From Seeing it to Experiencing it: Interactive Evaluation of Intersectional Voice Bias in Human-AI Speech Interaction
A novel study lets users experience AI bias firsthand by swapping their voice to different accents and genders.
A research team led by Shree Harsha Bokkahalli Satish, with collaborators from the University of Edinburgh and Apple, has published a novel study on arXiv that tackles a critical blind spot in AI evaluation: how bias manifests in real human-AI speech interactions. Their paper, 'From Seeing it to Experiencing it: Interactive Evaluation of Intersectional Voice Bias in Human-AI Speech Interaction,' moves beyond traditional automated metrics to probe how SpeechLLMs—AI models that process spoken language directly from audio—treat users differently based on vocal cues like accent and perceived gender.
The researchers' methodology is two-fold. First, they conducted a controlled, automated analysis using a test cohort spanning six different accents and two gender presentations, measuring disparities in AI response quality (like off-topic replies) and content. Second, and most innovatively, they designed an interactive study where 24 participants used voice conversion technology to experience how identical queries were processed when their voice was synthetically altered to sound like different accents and genders. This 'perspective-taking' experiment revealed that voice conversion significantly increased user trust in benign AI responses and fostered a deeper understanding of the bias issue.
The findings were stark. Automated analysis uncovered clear intersectional bias—specific combinations of accent and gender led to measurable disparities in how well the AI aligned with user requests and the verbosity of its responses. The study highlights that current bias evaluations often miss these nuanced, experiential dimensions of discrimination in end-to-end speech interactions. By combining quantitative metrics with qualitative, user-centered experience, the team provides a more comprehensive evaluation suite for developers building spoken conversational AI, pushing the field toward systems that are fair and equitable for all users, not just those with 'standard' accents.
- The study used voice conversion tech to let 24 users experience AI bias firsthand by swapping their vocal identity across 6 accents and 2 gender presentations.
- Automated analysis of SpeechLLMs revealed measurable 'accent x gender' disparities in response alignment and verbosity, showing intersectional bias.
- The interactive method increased user trust in AI responses and encouraged perspective-taking, offering a richer evaluation framework than metrics alone.
Why It Matters
As voice AI becomes ubiquitous, this research provides tools to build fairer systems by exposing and quantifying real-world, experiential bias that standard tests miss.