(Credit: © patpitchaya - stock.adobe.com)
AI Graders May Penalize Accents That Sound Less American
In A Nutshell
- Several AI systems that grade speech gave Chinese and Nigerian accents lower delivery scores than American and British ones, even when everyone read the same script.
- The pattern also showed up with real human voices in hiring scenarios, and the AI mishearing words did not explain it.
- The farther a voice sounded from American English, the lower its delivery score tended to be.
- Content scores sometimes slipped in the same direction, but the study did not measure real hiring or visa outcomes.
Several AI systems that listen to speech and grade it gave Chinese and Nigerian accents lower delivery scores than American and British ones, even when every speaker read the exact same script word for word. The models could transcribe the words accurately and still marked down how they sounded.
In a study published in the Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, researchers at the University of Washington and partner institutions tested six audio AI systems, including Google’s Gemini and OpenAI’s GPT-Audio-1.5. Speakers represented five accent groups, American, British, Chinese, Indian and Nigerian English, chosen because English is commonly taught or spoken in each of those regions.
Stakes are real. AI-powered tools already help screen job candidates, who may record answers to automated interview questions and get scored by software before hiring shortlists are drawn up. At least one widely used English test, the kind of exam often required for visas into countries like Canada, Australia and the UK, uses speech AI to score some spoken answers. Its provider alone administered about 1.2 million exams worldwide in 2023, double the 2022 total. Which models power such systems is not publicly known, and to the researchers’ knowledge, no public audits exist.
AI Accent Bias Appeared With Synthetic and Real Human Voices
Researchers built three mock scenarios modeled on real uses of voice AI: an automated job interview, academic presentations, and an English proficiency test. Scripts included a finance-job self-introduction, a client explanation of retirement accounts, and a line chart description.
Most recordings came from synthetic, AI-generated voices, six per accent group pulled from a commercial voice library, with each voice informally checked by someone who has lived experience of that accent. To confirm the results held up with natural speech, the team also recorded 52 volunteers: 13 American, 4 British, 11 Chinese, 10 Indian and 14 Nigerian. Nearly all were young adults between 18 and 34, and they answered both scripted and unscripted hiring questions.
Each recording went to the AI systems, which scored two things on a seven-point scale: content, meaning how good and complete the answer was, and delivery, meaning how clear, confident and fluent the speech sounded. Four of the six gave reliable results, and the other two were set aside as inconsistent.
Lower delivery scores for Chinese and Nigerian speakers held up whether the AI was asked to critique the speaker, compare them with a native English speaker, or compare them with an ideal employee or presenter. English test scenarios produced the widest gaps in the Google and OpenAI models, and in hiring scenarios, real human voices showed the same pattern as synthetic ones.
Less American-Sounding Voices Tended to Score Lower on Delivery
Researchers next measured how far each speaker’s pronunciation, stress and rhythm strayed from American English, using a computer analysis of the speech sounds themselves. British and Indian accents landed close to American English. Chinese and Nigerian accents landed much farther away. Among the real human speakers, the farther a voice sat from American English, the lower its delivery score tended to be, a moderate to fairly strong link in all four models.
A simpler explanation, that the AI could not understand certain accents, did not hold up. Word error rates, which count misheard words, stayed under 4% for every accent group in every model tested.
Written feedback told a similar story. Models rarely mentioned accent outright, but when they did, it was usually for Chinese, Nigerian or Indian speakers, as in a comment about “a distinct non-native accent that occasionally muddies their speech.” American and British speakers were described with words like clear, excellent and professional, while explanations for Chinese and Nigerian speakers more often cited mumbling or unclear articulation.
AI Accent Bias Sometimes Skewed Content Scores, Too
Models were told to ignore delivery when judging content, yet that instruction did not reliably hold. Content scores sometimes drifted in the same direction as delivery scores for Chinese and Nigerian speakers.
One model bucked the trend. In some presentation and English test scenarios, Qwen3-Omni gave Nigerian, Chinese and Indian speakers similar or slightly higher scores than American and British speakers. In hiring scenarios, though, its written feedback grew far more uneven across accents with real human voices than with synthetic ones, possibly because synthetic speech reduces some of the vocal cues that drive the disparities. A model that looks fair in one setting can behave differently in another, so the authors argued that bias testing should focus on how these tools are actually used.
One English test provider cites the goal of “reducing the risk of human bias” for its speech AI. In these experiments, sounding less American cost speakers points even though their words were identical. The researchers could not say why, and they did not measure how often such gaps change real hiring or visa outcomes. Any tool that grades a voice should be shown to treat accents fairly before it grades a job candidate or a visa applicant.
Disclaimer: This article reports on findings from a single research study and is intended for general informational purposes only. It is not a substitute for professional advice, and results may not apply to every person or setting.
Paper Notes
Limitations
Researchers acknowledged that human sample sizes per accent were small (4 British speakers, 10 to 14 for the other groups) and may not represent the full range of variation within each accent. Collecting human recordings takes considerable time, so real voices were tested only in hiring scenarios, and the main experiments used scripted speech, while real-world speech is typically more spontaneous. Synthetic voices may also over- or under-represent certain pronunciation features. Accent groups were defined by national origin, a coarse label the authors say hides substantial variation within each group. The researchers could not tie the acoustic distances to any single pronunciation feature, and they did not establish why the models developed this tendency. Finally, how voice AI is built into real hiring and exam-scoring systems is usually not public, and such systems are often inaccessible to outside researchers, so commercial products were not tested directly.
Funding and Disclosures
Gift funding from Google and OpenAI supported the work in part, and Google’s Gemini models and OpenAI’s GPT-Audio-1.5 were among the systems tested. The authors’ organization determined the study was exempt from Institutional Review Board oversight. Participants gave explicit consent, with separate consent required for inclusion in a controlled-access data repository, and received $6 per session, described as roughly $18 per hour. The authors also disclosed that Claude Code was used to refactor code for their public repository.
Publication Details
Title: The Cost of Sounding Different: Accent Bias in Audio Language Models | Authors: Aruna Srivastava, Orevaoghene Ahia, Li Lucy, Ashley Christendat, Sameep Chattopadhyay, Farhan Samir, Tejumade Afonja, Valentin Hofmann, Sachin Kumar, Noah A. Smith, Yulia Tsvetkov | Affiliations: University of Washington, UW-Madison, University of Toronto, CISPA Helmholtz Center, TRI AI, Munich Center for Machine Learning, LMU Munich, The Ohio State University, Allen Institute for AI | Journal: Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 36660 to 36687, Budapest, Hungary, October 2026 | DOI: https://doi.org/10.18653/2026.emnlp-main.1612







