Outstanding Paper Award at EMNLP 2026 for Work on Accent Bias in Audio Language Models

The paper “The Cost of Sounding Different: Accent Bias in Audio Language Models” received an Outstanding Paper Award at the Conference on Empirical Methods in Natural Language Processing (EMNLP 2026) in Budapest, Hungary. The paper was written by MCML PI Valentin Hofmann and co-authors Aruna Srivastava, Orevaoghene Ahia, Li Lucy, Ashley Christendat, Sameep Chattopadhyay, Farhan Samir, Tejumade Afonja, Sachin Kumar, Noah A. Smith, and Yulia Tsvetkov.

Audio language models are increasingly used in high-stakes settings such as hiring, oral assessments, and language proficiency tests. The authors examined how six audio language models assess English speakers with American, British, Chinese, Indian, and Nigerian accents. They found that several models gave lower speech delivery scores to Chinese- and Nigerian-accented speakers, regardless of the quality of what was said. Scores also tended to drop the more a speaker’s accent differed acoustically from American English, suggesting that the models judge speech against an implicit American norm. Even though the models transcribed the speech correctly, their assessments could disadvantage speakers in real-world decisions. The authors release their framework and data to support further research on bias in audio language models.

Congratulations to the authors on this recognition!