Cocktail-party audio as a reverse Turing test: humans beat speech models, but not by enough [P]

Overview

Hello - I wanted to share a small project I worked on test whether picking one voice out of background chatter could separate humans from machines, in the spirit of an audio CAPTCHA. Setup 40 synthetic clips: one target sentence read by a synthetic voice, mixed over multiple background voices reading random short fragments (staggered starts, 4 dB attenuation) 146 human transcriptions from 18 sessions, unlimited replays allowed 8 model configurations, one pass per clip: Whisper base/small, Vosk,

SpinGraph analysis pending — check back after processing.