A reverse Turing test for the agentic AI era, using a skill your brain already has. Crosstalk is a CAPTCHA built on selective auditory attention - the trick your brain uses to track one voice in a crowded room. Humans do it without thinking; AI agents can't.
How it works
No distorted letters, no "select all the crosswalks." Just a sentence, spoken once, in a room full of other people talking.
We generate a fresh clip. One voice reads a sentence aloud. Several others talk over it — real background chatter, not static.
Your visitor listens once and types what they heard. No puzzle-solving, no second-guessing which images have a bus in them.
We verify it server-side, in real time. A drop-in API call — pass or fail, same shape as any CAPTCHA check you've integrated before.
How it sounds
These are real, unedited outputs from our pilot. Press play, then reveal actual AI transcriptions of the same clip.
How it looks
We ran a corpus of audio clips through multiple speech-recognition models spanning different architectures, from lightweight edge models to large multimodal transformers. None came close to human performance.
Early access
We're testing this with a small group before general availability. A couple quick questions help us figure out where it's actually useful.