Оценка качества медицинского распознавания речи

★ 7.6 · content

digital-health-clinical-asr-eval is a Claude Code skill that implements Stage 3 of the Clinical ASR Flywheel: it takes a NeMo-format manifest, transcribes each audio clip via the selected ASR NIM (default: nvidia/parakeet-tdt-0.6b-v2 over NVIDIA NVCF gRPC), and computes four metrics — WER, CER, KER, and SER. The primary output is a five-section KER leaderboard: headline scores, breakdown by entity_category, by ipa_source (the merriam-webster vs. magpie_g2p delta proves the SSML override pipeline is working), by noise_level, and a worst-first per-term table. A built-in decision tree routes users to fine-tuning (/digital-health-clinical-asr-finetune), dataset repair (/digital-health-clinical-asr-build), or eval hardening. Designed for clinical NLP and medical speech-to-text teams who need granular diagnostics on term-level recognition accuracy. Requires a manifest from /digital-health-clinical-asr-build; real patient audio and PHI must not be passed through this skill.