Abstract

AI-powered listening trainers increasingly dominate language education, yet their accent selection remains biased toward Inner Circle Englishes (e.g., General American, RP). This study evaluates the impact of accent-diverse AI trainers on listening comprehension, anxiety reduction, and pragmatic competence. Using a mixed-methods approach with 412 intermediate learners (A2-B2), we tested an AI system exposing learners to 8 Global English accents (Nigeria, India, Singapore, Jamaica, Scotland, etc.). Quantitative results showed 23.7% higher comprehension accuracy (p<0.01) for diverse accents in international communication scenarios. Qualitatively, 81% reported reduced "accent anxiety." We further propose a decolonial data curation framework to mitigate algorithmic accent bias. Findings challenge the monolingual paradigm in AI listening tools and advocate for intentional accent diversity as a pedagogical imperative.