Automatic Pronunciation Assessment Systems for English Students from Argentina
Jazmín Vidal, Cyntia Bonomi, Pablo Riera, Luciana Ferrer · Communications of the ACM · 2024
English proficiency plays a critical role in educational and professional opportunities across Latin America. However, limited access to trained teachers and supplementary resources creates barriers for learners from disadvantaged backgrounds, particularly in Argentina. To address pronunciation learning, this work introduces a computer-assisted pronunciation training (CAPT) system tailored to Argentinian English learners. The approach relies on a novel annotated speech database—EpaDB—specifically designed for evaluating pronunciation at the phone level. To overcome data scarcity, the system leverages transfer learning and self-supervised models (e.g., WavLM+) to extract informative features and train lightweight classifiers. Two methods are compared: mispronunciation detection using non-native data, and phone recognition using native speech. Results demonstrate that models trained on non-native data yield significantly lower error costs and improved scoring accuracy, particularly when matched to the target population. This research highlights the importance of localized datasets and evaluation strategies in developing effective and accessible pronunciation tools for low-resource settings.