The database consists of audio recordings, each ten seconds long. Each audio file contains between two and four acoustic events that do not overlap. In addition, various background noises are present in the recordings. The possible acoustic events (classes) include, among others: Shatter, Bark, Doorbell, Shout, Cough, Camera, Church_bell, Scratching (performance technique), Fireworks, Burping and eructation, Meow und Cheering.

For each audio file, there is an annotation file that contains the exact start and end times of the events as well as their respective classes. In total, the dataset comprises about 36 hours of audio material, with 28 hours designated for training and 8 hours for testing purposes.