Skip to content

fastaudio/Audio-Competition

Repository files navigation

Audio-Competition

Small datasets to test novel audio algorithms, heavily influenced by imagenette

The Datasets

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:

ESC-50 is a really nice starting dataset as it is especially clean (fixed-length, hand-labeled, single sample-rate) and well maintained. Many thanks to Karol Piczak for maintaining a really great Github Repo based around this dataset.

Usage

If you are already using the fastaudio library, you can download and access these quickly with commands like:

path = untar_data(URLs.ESC50)

where path now stores the destination to ESC-50.

Leaderboard

Generally you'll see +/- 1% differences from run to run since it's quite a small validation set. So please only send in contributions that are higher than the reported accuracy >80% of the time. Here's the rules:

  • No inference time tricks, e.g. no: TTA
  • Must be one of the split/#epoch combinations listed in the table
  • If you have the resources to do so, try to get an average of 5 runs, to get a stable comparison. Use the "# Runs" column to include this
  • In the URL column include a link to a notebook, blog post, gist, or similar which explains what you did to get your result, and includes the code you used (or a link to it), including the exact commit, so that others can reproduce your result.

ESC-50 Fold 1 Leaderboard

  • For this leaderbord, use only the data corresponding to Fold 1 to validate, and train on the other 4 folds.
Epochs URL Accuracy # Runs
80 fastaudio baseline w/ mixup 78.25% 1
20 fastaudio baseline 66.64% 5, mean
10 fastaudio baseline 62.69% 5, mean

ESC-50 All Folds Leaderboard

  • For this leaderbord, do 5-fold cross validation using the splits defined in the metadata, and report the mean accuracy.

  • The results here are also comparable with the official dataset leaderbods

Epochs URL Accuracy # Runs
80 no entries yet -- --
20 fastaudio baseline 67.35% 1
10 fastaudio baseline 64.94% 1

About

Small datasets to test novel audio algorithms

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published