Author Information
Xiang-Ling Chen, National Tsing Hua University, TaiwanChingching Lu, National Tsing Hua University, Taiwan
Abstract
This study aims to develop and evaluate an AI-assisted pronunciation feedback tool to enhance dialectal awareness among kindergarten Hakka language teachers in Taiwan. This research investigates whether real-time, AI-generated feedback can help teachers recognize and correct their own pronunciation patterns. Due to the dominance of the other dialect and teachers’ varied phonetic backgrounds, many of them unintentionally teach a blended variant. To tackle this problem,we developed an AI-assisted pronunciation evaluation tool designed to raise dialectal awareness among language teachers. Using the Wav2Vec 2.0 model pre-trained on the LibriSpeech corpus and fine-tuned with 300 annotated Hakka sound samples focused on single words, the system provides teachers with real-time feedback on pronunciation deviations. Key evaluation metrics included accuracy, recall, precision, and F1 score, with the current model achieving 81% classification accuracy on new input data. In practice, five Hailu-speaking kindergarten teachers tested the tool. While they noted that the system's precision still needed to be improved, all participants reported increased sensitivity to dialectal differences over time. The AI feedback gradually enhanced their ability to distinguish between Sixian and Hailu pronunciations— an awareness that had previously been lacking. These findings suggest that AI tools can support professional development by helping teachers self-monitor and correct pronunciation errors before they are modeled to children. Therefore, this study demonstrates a new pathway for language teacher training in dialectally diverse settings. It could lead to establish AI-assisted pronunciation training as a pre-instructional standard, particularly for languages and dialects vulnerable to phonological convergence and loss.
Comments
Powered by WP LinkPress