All projects

Perceptual and Acoustic Correlates of Race in TTS

Read the Interspeech 2026 paper (PDF) · DOI: 10.21437/Interspeech.2026-1479

Code and stimuli available here: GitHub

Companies are beginning to claim that they can encode race in their synthetic voices. We conducted a large-scale perceptual study to test whether listeners actually hear those intended racial identities.

We found that listeners perceived race more consistently in male synthetic voices. We also found that when users did reliably perceive race, they often applied racial stereotypes to those voices. Lastly, we found that perceived race could be effectively parameterized using a low-dimensional set of spectral features.

Listen for yourself

African American synthetic voice

White synthetic voice

Both synthetic voices are male and read the same sentence.

Mean personality ratings for female and male synthetic voices grouped by perceptually assigned race
Personality ratings by perceptually assigned race and gender.
Bar chart ranking spectral features used to classify the perceived race of synthetic voices
Top acoustic features associated with listeners’ perception of race.

I collaborated on this project with Nicole Holliday (UC Berkeley) and Sarah Creel (UC San Diego).