Automatic Pronunciation Assessment Systems for English Students from Argentina

Abstract

English proficiency is crucial for education, work, social mobility, and global engagement. Despite government efforts across Latin America to expand language learning opportunities, teacher shortages create persistent disparities, particularly affecting low-income and rural students who cannot supplement public educa-tion with private lessons. Consequently, many graduate without basic English conver-sational skills. Computer-Assisted Language Learning (CALL) has improved lan-guage education by offering remote learning solutions, reducing teacher workloads, and providing stress-free practice opportunities. However, these systems remain suboptimal for pronunciation learning due to poor performance in detecting errors in short speech segments. Additionally, they historically emphasized native-like pronunciation rather than intelligibility understood as those mispronunciations that cause communication breakdowns. Our long-term goal is to develop a free mobile and web application tailored to the needs of Argentinian English learners. We focus on segmental-level pronunciation (phones or syllables), which research shows is more effec-tive for novice learners than phrase-level evaluation. We prioritize errors that are most impacting intelligibility. To address these challenges, we created EpaDB, a database of non-native English speech by Argentinian speakers for developing phone-level pronunciation scoring systems. We then explored two strategies for handling extreme data scarcity: first, a transfer learning approach that demonstrated significant im-provements over standard methods. Second, we compared different self-supervised learning speech models for the task. All our code is available for research purposes in an opensource repository.