digital-health-clinical-asr-finetune

This is Stage 4 of NVIDIA's Clinical ASR Flywheel, the "adapt-and-measure" step. It solves the problem of actually closing the improvement loop: once Stage 3 eval shows priority-category KER above 0.3 on a sufficiently large manifest, this skill runs stock NeMo supervised fine-tuning (SFT) on a Parakeet TDT v2 base model and re-evaluates offline as cycle N+1 to measure that the model quality genuinely improved. It reports an empirically verified result on a reference manifest of baseline KER 0.513 dropping to 0.128 after three epochs, a 75 percent relative reduction, with the largest gains on drug names.

The skill is self-contained and prescriptive: it supplies gate criteria (fire only when priority KER exceeds 0.3 and the manifest has at least 100 rows), a base-model selection table, a stock-NeMo-SFT recipe using /opt/NeMo/examples/asr/speech_to_text_finetune.py inside the nvcr.io/nvidia/nemo:25.11.01 container, a hyperparameter table with rationale, and a cycle-N+1 decision table. It carries explicit anti-footguns: do not use the broken adapter-mixin path on TDT/RNNT decoders, and do not fine-tune the streaming Nemotron base whose SFT path collapses. A deep-dive reference covers hyperparameter reasoning, when to stop tuning, and full Brev GPU provisioning. Notably, where a cloud install uses a curl-piped install script, the reference calls this an antipattern and prescribes mitigations: download first and verify with shasum before running, pin to a release tag, or use Homebrew on macOS.

Target users are clinical AI/ML and healthcare TME engineers with access to a CUDA host (24 GB VRAM comfortable, 16 GB workable) or a Brev cloud GPU. NVIDIA_API_KEY is required for the offline re-eval round-trip and any NIM deploy, and the finetune-asr and riva-asr-custom skills are expected alongside it.