Cross-lingual transfer from Sinhala improves Dhivehi ASR

A study investigates using cross-lingual transfer learning from Sinhala to improve automatic speech recognition for under-resourced Dhivehi. Researchers conducted seventeen experiments across five paradigms, including baselines and multilingual fine-tuning.

• The strongest system used continual pre-training on Sinhala followed by fine-tuning on Dhivehi with KenLM.
• This approach achieved 12.89% word error rate (WER) and 2.70% character error rate (CER).
• It outperformed the Dhivehi-only baseline by 13.50% WER and 3.02% CER.
• A Turkish control confirmed improvements stem from linguistic relatedness rather than general adaptation.

The results demonstrate that transfer learning from a linguistically related language significantly boosts low-resource ASR performance when combined with appropriate decoding configurations.