A large joint sound scene and sound event dataset for source separation of foreground sound events

This large scale data set contains 10000 samples of sound scenes generated from real world recordings, and the original source recordings. It includes 10 different backgrounds with 6-9 appropriate foreground sound events. Strong labels (timed annotations) are provided in four formats for all samples. The original sourceids to identify the class type, a two source method to simply separate foreground and backgrounds, a 32 source annotation for all distinct foregrounds, and a by background (scene) type annotation where sources are according to the background. 

Baseline results will be presented later in 2020. Further evolutions of this dataset will also be produced with more complex, polyphonic foreground sound events. Please email h.bear@qmul.ac.uk with any questions.

Data is free to use for Research purposes only.