Open-Set Tagging Dataset (OST)

<p>Open-set Tagging (OST) is a synthetic dataset of 1s clips used to evaluate source-centric representation learning models in the paper&nbsp;<a href="https://ieeexplore.ieee.org/document/10890242">Compositional Audio Representation Learning</a>.</p> <p>Due to the size of the dataset, we only share the source files, and provide the scripts to generate the dataset are available <a href="https://github.com/sripathisridhar/moads">here.</a><br><br>The dataset generation process is as follows:<br>1. From single-source FSD50K audio files, we generate a dataset of 10s soundscapes called Open-set Soundscapes (OSS) using <a href="https://github.com/justinsalamon/scaper">Scaper</a>.</p> <p>2. We then center a 1s window around the center of each sound event in the 10s soundscapes to generate Open-set Tagging (OST), which contains ~500k clips.&nbsp;</p> <p>If you are not going to use OSS, you can choose to synthesize it without audio-- this will synthesize only the <a href="https://github.com/marl/jams">JAMS</a> annotation files needed for the 1s clips. Using the OSS JAMS files, OST clips can be generated deterministically.</p> <p>There are five dataset variants (~17GB each), each with a different random assignment of classes to the known and unknown class categories. For further details, refer to our previous paper&nbsp;<a href="https://dcase.community/documents/workshop2023/proceedings/DCASE2023Workshop_Sridhar_11.pdf">Multi-label open-set audio classification</a>. In <a href="https://ieeexplore.ieee.org/document/10890242">this work</a>, OST dataset variant 1 is referred to as OST for simplicity.&nbsp;<br><br>We also introduce a tiny version of the dataset called OST-Tiny, which contains ~20k clips and only 10 known classes. This is convenient for faster prototyping and to evaluate models in a more challenging open-set classification scenario.</p> <p>&nbsp;</p>