Open-Set Tagging Dataset (OST)

OST is a synthetic dataset of 1s clips to study multi-label open-set audio classification. The scripts generate an Open-Set Soundscapes dataset (OSS) of 10s soundscapes generated using Scaper. If you are not going to use this dataset, you can choose to synthesize it without audio-- this will synthesize only the JAMS annotation files needed for the 1s clips. Using the JAMS annotation files in OSS, 1s clips in OST are generated deterministically. There are five dataset variants (~17GB each), each with a different combination of classes assigned to the known and unknown class categories. In case of space and/or computational constraints, OST dataset variant 1 is preferred for evaluation. Due to the size of the dataset, we only share the source files as a zip file here. The scripts to generate OSS and OST are available on Github: Multi-label open-set dataset synthesis tool (MOADS)