Translating Low-Level Features into Student Friendly Explanations using XAI

An alarming decline in Gaeilge proficiency has renewed education efforts, fueled by the rise of AI in recent years. This research presents the development of a deep learning framework designed to evaluate spoken Gaeilge (henceforth Irish) proficiency of learners by quantifying student voice sample similarity to everyday Irish speaker truth samples, combined with explainable AI (XAI) techniques used to generate proof-of-concept student-friendly feedback. While automatic speech recognition (ASR) systems for high-resource languages like English have access to large, detailed datasets, the Irish language relies on a limited availability of speech resources, that challenges traditional model development. We address these gaps through a human-centered approach that employs a triplet-loss based Convolutional Neural Network (CNN) using spectrograms for similarity scoring, and XAI methods like using Mel-frequency cepstral coefficients (MFCCs) and Gradient-weighted Class Activation Mapping (Grad-CAM) to compare and contrast student attempts. Our evaluation combines technical metrics with visualisations, to make the learning process intuitive and simple for students. Results demonstrate that the model can successfully identify the similarity between student samples and truth samples, supplemented by power-transformed similarity scoring to help interpretability. This work contributes a scalable foundation for low-resource language education, demonstrating the potential for AI-assisted tools to aid in preserving languages like Irish while positioning XAI as a bridge between deep learning techniques and personalised, accessible learning for all students.