DESED — recorded subset (redistribution)

This is a redistribution, not the official release. The audio was re-downloaded from the YouTube identifiers in the official DESED metadata, because the official downloader was blocked on our network. It is provided for reproducibility of our experiments, with attribution to the original authors. If you use it, cite the original DESED papers and follow their terms.

Original dataset

• DESED: Domestic Environment Sound Event Detection dataset.
• Authors: Nicolas Turpault, Romain Serizel, Ankit Parag Shah, Justin Salamon.
• Project page: https://project.inria.fr/desed/
• Repository: https://github.com/turpaultn/DESED
• Synthetic set: https://zenodo.org/record/3702397
• Public evaluation set: https://zenodo.org/record/3588172
• Papers:
  • Turpault et al., Sound event detection in domestic environments with weakly labeled data and soundscape synthesis, DCASE 2019.
  • Serizel et al., Sound event detection in synthetic domestic environments, ICASSP 2020.

What is included

Re-fetched from the official metadata (weak.tsv, unlabel_in_domain.tsv, validation.tsv), segmented with the official start/end times and stored as 16 kHz mono PCM16 WAV.

Split	Clips	Labels
audio/train/weak	1,578	clip-level (weak)
audio/train/unlabel_in_domain	14,412	unlabeled
audio/validation	1,168	strong (onset/offset)

Metadata TSVs are included under metadata/. Files that were unavailable on YouTube are listed in missing.tsv; coverage is ~83% of the official list.

Licence and redistribution note

• The original audio comes from AudioSet/YouTube and is subject to the terms of the original DESED release and the underlying recordings.
• This copy is shared for non-commercial research only. It does not grant broader rights than the original sources.
• No authorship of the recordings is claimed. All credit belongs to the DESED authors and the original recordists.
• If the original authors object to this redistribution, the repository will be taken down.

Intended use

Training and evaluating streaming sound event detection on the recorded DESED subsets. Companion synthetic data and the public evaluation set are in the official Zenodo releases linked above.

Citation

@inproceedings{turpault2019desed,
  title     = {Sound event detection in domestic environments with weakly labeled data and soundscape synthesis},
  author    = {Turpault, Nicolas and Serizel, Romain and Shah, Ankit Parag and Salamon, Justin},
  booktitle = {Workshop on Detection and Classification of Acoustic Scenes and Events (DCASE)},
  year      = {2019}
}