Research
Study finds frontier audio models struggle to tag physical ASMR triggers
Eight judges, including Gemini 3.1 Pro at F1 0.325 against human labels, scored poorly on physical ASMR sounds like brushing and paper, while whispering was detected reasonably well. The author is building a human-labelled ASMR dataset and a CLAP-based student model.
Every outlet that ran it 1 article
r/MachineLearningCommunityLead
Frontier models are bad at hearing ASMR: 8 judges vs human labels on 480 clips [P]