ESC-50 - Imagenette like competition for audio

I want to invite everyone to have some fun participating in a competition to do audio classification, using the ESC-50 dataset. This is a opportunity to practice the concepts learned in the course in a different setting.

ESC-50: Environmental Sound Classification

The ESC-50 dataset is a labeled collection of 2000 environmental audio recordings suitable for benchmarking methods of environmental sound classification


The dataset consists of 5-second-long recordings organized into 50 semantical classes (with 40 examples per class) loosely arranged into 5 major categories


Fastaudio library

Fastaudio is a community contributed extension to fastai to help you build audio machine learning applications while minimizing the need for audio domain expertise. It’s not obligatory to use this library to enter the competition, but strongly recommended.

How to compete?

The repository with the full instructions and leaderboards is here. There, you can also find notebooks with baseline models to help you start competing. When you have a result that you want to submit to the leaderboards, just open a pull request with your results linked on the table.

Feel free to use this post to discuss anything about the competition.


Do we have to submit the final model accuracy (i.e. after all epochs) or can we submit accuracy from the best epoch?

Yes, submit only the final model accuracy. If you have the resources, it’s better to run a couple of times and take the average of the results, to get a more stable value to compare.

1 Like