Title: Bridging the Divide: On-Device Continual Adaptation of ASR for Clinical Telephony

Abstract:

While Automatic Speech Recognition (ASR) has the potential to substantially alleviate documentation pressures in clinical workflows, its performance often suffers in actual telephony environments. This degradation is driven by factors such as acoustic noise, dialectal diversity, and stringent data residency requirements that hinder cloud-based adaptation strategies. To investigate this "reality gap," we utilize Gram Vaani—a telephonic Hindi dataset covering rural healthcare and agricultural helplines—as the most viable proxy for clinical speech under strict on-device limitations. Our analysis reveals that the robust multilingual IndicWav2Vec model experiences a significant performance drop, with Word Error Rate (WER) rising from 11.59% on clean standard Hindi to 41.71% when applied to this proxy telephony data.

We assess a series of on-device adaptation strategies that adhere to realistic constraints, ranging from full fine-tuning to parameter-efficient LoRA and stream-based continual learning. These evaluations span various baselines, datasets, and random seeds. Our primary focus on continual learning uncovers a pivotal interaction between Experience Replay (ER) and Elastic Weight Consolidation (EWC), particularly regarding the regularization strength $\lambda$. We demonstrate that conventional positive EWC ($\lambda > 0$) can conflict with updates driven by replay, thereby restricting the model's ability to adapt. Conversely, inverting EWC’s strength to negative values ($\lambda < 0$) reveals its potential as a directional control mechanism within ER-guided adaptation. Specifically, a negative $\lambda$ enhances the plasticity initiated by replay, while a scheduled $\lambda$ allows for phase-dependent management of the stability-plasticity balance.

Our multi-dataset evaluations indicate that while multi-domain replay establishes a robust foundation for adaptation, EWC serves to modulate the stability-plasticity dynamics without changing the ultimate performance outcomes. These findings underscore that successful on-device adaptation relies not on selecting methods in isolation, but on comprehending the interplay between data-driven updates and parameter-level learning signals.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC