UW study finds leading AI models score Chinese and Nigerian accents lowest in interviews and tests
Researchers tested multimodal models from Google, OpenAI and Alibaba on identical scripts read in five accents and found lower delivery and sometimes content scores even when transcription was accurate.
Why it matters
Audio models are increasingly used to screen recorded job interviews and score oral exams and English tests, so a systematic accent penalty has direct consequences for hiring and education. The team compared American, British, Chinese, Indian and Nigerian accents on the same scripts, which isolates accent from content. The models often said Chinese and Nigerian speakers mumbled even though they transcribed them accurately. The work was presented on 25 October at EMNLP in Budapest; watch for responses from the model makers and test providers.
Key facts
- Five accents: American, British, Chinese, Indian, Nigerian
- Models from Google, OpenAI and Alibaba tested
- Settings: job interviews, conference talks, English proficiency tests
- Chinese and Nigerian accents ranked lowest
- Presented 25 October at EMNLP in Budapest