Gemini and GPT-Audio scored Chinese and Nigerian accents lowest in simulated interviews and tests
A University of Washington-led team tested four audio AI models on five English accents, and American and British speakers scored highest.
In brief
AI judges gave some accents lower scores
Four audio AI models scored mock interviews and tests. American and British accents ranked highest. Chinese and Nigerian ranked lowest, even with identical words.
Companies now use AI to screen recorded job interviews, and some English tests used for immigration rely on it to score spoken answers. A University of Washington-led team found that four leading models gave lower marks to certain English accents in mock interviews, conference talks and spoken English tests. The team presented the work at the EMNLP conference in Budapest.
Why it matters
Take two applicants who read out the same answer to the same question. If a scoring tool marks one down because of the sound of their voice, the content did not decide the outcome.
Co-lead author Orevaoghene Ahia, a UW doctoral student, said people now trust audio models as third-party evaluators in hiring, exams and some immigration English tests. Yulia Tsvetkov is the senior author.
What they found
American and British voices got the best scores. Indian voices landed in between, and Chinese and Nigerian voices got the worst.
The pattern persisted when every speaker said identical words and the models transcribed the speech accurately. According to the UW release, American and British voices scored higher on both delivery and content.
The second source, a Kurt Schlosser article, adds that the team also measured how far each accent sits from American English in sound. Chinese and Nigerian accents were the furthest away.
What the team tested
The models were Google's Gemini 2.5 Pro and Gemini 2.5 Flash, OpenAI's GPT-Audio 1.5, and the open-source Qwen 3 Omni. The researchers say these are the kind of systems companies would actually deploy.
Five accents were compared: American, British, Chinese, Indian and Nigerian. The team chose regions with large English-speaking populations and where the researchers had connections.
How the test worked
To test at scale, the team used AI voices from ElevenLabs to read the same scripts: interview answers, a class presentation, a conference talk and English test responses. Male and female voices were used. The models gave scores from one to seven. The release says the results were checked against 52 human speakers from the five countries.
Limits
The tests were simulated, and most used synthetic voices. Only five accents and four models were covered, so other accents and systems were not tested. One aggregator reports further details, but the UW release and the Schlosser article do not confirm them. Independent coverage so far is thin, and the sources do not say whether any company has responded.