VQ-Bench study shows speech AI judges salary, leadership by voice quality

A new study from KTH Royal Institute of Technology, led by Harm Lameris and colleagues, tackles a blind spot in speech AI: how models respond to non-lexical voice cues. The paper, accepted at Interspeech 2026 and posted on arXiv (2510.25577), introduces VQ-Bench, a controlled evaluation suite with a parallel dataset of synthesized modal, breathy, creaky, and end-creak phonation types. The researchers tested Speech Foundation Models (SFMs) through open-ended generation across four ecologically valid domains, plus speech emotion recognition, to measure sensitivity to paralinguistic variation that raw audio models are now able to process.

The results are concerning. A leading commercial API failed basic biometric sanity checks, suggesting it cannot reliably distinguish between speakers or voice qualities. Other models showed systematic shifts in perceived agency, empathy, and leadership based solely on phonation type. Crucially, the study found gender asymmetries in salary and leadership endorsements: the same content delivered with different voice qualities led models to recommend different compensation and leadership potential depending on gendered vocal cues. This demonstrates that SFMs may mirror or amplify human social biases, raising the stakes for deploying speech-based AI in hiring, virtual assistants, and other high-stakes contexts. The VQ-Bench framework provides a reproducible way to audit these behaviors, pushing the field toward responsible paralinguistic interpretation and more equitable speech technology.