FSD50K
Fonseca, Favory, Pons, et al. (Universitat Pompeu Fabra)
51,197 Freesound clips labeled with 200 sound-event classes drawn from the AudioSet Ontology.
An open set of human-labeled sound events: 51,197 Freesound audio clips spread unevenly over 200 classes from the AudioSet Ontology, totaling about 108 hours. Labels were gathered with crowd annotation and validated. It is used to train and benchmark audio tagging and sound event detection.
Access
- Cost
- FreeFree download from Zenodo; no account needed.
- License
- Provider's own termsPer-clip Creative Commons licenses (CC0, CC BY, CC BY-NC, and CC Sampling+) plus dataset terms
- Sign-up
- No sign-up
- How you get it
- Bulk downloadWAV, ZIP (split archive)
What's in it
- One row is
- One audio clip with one or more sound-event labels
- Covers
- Not geographic
- Years
- Clips from Freesound (release 2020)
- Updated
- One-time release
- Size
- 51,197 clips in 200 classes
Quality
Well documented with a published paper and standard train and test splits; static since 2020.
- Mixed clip licenses; check each clip's terms for commercial use
- Strong class imbalance
Connects to
- AudioSet: Larger weakly labeled YouTube set that shares the ontology
- Freesound: Source repository of the audio clips
- FSDKaggle2018: Earlier smaller subset from the same project
Questions it could answer
- Which sound classes are hardest to tag?
- How does a model pretrained on AudioSet do on FSD50K labels?