ESC: Environmental Sound Classification Dataset with 50 Classes

2,000 labeled 5-second audio clips comprise the ESC-50 dataset, organized into 50 classes with 40 clips each. It was created by Karol J. Piczak of Warsaw University of Technology from public field recordings on Freesound.org. The collection also includes a 10-class subset (ESC-10) and a larger unlabeled set (ESC-US) of 250,000 clips for unsupervised learning.

Use Cases

• Train supervised audio classification models based on the 50 labeled environmental sound categories.
• Benchmark model performance using the provided 5 uniformly sized folds for cross-validation.
• Conduct unsupervised or self-supervised learning experiments based on the 250,000 unlabeled clips in ESC-US.
• Explore weakly-supervised learning techniques based on the user-submitted tags accompanying the unlabeled data.

Strengths

• Provides 2,000 pre-labeled, standardized audio clips (5 seconds, 44.1 kHz, single channel).
• Includes a large unlabeled set of 250,000 clips for pre-training.
• Data is structured into 5 folds to prevent data leakage from the same source recording.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Last update date is unknown; freshness unverified.

Provenance

Source

Freesound.org project

Collection Method

Clips extracted from public field recordings available on Freesound.

Time Range

null

Freshness

Last updated date is unknown.

Geography

null