Modern artificial intelligence systems that analyze spoken language may be biased against people with certain accents. That is the conclusion of a group of researchers led by the University of Washington (USA). They tested how several advanced AI models evaluate answers in job interviews, at conferences and on English language tests.
The experiment involved four models: two from Google (Gemini 2.5 Pro and Gemini 2.5 Flash), one from OpenAI (GPT-Audio 1.5) and one open model (Qwen 3 Omni). All of them can work with both text and sound - such models are called multimodal, meaning they can process different types of information. The researchers had them listen to the same texts spoken in voices with five different accents: American, British, Chinese, Indian and Nigerian. For the sake of experimental rigor, synthesized voices from ElevenLabs were used, and then the results were checked with the participation of 52 live speakers of these accents.
It turned out that the models generally gave higher scores to American and British pronunciation, and the lowest to Chinese and Nigerian. The Indian accent landed somewhere in the middle. Moreover, this difference persisted even when the content of the speech was identical and the models correctly recognized all the words. That is, the issue is not that they had difficulty making out the speech - they heard and understood what was said, but still rated speakers of some accents lower than others.
The researchers also measured the so-called "acoustic distance" - how much a pronunciation differs from the American reference. It turned out that the farther an accent is from American, the lower the score usually was. Moreover, in their explanations the models sometimes called speech with a Chinese or Nigerian accent "unintelligible" or "difficult to understand," even though the transcription (conversion of speech to text) was in fact accurate.
The work was presented on October 25 at the Conference on Empirical Methods in Natural Language Processing (EMNLP) in Budapest (Hungary). The authors emphasize that bias may arise from training data in which some accents are underrepresented, or from models absorbing stereotypical ideas about "correct" pronunciation. For example, English is the official language in Nigeria and the language of instruction in schools in India, but the models still rated these accents lower than British and American.
The scientists note that this is not a verdict but a reason for improvement. They plan to expand the study by adding more human accents, including regional variants within countries, and to study how AI analyzes speech in medical scenarios, for example in automated handling of emergency calls. According to the authors, developers should publish model evaluations not only by average metrics but also broken down by groups, dialects and accents - so that high average accuracy does not hide failures in individual subgroups.