AudioSet
Google Research (Sound Understanding group)
Two million 10-second YouTube clips labeled with 527 kinds of sound, from barking to bagpipes.
Human-verified labels for sound events in 10-second segments of YouTube videos, organized by a hierarchical ontology of human, animal, music, and environmental sounds. Google distributes the labels and precomputed audio features rather than the audio itself. It is the standard benchmark for audio event classification.
Access
- Cost
- FreeFree download of labels and features; audio must be fetched from YouTube.
- License
- Open licenseDataset CC BY 4.0; ontology CC BY-SA 4.0
- Sign-up
- No sign-up
- How you get it
- Bulk downloadCSV, TFRecord, JSON ontology
What's in it
- One row is
- One 10-second YouTube segment with one or more sound labels
- Covers
- YouTube videos worldwide
- Years
- Released 2017; strong (timed) labels added 2021
- Updated
- One-time release
- Size
- 2,084,320 clips; about 5,800 hours; 527 labeled classes in a 632-class ontology
Quality
Large and carefully built, but label quality varies by class and the audio itself is not distributed.
- Many source videos have been deleted
- Music and speech dominate; rare classes have few examples
- Label accuracy varies by class
Connects to
- Kinetics-700: Same 10-second YouTube clip structure, labeled for visual action
- xeno-canto: Expert-identified bird recordings to go deeper than AudioSet's bird classes
Questions it could answer
- Which sound classes co-occur most often in the same clip?
- What share of the evaluation clips can still be downloaded?