Human and LLM accent rating of English-L2 speech by Brazilian speakers

This study examines whether Large Language Models (LLMs), specifically the free (Flash) and paid (Pro) versions of Google Gemini, can approximate human judgments of English-L2 accentedness produced by Brazilian speakers. Using telephone recordings from a subcorpus of the CSLU: 22 Languages Corpus, accent ratings from three native human judges were compared with AI-generated ratings on a four-point scale. Cumulative Link Mixed Models revealed systematic divergence across raters: AI systems, particularly the Pro version, consistently overestimated accent strength, assigning severe ratings to samples humans judged as moderate. These findings suggest that while human judgments integrate sociolinguistic and contextual cues, AI relies primarily on acoustic deviation, lacking the sociophonetic tolerance required for accent assessment.