Aspects of Selecting the Right ASR Training Languages for Under-Resourced Languages

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computational Linguistics · Depth: Expert, medium

Summary

A study presented at ComputEL-9 in July 2026 by J. Elizabeth Liebl, Summer Chambers, Matthew Kelley, and Géraldine Walther investigates optimal training language selection for cross-lingual IPA Automatic Speech Recognition (ASR) on under-resourced languages. The researchers trained multilingual IPA-based ASR models for Upper Sorbian, Luganda, and Tatar using Common Voice audio and Vox Communis phonetic transcripts. They evaluated three linguistically motivated strategies—genealogical relatedness, geographic proximity, and phonological inventory overlap—against a random baseline, measuring performance with phone error rate. While linguistically informed selection generally improved transfer, no single strategy proved consistently optimal. Geographic proximity was most effective for Luganda, phonological overlap slightly better for Tatar, but none of the strategies surpassed random selection for Upper Sorbian. The findings indicate that linguistic similarity aids low-resource ASR transfer, though the most beneficial similarity dimension varies by target language.

Key takeaway

For NLP Engineers developing ASR systems for under-resourced languages, you should move beyond generic cross-lingual transfer approaches. Your selection of training languages must be tailored, as the most effective linguistic similarity dimension—whether genealogical, geographic, or phonological—is highly language-dependent. Systematically evaluate different similarity metrics for your specific target language, as a "one-size-fits-all" strategy may not outperform random selection, particularly for languages like Upper Sorbian.

Key insights

Optimal linguistic similarity for low-resource ASR transfer is target-language dependent.

Principles

• Linguistically informed selection generally improves ASR transfer.
• No single linguistic similarity strategy is universally optimal.
• Optimal similarity dimension varies per target language.

Method

Train multilingual IPA-based ASR models using Common Voice audio and Vox Communis phonetic transcripts, evaluating genealogical, geographic, and phonological similarity strategies against a random baseline with phone error rate.

In practice

• Consider geographic proximity for languages like Luganda.
• Explore phonological overlap for languages like Tatar.
• Be aware that some languages may not benefit from these strategies.

Topics

• ASR
• Under-resourced Languages
• Cross-lingual Transfer
• IPA ASR
• Language Similarity
• Phonological Inventory

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• Towards Truly Multilingual ASR: Generalizing Code-Switching ASR to Unseen Language Pairs — Code-switching ASR generalization to unseen language pairs remains limited despite model merging and domain generalization efforts.
• Hard to Be Heard: Phoneme-Level ASR Analysis of Phonologically Complex, Low-Resource Endangered Languages — Data scarcity, not phonological complexity, primarily drives ASR errors in low-resource, typologically complex languages.
• Multilinguality of Large Language Models From a Structural Perspective — LLMs' multilinguality involves structural differences, with low-resource languages diverging more from English.
• Improving low-resource ASR using bilingual fine-tuning with language identification: a cross-linguistic evaluation — Bilingual fine-tuning enhances low-resource ASR, particularly when language identification is accurate or explicitly guided during inference.
• 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.
• 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 Paper Index on ACL Anthology.