ARCA23K

This is the software used to create the ARCA23K and ARCA23K-FSD datasets. A description of these datasets, along with download links, can be found on the Zenodo page. Details of how the datasets were created can be found in our DCASE2021 paper¹.

Due to the mutable nature of the Freesound database (the source of the audio data), this software is unlikely to reproduce ARCA23K and ARCA23K-FSD exactly. Nevertheless, we hope this code can serve as a useful reference.

The source code for the baseline system can be found here.

Requirements

This software requires Python >=3.8. To install the dependencies, run:

poetry install

or:

pip install -r requirements.txt

You are also free to use another package manager (e.g. Conda).

FFmpeg is required too for converting the audio files.

Configuration

Some of the scripts require access to the Freesound API. To use the API, access credentials are required, which can be applied for here. Once a client ID and a client secret key are obtained, they need to be added to the client.json file. An access token is also needed to download clips from Freesound. To obtain an access token, follow the instructions given here. Note that an access token is only valid for 24 hours. To use the API without request limitations, you may need to contact the Freesound developers.

Usage

Each Python script has the following usage:

python SCRIPT [--work_dir DIR] [other-options...]

The --work_dir option is used to specify the directory in which the output files are to be written. By default, it is set to _output/. As some scripts depend on the output files of other scripts, please ensure that this option is set to the same value across scripts.

For details of the other arguments and options, use --help.

This software provides six scripts:

1. src/create_fsd50k_subset.py: Creates a tentative version of ARCA23K-FSD. It selects a single-label subset of FSD50K and saves the ground truth data of the subset.
2. src/query_freesound.py: Uses the Freesound API to search the database for all clips that are up to 30 seconds in duration. The search results, which include various metadata, are saved to disk.
3. src/retrieve.py: Based on the search results of the previous script, the results are narrowed down to clips that can be assigned a label, which is determined using an automated procedure.
4. src/download_clips.py: Uses the Freesound API to download clips from Freesound. The clips that are downloaded depend on the results of the previous script.
5. src/convert_audio.py: Converts the downloaded Freesound clips to mono 16-bit 44.1 kHz WAV files.
6. src/curate_datasets.py: Creates the final ground truth data for ARCA23K and ARCA23K-FSD.

Ensure that the scripts are run in the given order.

Attribution

src/extern/freesound.py is from MTG/freesound-datasets.

metadata/ontology.json is from audioset/ontology.

Citing

If you wish to cite this work, please cite the following paper:

BibTeX:

@inproceedings{Iqbal2021,
    author = {Iqbal, T. and Cao, Y. and Bailey, A. and Plumbley, M. D. and Wang, W.},
    title = {{ARCA23K}: An audio dataset for investigating open-set label noise},
    booktitle = {Proceedings of the Detection and Classification of Acoustic Scenes and Events 2021 Workshop (DCASE2021)},
    pages = {201--205},
    year = {2021},
    address = {Barcelona, Spain},
}

---

1. T. Iqbal, Y. Cao, A. Bailey, M. D. Plumbley, and W. Wang, “ARCA23K: An audio dataset for investigating open-set label noise”, in Proceedings of the Detection and Classification of Acoustic Scenes and Events 2021 Workshop (DCASE2021), 2021, Barcelona, Spain, pp. 201–205.↩