From Sinhala to Dhivehi: Cross-Lingual Transfer Learning for Low-Resource Speech Recognition

2026-07-07

AI Summary

This study addresses the scarcity of resources for Dhivehi automatic speech recognition (ASR) by systematically investigating cross-lingual transfer learning from the closely related Sinhala language. The authors evaluate five transfer paradigms—including continued pretraining, multilingual fine-tuning, and sequential fine-tuning—and integrate a KenLM language model to optimize decoding. Through comprehensive experiments, they assess the impact of linguistic proximity, adaptation strategies, and decoding configurations on transfer performance. The best-performing system achieves a word error rate (WER) of 12.89% and a character error rate (CER) of 2.70% on a Dhivehi test set, representing a significant 13.50% relative WER reduction over the baseline. These results demonstrate the strong potential of leveraging Sinhala data for effective ASR transfer to Dhivehi.

Abstract

Dhivehi, the national language of the Maldives, is currently under-resourced for automatic speech recognition (ASR) and other NLP tasks. This study investigates whether cross-lingual transfer learning from Sinhala, a linguistically related, relatively well-resourced Insular Indo-Aryan language, can improve Dhivehi ASR. We conduct seventeen experiments across five transfer learning paradigms: Dhivehi-only baselines, sequential fine-tuning, multilingual fine-tuning, continual pre-training, and a control using Turkish as an unrelated language. The strongest system, continual pre-training on Sinhala followed by fine-tuning on Dhivehi with KenLM, achieves 12.89% WER and 2.70% CER, outperforming the Dhivehi-only baseline by 13.50% WER and 3.02% CER. However, the adaptation strategy and decoding configuration are equally critical for a successful transfer learning experiment. We conduct seventeen controlled experiments spanning five transfer learning paradigms: Dhivehi-only baselines, sequential fine-tuning, multilingual fine-tuning, continual pre-training, and a control experiment using Turkish as an unrelated language. The strongest system, continual pre-training on Sinhala followed by fine-tuning on Dhivehi with KenLM, achieves 12.89% WER and 2.70% CER, outperforming the Dhivehi-only baseline by 13.50% WER and 3.02% CER. The Turkish control experiment confirms that observed improvements stem from linguistic relatedness; adaptation strategy and decoding configuration are also critical.

Problem

• Dhivehi
• low-resource
• automatic speech recognition
• cross-lingual transfer

Innovation

• cross-lingual transfer learning
• low-resource ASR
• continual pre-training
• language relatedness
• KenLM