[AUDITORY] Releasing the test set of FSDKaggle2019 dataset (used in DCASE 2019 Task2)

[Eduardo Fonseca <eduardo.fonseca@xxxxxxx>]

=== Apologies for cross-posting ===

Dear list,

We’re glad to announce we have released the full test set & labels of FSDKaggle2019. This dataset was used for DCASE 2019 Task 2, which was hosted on the Kaggle platform as a competition titled Freesound Audio Tagging 2019.

FSDKaggle2019 includes almost 30k audio clips amounting over 100h of audio, encompassing 80 classes drawn from the AudioSet Ontology. It includes a human curated train set from Freesound (~5k clips, ~11h), a noisy train set from Flickr (~20k clips, ~80h), and a test set from Freesound (~4.5k clips, ~13h). The dataset allows development and evaluation of machine listening methods in conditions of label noise, minimal supervision, and real-world acoustic mismatch.

FSDKaggle2019 is freely available from Zenodo: https://doi.org/10.5281/zenodo.3612637

You can find more details in our DCASE 2019 paper: E. Fonseca, M. Plakal, F. Font, D. P. W. Ellis, and X. Serra. Audio tagging with noisy labels and minimal supervision. Detection and Classification of Acoustic Scenes and Events (DCASE) Workshop, NYC, USA, 2019

Both competition and dataset have been a collaboration between the Music Technology Group of Universitat Pompeu Fabra, and the Sound Understanding team at Google AI Perception. This effort was kindly sponsored by a Google Faculty Research Award 2018.

Best,

Eduardo, Manoj, Frederic, Dan and Xavier

--

Eduardo Fonseca
Music Technology Group
Universitat Pompeu Fabra

--