Longitudinal Comparison of Artificial Intelligence-Based and Human Evaluations of English-as-a-Foreign-Language Pronunciation

This study examines pronunciation assessment in English as a Foreign Language, comparing holistic and atomistic evaluations by human raters—native and non-native speakers—and the artificial intelligence (AI) tool ELSA Speak. Over six months, university students in English as a Foreign Language were assessed across four phases. Two questions guided the research: How do ELSA Speak's holistic scores align with human judgments? How do atomistic evaluations of phonological errors differ between AI and humans? This study used a mixed-methods approach to analyze quantitative scores and qualitative error patterns. Findings revealed asymmetry: ELSA Speak consistently tracked progress, while human ratings were less stable and time sensitive. AI-powered tools focused on segmental features, whereas humans, especially natives, captured suprasegmental aspects. Rather than replacement, the study advocates integration—combining AI's diagnostic precision with human perceptual depth. A hybrid model emerges as the most pedagogically sound approach in the evolving landscape of language assessment.