Leading audio AI models produce systematic score gaps across English accents

A University of Washington-led team tested Gemini 2.5 Pro, Gemini 2.5 Flash, GPT-Audio 1.5 and Qwen 3 Omni on five accents in interviews, presentations and English tests, validating the setup with 52 human speakers. Chinese and Nigerian accents generally ranked lowest, Indian in the middle, and American and British highest, affecting both delivery and content scores. Hiring, education and immigration systems should publish subgroup evaluations and keep human review rather than relying on average accuracy.