Hybrid Continual Learning for Low-Resource Australian Aboriginal Language Identification

Short summary

This paper proposes two hybrid continual learning methods—Replay Augmented Elastic Weight Consolidation and Constraint Guided Knowledge Distillation—to adapt pretrained speech models for identifying endangered Australian Aboriginal languages despite extreme data scarcity. Evaluated on Warlpiri, Dalabon, and Dharawal, the methods outperform standard fine-tuning and existing CL baselines by improving adaptation to new low-resource languages while preserving performance on previously learned high-resource languages.

• •Two hybrid CL methods combine replay-augmented EWC and constraint-guided knowledge distillation
• •Outperforms fine-tuning and CL baselines on Warlpiri, Dalabon, and Dharawal
• •Addresses catastrophic forgetting in extreme low-resource language identification

Generated with AI, which can make mistakes.