Modeling L1 Influence on L2 Pronunciation: An MFCC-Based Framework for Explainable Machine Learning and Pedagogical Feedback

Machine learning can identify specific acoustic features that reveal a speaker's native language influence on English pronunciation: The study found that just three acoustic measurements (related to energy, formant regions, and voicing) were more effective at detecting whether someone was a Mandarin or English native speaker than using all thirteen available measurements, demonstrating that simpler, more focused models can outperform complex ones.

Explainability matters for practical AI applications in education: By using interpretable machine learning methods and identifying which specific acoustic features drive predictions, the researchers created a system that teachers and curriculum designers can actually understand and act upon—rather than a "black box" that makes predictions without explanation.

Smaller, targeted datasets can be more efficient than larger ones: The reduced-feature model using only three key acoustic measurements significantly outperformed the full-feature model, suggesting that in AI development, identifying the most relevant data points can be more valuable than simply collecting more data.

AI-driven speech analysis could personalize language instruction: The findings suggest that automated systems could identify which specific pronunciation features are influenced by a student's native language, enabling teachers to provide targeted feedback and customize instruction rather than using one-size-fits-all approaches.

This research demonstrates responsible AI methodology for sensitive applications: The study combined multiple analytical approaches, tested results rigorously, and grounded findings in established linguistic theory—providing a model for how AI should be developed when it affects educational outcomes and potentially influences perceptions of speakers from different language backgrounds.