🎯 "Walk a Mile in My Voice”: Voice Conversion Shapes Trust, Attribution, and Empathy in Human–AI Speech Interactions
Exploring intersectional bias in Speech LLMs across accents and gender
🔬 Participate in Our Research Studies
We're investigating how voice characteristics influence AI system responses. Join our studies to contribute to important research on fairness and bias in AI systems.
📁 View Study Results & Data
🎙️ Voice Adaptive AI Study - Voice Conversion
Experience voice conversion technology firsthand!
In this study, you will:
- Record yourself asking an AI assistant 4 conversational questions
- Hear your voice converted to different gender/accent profiles
- Compare how the AI responds to your original voice vs. converted voices
- Reflect on differences you notice in the AI's behavior
📋 Requirements:
- Computer with working microphone (recording deleted after exit)
- Quiet environment • English proficiency
- Chrome, Firefox, or Safari browser
🖥️ Desktop
📱 Tablet
🔊 Audio
🎤 Microphone
🚀 Start Voice Study
🎧 Voice Adaptive AI Study - Listening Task
Evaluate AI conversations with different voices!
In this study, you will:
- Listen to 4 pre-recorded conversations between people and an AI
- Hear speakers with different genders and accents
- Rate the AI's responses on harm, acceptability, and trust
- Share observations about potential differences in AI responses
📋 Requirements:
- Computer with speakers or headphones
- 15-20 minutes of uninterrupted time
- No recording needed – just listening and evaluation
- Chrome, Firefox, or Safari browser
🖥️ Desktop
📱 Tablet
🔊 Audio
🚀 Start Listening Study
📧 Questions? Contact: shbs@kth.se
📊 The Speech Test Suite - SpeechLLM Responses
The Speech Test suite, SpeechLLM responses can be found below.
Select a Question Type to Explore:
🔍 Understanding the Visualizations
PCA Plots: Each visualization shows 4 subplots examining different aspects of bias:
- Accent (with Gender): Colors = accents, Shapes = gender (★ female, ● male)
- Model: Compare responses across different AFMs
- Gender (with Hesitation): Colors = gender, Shapes = hesitation (▲ with, ● without)
- Hesitation: Direct comparison of hesitation patterns