clinical ASR Flywheel, the bootstrap step. Its single job is to prove that a user can reach NVIDIA's hosted speech stack with the NVIDIA_API_KEY they currently hold, by round-tripping one clinical sentence through Magpie TTS and Parakeet/Nemotron ASR. Once that smoke test passes, the user is cleared to advance to Stage 2 (build). The flywheel as a whole exists to drive down KER (keyword error rate) on clinical entities such as drugs, procedures, anatomy, conditions, labs, and roles, where averaged WER would otherwise hide clinically important failures.

Stage 1 is three concrete steps: verify the API key, install Python dependencies (Python 3.10+), and run the TTS-to-ASR smoke test. The skill is explicit that there is no installer script hidden anywhere in it, and that everything beyond Stage 1 is composed from sibling skills. A dependency-ownership reference draws clear boundaries around what the flywheel family owns (the clinical-ASR methodology, KER metric, two-tier IPA tagging, manifest schema extension, base-model table, inlined offline ASR gRPC recipe) versus what it defers to riva-tts, riva-asr, finetune-asr, riva-asr-custom, riva-nim-setup, and data-designer.

Most prominently, the skill front-loads an outbound-data disclosure: two external endpoints (NVIDIA NVCF over grpc.nvcf.nvidia.com and Merriam-Webster) receive synthetic sentences, transcribed audio, and individual clinical terms, and the user must acknowledge both against their organization's data-governance policy before Stage 2 begins. It repeatedly warns against sending real patient transcripts, recorded clinical audio, or any PHI. Target users are clinical AI/ML and healthcare TME engineers standing up a clinical ASR benchmarking cycle. Its pre-publication benchmark returned an overall PASS verdict.