Hybrid Continual Learning for Low-Resource Australian Aboriginal Language Identification

arXiv cs.CL · 18 hours ago

Researchers developed two continual learning methods, Replay Augmented Elastic Weight Consolidation and Constraint Guided Knowledge Distillation, to identify endangered Australian Aboriginal languages using pretrained speech models. The methods were tested on Warlpiri, Dalabon, and Dharawal languages and showed better performance than fine-tuning baselines while preventing catastrophic forgetting. This enables speech technologies to support language revitalization for low-resource Aboriginal languages without losing capability on high-resource languages.