About & credits
VoxParity is an open research benchmark. Each caller line is heard in two different deliveries with the same words; the right action depends on how it sounds. People and AI systems face the same calls and the same action menus, and are scored the same way.
What we collect
- Your answers (action, typed details, and how you think the caller sounds).
- Timings (how long each choice took) and how many times you played each clip.
- A random player id and session id stored in your browser, your 18+ confirmation, and an optional age bracket.
- No name, email, IP address or audio. Nothing is recorded from your microphone.
Anonymous answers may be released as part of an open research dataset.
Audio credits
- All caller voices are synthetic or public-domain recordings. No child's voice is real.
- DEMAND: J. Thiemann, N. Ito, E. Vincent, 'DEMAND: a collection of multi-channel recordings of acoustic noise in diverse environments' (scene STRAFFIC, ch01), Zenodo, https://doi.org/10.5281/zenodo.1227121. Licensed CC BY 4.0. (CC-BY-4.0)
- Most voices rendered with Google Gemini text-to-speech.
- Some voices rendered locally with Kokoro-82M (hexgrad, Apache-2.0).