AI xPress

AI and computer science research.

Back to Foragentis

Speech & Audio

Five automated speech recognition systems made more transcription errors for Black speakers than for white speakers

Speech & Audio November 15, 2026

A field recorder on a tripod in a forest

In a 2020 paper in Proceedings of the National Academy of Sciences, researchers at Stanford University and Georgetown University tested five automated speech recognition systems. The systems came from Amazon, Apple, Google, IBM, and Microsoft. They transcribed structured interviews with 42 white speakers and 73 Black speakers in five US cities. The audio totaled 19.8 hours and matched speakers by age and gender. The average word error rate, the share of words a system transcribes incorrectly, was 0.35 for Black speakers and 0.19 for white speakers. All five systems showed the gap.

How large was the racial disparity in speech recognition accuracy?

The average word error rate was 0.35 for Black speakers and 0.19 for white speakers. That means the systems made more transcription mistakes when the speaker was Black. The gap appeared in all five systems.

How did acoustics contribute to the speech recognition gap?

The authors traced the disparity to the acoustic models that the systems use. The gap was just as large when Black and white speakers said identical phrases. That finding pointed to the sound-processing parts of the systems rather than to the words alone.

How is speech recognition different from speech synthesis?

Speech recognition converts spoken language to text. Speech synthesis creates spoken language from text. This study measured errors in speech recognition.

What did the authors suggest to reduce the gap?

The authors proposed using more diverse training datasets, including African American Vernacular English. The abstract does not report results from testing that approach.

Limits and what's next

The abstract describes a study of five speech recognition systems on structured interviews with 42 white speakers and 73 Black speakers in five US cities. It reports 19.8 hours of audio matched on age and gender. The abstract reports no independent replication, so this is one study. The authors propose using more diverse training datasets, including African American Vernacular English, but the abstract does not report testing that strategy.

Source: Allison Koenecke et al., “Racial disparities in automated speech recognition,” Proceedings of the National Academy of Sciences, 2020.