Find your next open-source project

Search by name, purpose, or keyword — or paste a GitHub repository URL
All skills

digital-health-clinical-asr-eval

Transcribe a clinical ASR manifest and produce a five-section KER leaderboard with routing advice.

  • Part of NVIDIA/skills
  • 1.9k installs
  • Machine Learning · Speech models

When to use it

You already have a NeMo-format manifest.jsonl with clinical fields and need to know whether to fine-tune, rebuild, or stop.

For example

You say

Score my manifest.jsonl with 200 clinical rows using digital-health-clinical-asr-eval and tell me if I should fine-tune.

You get
  • per_sample.json
  • results.csv
  • leaderboard_cycle1.md

How it works

  • KER checks whether each flagged term appears in order and adjacent, so 'cefazolin' split into 'cefa zolin' counts as a miss.
  • The by-ipa_source split shows whether merriam-webster rows beat magpie_g2p rows, revealing a pronunciation-coverage gap rather than a model gap.
  • Routing depends on priority-category KER: above 0.3 suggests fine-tuning, below 0.1 suggests pushing the eval harder.

Before installing

Requires NVIDIA_API_KEY for hosted ASR NIMs and a NeMo-format manifest with clinical-extension fields; audio must be synthetic, not real patient recordings.