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

2026-07-07 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

This study explores cross-lingual transfer learning to improve automatic speech recognition (ASR) for Dhivehi, a low-resource language, by leveraging Sinhala, a linguistically related and better-resourced language. Researchers conducted seventeen experiments across five transfer learning paradigms: Dhivehi-only baselines, sequential fine-tuning, multilingual fine-tuning, continual pre-training, and a control using Turkish. The most effective approach involved continual pre-training on Sinhala followed by fine-tuning on Dhivehi with KenLM, which achieved a 12.89% Word Error Rate (WER) and 2.70% Character Error Rate (CER). This significantly outperformed the Dhivehi-only baseline by 13.50% WER and 3.02% CER. The findings confirm that linguistic relatedness is crucial for improvements, and emphasize the importance of adaptation strategy and decoding configuration for successful transfer learning experiments.

Key takeaway

For NLP Engineers developing ASR systems for low-resource languages, you should prioritize cross-lingual transfer learning from linguistically related, better-resourced languages. Specifically, consider continual pre-training on the related language before fine-tuning on your target language, integrating robust decoding configurations like KenLM. This approach can yield significant performance improvements, as demonstrated by a 13.50% WER reduction for Dhivehi ASR. Carefully evaluate adaptation strategies and decoding settings for optimal results.

Key insights

Cross-lingual transfer from linguistically related languages significantly improves low-resource ASR, but adaptation strategy and decoding configuration are equally critical.

Principles

• Linguistic relatedness drives transfer gains.
• Adaptation strategy is crucial for success.
• Decoding configuration impacts ASR performance.

Method

The most effective method involved continual pre-training on a related language (Sinhala), followed by fine-tuning on the target low-resource language (Dhivehi) with KenLM for decoding.

In practice

• Apply continual pre-training from related languages.
• Test various adaptation strategies.
• Optimize ASR decoding configurations.

Topics

• Dhivehi ASR
• Cross-Lingual Transfer
• Low-Resource NLP
• Continual Pre-training
• KenLM
• Speech Recognition

Best for: Research Scientist, AI Scientist, NLP Engineer, Machine Learning Engineer

Related on AIssential

• Low-resource Language Discrimination Towards Chinese Dialects with Transfer learning and Data Augmentation — Transfer learning and data augmentation effectively overcome scarce annotation resources for low-resource Chinese dialect discrimination.
• Zero-Shot to Full-Resource: Cross-lingual Transfer Strategies for Aspect-Based Sentiment Analysis — Cross-lingual transfer strategies for ABSA vary in effectiveness based on model architecture and resource availability.
• The Tatoxa System for Text Detoxification in Low-Resource Languages: The Case of Tatar — The Tatoxa system achieves state-of-the-art text detoxification for Tatar, emphasizing native data's superiority over cross-lingual transfer.
• Giving Voice to the Constitution: Low-Resource Text-to-Speech for Quechua and Spanish Using a Bilingual Legal Corpus — Cross-lingual transfer in TTS effectively mitigates data scarcity for low-resource languages like Quechua.
• KIT's Submission to Cross-Lingual Voice Cloning in IWSLT 2026 — Explicit language tags and lexical matching significantly improve cross-lingual voice cloning performance and reduce accent leakage.
• Low-resource Language Discrimination Towards Chinese Dialects with Transfer learning and Data Augmentation — Transfer learning and data augmentation effectively overcome low-resource challenges in Chinese dialect discrimination.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.