Study finds Gemini and GPT audio models rate Chinese and Nigerian accents lower

On November 5, 2026, Solidot covered a study examining how advanced multimodal models evaluate speech in job interviews, conference presentations, and English speaking tests. The models tested included Google's Gemini 2.5 Pro and Gemini 2.5 Flash, OpenAI's GPT-Audio 1.5, and the open-weights model Qwen 3 Omni. Researchers asked these models to score English speech produced with five accents: American, British, Chinese, Indian, and Nigerian. The results showed that the models systematically rated the Chinese and Nigerian accents lowest, with the Indian accent in the middle, while the American and British accents received the highest scores. Even when the spoken content was identical, the systems still judged the American and British accents as superior in both expression and content. The research team also found that this clear hierarchy carried over into the models' explanations. The models repeatedly described Chinese and Nigerian English speakers as mumbling or hard to understand. The researchers point out, however, that the models are actually very good at transcribing accented speech and understand the speakers well. Their problem, the team argues, is that they hold a tacit hierarchy about speakers based on accent alone. As of this report, no further public developments have followed.