Environmental Sound Classification Dataset With 50 Classes

ESC-50 is a labeled collection of 2,000 environmental audio recordings. It contains 50 distinct sound classes, each with 40 examples, created by K. J. Piczak. The dataset was published for the 23rd ACM Multimedia Conference in 2015.

Use Cases

• Train audio classifiers to identify sound categories like 'dog bark' or 'helicopter' from 5-second audio clips.
• Benchmark model performance on the predefined 50-class classification task using the provided fold indices.
• Perform audio feature extraction and analysis on recordings sampled at 44.1 kHz.
• Study transfer learning by using the pre-labeled sounds as a source dataset for other audio tasks.

Strengths

• Contains 2,000 individual audio recordings.
• Provides a standardized benchmark with 50 distinct sound classes.
• Includes pre-defined 5-fold cross-validation splits for consistent evaluation.

Limitations

• Relatively small scale with only 40 examples per class.
• Audio clips are short, fixed at 5 seconds each.
• Dataset is from 2015 and may not reflect contemporary sound environments.

Provenance

Source

K. J. Piczak, Proceedings of the 23rd Annual ACM Conference on Multimedia, 2015.

Collection Method

Curated collection of environmental sound recordings.

Freshness

The core dataset was published in 2015; a platform entry was updated in January 2023.