Voice Quality Biases in Speech Foundation Models

Evaluation via VQ-Bench reveals that speech foundation models exhibit systematic shifts in perceived agency, empathy, and leadership based on phonation types like breathy or creaky voice. Results show that these models can mirror or amplify human social biases, including gender-based asymmetries in leadership endorsements.