ASR for Noisy Radio Speech, TTS for Fine-Tuning
Turning noisy EMS radio into structured clinical data before the patient arrives.
A current project (Harvard collaboration, PI: Gabriel Brat, MD) building an agentic voice interface for pre-hospital neurological emergency triage (NEI-6).
The pipeline converts noisy EMS radio audio into structured clinical markers before the patient arrives:
- Synthetic data generation via TTS+LLM to overcome scarce, sensitive real-world audio,
- fine-tuned ASR on a synthetic–real mix,
- LLM-based transcript correction to recover clinically meaningful content from degraded transmissions.
The evaluation target is downstream clinical-marker accuracy, not just transcript word-error-rate.