Researchers at the University of Washington ran a simple test: have the same script read aloud in five different accents, feed the recordings to four AI models, and see if the scores change. They did. American and British accents came out on top, Indian accents landed in the middle, and Chinese and Nigerian accents were ranked lowest, every time, even though the transcripts the models produced were accurate and the content never varied.
The study, in plain terms
The paper, titled "The Cost of Sounding Different: Accent Bias in Audio Language Models," was presented at EMNLP 2026 in Budapest, one of the top conferences in natural language processing, where it won an Outstanding Paper award. It was led by UW doctoral student Orevaoghene Ahia and senior author Yulia Tsvetkov, an associate professor in the Allen School of Computer Science & Engineering, along with co-authors from the University of Toronto, University of Wisconsin-Madison, the CISPA Helmholtz Center, and Koel Labs.
The team tested four audio-capable AI models: Google's Gemini 2.5 Pro and Gemini 2.5 Flash, OpenAI's GPT-Audio 1.5, and the open-source Qwen 3 Omni. They generated speech in American, British, Chinese, Indian, and Nigerian accents using AI voices from ElevenLabs reading identical scripts, then validated the results against recordings from 52 real human speakers across those five countries. Each model scored every recording on a 1 to 7 scale for both delivery and content, the same categories used to judge job interviews, class presentations, and spoken English tests.
What actually changed, and what didn't
The content never changed. The words were the same, and the models transcribed them correctly regardless of accent. What changed was the score. Four models rated American and British accented speakers higher on delivery and content than Chinese and Nigerian accented speakers reading the exact same lines. Indian accented speakers landed somewhere in between.
That detail is what makes this different from a typical speech-recognition bug. The models understood what was said. They just did not rate it the same way once it was said differently.
"They understood the speakers, but they just had this implicit notion of hierarchy among them based on accent," Ahia said.
Where this kind of scoring already shows up
This is not a hypothetical lab exercise. AI voice scoring is already being used, or piloted, in exactly the settings where a biased score has real consequences:
- Hiring platforms that use AI to screen recorded or live video interview responses
- Schools and universities scoring oral presentations or spoken assignments
- English proficiency tests used for immigration and academic admissions
Ahia pointed directly at that overlap: "People are now trusting audio AI models as sort of third-party evaluators, for hiring, for oral exams in education, for certain English tests in immigration, contexts where fairness really matters."
In other words, the bias the researchers measured in a controlled study is the same kind of scoring already deciding who moves forward in a hiring pipeline or who passes a language exam, just without anyone checking the accent breakdown of the results.
Why this keeps happening
The likely causes are not exotic. Training data for these models leans heavily on certain accents being labeled as the clear, default standard, while other accents are underrepresented or get flagged in training sets as harder to understand. A model trained mostly on American and British speech, with little exposure to the full range of Nigerian or Chinese English, ends up treating unfamiliarity as a quality problem instead of just a difference.
Tsvetkov framed the core issue in terms of what gets measured and reported. Companies tend to publish one overall accuracy number for a model, which can look fine on average while hiding a much worse result for specific groups. "We don't want the average performance to be 97% accurate for the general population, but for a smaller group only 20%," she said.
That is the real structural problem: an average score can look great while quietly failing a subset of real users, and nobody notices unless someone specifically goes looking, the way this study did.
What the researchers say would actually fix it
The paper's recommendations are not about banning these models from scoring speech. They are about changing what gets built and what gets reported:
- Train on more representative data sets that include a wider range of accents and dialects, not just the ones that happen to dominate existing audio datasets
- Publish performance broken out by subgroup, not just an overall average, so a company or developer can see if a model quietly underperforms for certain accents before deploying it
- Treat accent fairness as a deployment requirement in high-stakes settings like hiring and testing, not an afterthought discovered after the fact
None of that requires new regulation or a redesign of how these models work. It requires companies deploying them for hiring, education, or testing to actually check for this kind of gap before trusting the score.
The takeaway
The unsettling part of this study isn't that AI models can misjudge accented speech, most people could have guessed that. It's that the models get the words right and still downgrade the speaker, which means the usual fix (better transcription) does nothing here. If you are on the receiving end of an AI-scored interview, presentation, or language test, the accuracy of the transcript tells you nothing about whether the score itself was fair. Until companies start publishing results by subgroup instead of one clean average, there is no way to tell from the outside whether a given tool has this problem or not.