INTERNATIONAL EDITION
INDEPENDENT WORLD NEWS
Technology

Study finds AI speech evaluators favor some English accents

University of Washington researchers found four AI models tended to score American and British English accents above Chinese and Nigerian accents, even for matching words.

View of 2024 photograph of microphone and headphone use in a recording studio; not participants in the study.
Contextual 2024 photograph of microphone and headphone use in a recording studio; not participants in the study.
Photo credit and license Photo: Asikironalio / Wikimedia Commons / CC BY 4.0; contextual studio photograph from 2024, not research participants. · Image source · License: CC BY 4.0
Text size

Artificial intelligence systems used to assess speech can produce unequal results for people saying the same words with different accents, according to research described by the University of Washington on November 4. The findings raise questions about using automated scores in hiring, presentations and language assessments. They do not prove that every AI model performs in the same way.

Researchers examined four models: Google Gemini 2.5 Pro and Gemini 2.5 Flash, OpenAI GPT-Audio 1.5 and the open-source Qwen 3 Omni. The study compared voices speaking English with American, British, Chinese, Indian and Nigerian accents. Across evaluated tasks, American and British accents generally received higher scores, Chinese and Nigerian accents lower scores, with Indian accents typically between them.

To isolate how accents influenced ratings, the researchers had synthetic voices read identical passages. They also tested the approach with 52 human speakers from the five countries. The systems scored delivery and content, not merely whether they could transcribe the words. Researchers reported that the models often understood the spoken material accurately even when they rated some accents as less clear or professional.

The work was presented on October 25 at the Conference on Empirical Methods in Natural Language Processing in Budapest. Its authors call for more representative training and evaluation data and for performance to be reported by speaker groups rather than only as an overall average. The study has limits: it examined selected languages, accents and models under specific tasks. Still, it points to a practical risk when an automated score can affect access to a job or educational opportunity. Decisions with serious consequences require scrutiny beyond a single numeric rating.