digital-health-clinical-asr-finetune

Fine-tuning clinical ASR requires applying stock NeMo speech_to_text_finetune on Parakeet TDT 0.6B. This process reduces Key Error Rate (KER) for healthcare vocabulary like drug names, conditions, and procedures using supervised training.

Fine-tuning speech recognition for medical terms requires a sizable manifest of roughly 100 or more rows. The data must be split into stratified term-aware train and validation sets before running the NeMo SFT workflow.

Offline cycle N+1 evaluation confirms loop closure by testing the fine-tuned clinical ASR model after training. It measures KER improvements on healthcare vocabulary to verify the Parakeet TDT base model adapted correctly.

Yes, you can deploy fine-tuned NeMo ASR models to Riva NIM for production serving. This optional step follows the 3-epoch bf16-mixed SFT and offline evaluation phases in the end-to-end workflow.

ASR models misrecognize drug names and medical conditions due to limited clinical vocabulary in the base model. Fine-tuning Parakeet TDT 0.6B with NeMo SFT on domain-specific manifests reduces these Key Error Rate (KER) issues.