Continual Adaptation for Pacific Indigenous Speech Recognition

The study presents an empirical analysis on adapting speech foundation models to low-resource Pacific Indigenous languages, highlighting the challenges posed by data scarcity and the risk of catastrophic forgetting during full fine-tuning. It examines the effects of data volume, adaptation strategies, and representational drift, revealing that while LoRA initially adapts effectively, it ultimately suffers from severe internal representational drift and memory issues in sequential learning contexts. This research underscores the necessity for robust adaptation strategies to enhance model performance for underrepresented languages, which is critical for practitioners developing AI systems in multilingual settings.