Клинический ASR: синтез обучающих данных для NeMo

★ 7.6 · content

digital-health-clinical-asr-build is a Claude Code skill that implements Stage 2 of the NVIDIA Clinical ASR Flywheel: it curates a clinical-specialty term list, resolves IPA pronunciations through a two-tier pipeline (override → Merriam-Webster → Magpie G2P), synthesizes audio via Magpie TTS over NVIDIA NVCF, and writes a NeMo-format manifest.jsonl tagged with clinical extension fields (term, entity_category, ipa_source, voice_id, noise_level, context_type). The skill is conversational — it asks specialty-aware clarifying questions, walks through an IPA audition gate before full Cartesian synthesis, and discloses external data transmission to Merriam-Webster and NVIDIA NVCF before any calls are made. Output includes audio/*.wav, manifest.jsonl, term_seed.csv, and pronunciation_overrides.csv, all ready for scoring at /digital-health-clinical-asr-eval. Requires /digital-health-clinical-asr-setup completed and a valid NVIDIA_API_KEY.