Navigating the Reality Gap: On-Device Continual Adaptation of ASR for Clinical Telephony
The research addresses how Automatic Speech Recognition (ASR) models degrade severely when moving from clean benchmarks to noisy, real-world clinical telephony audio due to a 'reality gap.' The study investigates on-device continual learning methods, specifically combining multi-domain Experience Replay (ER) and Elastic Weight Consolidation (EWC), to stabilize model adaptation and prevent catastrophic forgetting in privacy-sensitive settings.