Which Languages Transfer Best to Warlpiri? A Similarity-Based Study for Low-Resource ASR

2026-07-11

AI Summary

This study addresses the performance bottleneck in automatic speech recognition for the extremely low-resource language Warlpiri, caused by severe scarcity of labeled data, by proposing a source language selection framework that integrates acoustic similarity with multidimensional linguistic relatedness—including typological, phonemic, morphosyntactic, and syntactic features—to guide cross-lingual transfer learning. Leveraging the Whisper pre-trained model, the work systematically evaluates the predictive power of various similarity metrics—derived from acoustic embedding distances, typological databases, and phoneme inventory comparisons—on transfer effectiveness. Experiments demonstrate that Assamese and Hindi as source languages substantially reduce word and character error rates. Acoustic similarity emerges as the strongest predictor under fine-tuning, whereas phonemic and typological similarities prove more effective in zero-shot transfer. This research presents the first systematic integration of acoustic and multifaceted linguistic similarity, revealing their distinct roles across different transfer paradigms.

Abstract

This paper investigates how language similarity can improve cross-lingual transfer for automatic speech recognition (ASR) in extremely low-resource settings. Warlpiri, an Australian Aboriginal language, has very limited transcribed speech data, making transfer learning essential. We propose a framework combining acoustic similarity from pre-trained speech models with linguistic similarity based on typology, phoneme inventories, grammatical, and syntactic features to rank high-resource source languages and evaluate their effectiveness for ASR transfer to Warlpiri. Experiments with Whisper show that acoustically and typologically similar languages outperform monolingual and multilingual baselines. Assamese and Hindi achieve substantial reductions in word and character error rates. Correlation analysis further indicates that acoustic similarity is the strongest predictor of fine-tuning performance, while phoneme inventory and typological similarity better explain zero-shot transfer.

Problem

• low-resource ASR
• cross-lingual transfer
• language similarity
• Warlpiri
• automatic speech recognition

Innovation

• cross-lingual transfer
• language similarity
• low-resource ASR
• acoustic similarity
• typological features