A manually curated dataset, labeled by 7 experts across 11 recordings. Each recording was annotated by 2 to 5 people, with an agreement rate of at least 80% among individual curators.