Research
Study finds humans beat speech models at picking one voice from cocktail-party audio, but the gap is narrow
Across 40 synthetic clips, humans averaged 0.71 word accuracy versus 0.54 for instructed Gemini, a gap too small to make a reliable audio CAPTCHA. The sample is small, with 18 sessions and a single synthetic setup.
Every outlet that ran it 1 article
r/MachineLearningCommunityLead
Cocktail-party audio as a reverse Turing test: humans beat speech models, but not by enough [P]