digital-health-clinical-asr-build

To build a NeMo manifest for clinical ASR, curate clinical terms, annotate IPA pronunciations, and synthesize a manifest with evaluation audio. This process outputs term_seed.csv, pronunciation_overrides.csv, and manifest.jsonl for Stage 3 evaluation.

The IPA annotation pipeline for medical speech recognition uses a two-tier process with override and Merriam-Webster lookups, supplemented by Magpie G2P. It generates pronunciation annotations for clinical terms to ensure accurate ASR evaluation.

Yes, you must complete the digital-health-clinical-asr-setup stage and secure Magpie TTS access via NVCF before generating clinical ASR evaluation audio. These prerequisites enable the two-tier IPA pipeline required for manifest synthesis.

Curating clinical terms for ASR datasets outputs three primary files: term_seed.csv for the initial term pool, pronunciation_overrides.csv for manual IPA corrections, and manifest.jsonl containing the NeMo-format evaluation audio and clinical extension fields.

To start an ASR benchmark manifest, create a term_seed.csv containing 4 to 10 clinical terms. This seed file initiates the cycle-based workflow, allowing you to run the IPA pipeline and generate evaluation audio.

A NeMo manifest for medical ASR includes clinical extension fields such as term, entity_category, ipa_source, voice_id, noise_level, and context_type. These fields support reproducible benchmarking across medical specialties.