Datasets:

schism-audio
/

gise-51

license: cc-by-4.0
task_categories:
  - audio-classification
language:
  - en
pretty_name: GISE-51
tags:
  - audio
  - sound-event-classification
  - environmental-sound
  - sound-event-detection
  - scaper
  - fsd50k
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path:
          - isolated_events/train/metadata.csv
          - isolated_events/train/**/*.wav
      - split: validation
        path:
          - isolated_events/val/metadata.csv
          - isolated_events/val/**/*.wav
      - split: test
        path:
          - isolated_events/eval/metadata.csv
          - isolated_events/eval/**/*.wav
    drop_labels: false

GISE-51

This is a Hugging Face mirror of the core GISE-51 isolated sound event dataset from Zenodo record 10.5281/zenodo.4593514.

GISE-51 contains 51 isolated sound event classes derived from FSD50K. The upstream release also provides synthetic GISE-51-Mixtures assets. This mirror includes the extracted isolated events, label metadata, noise files, the original JAMS annotation archive for regenerating mixtures, the dataset license file, and silence_thresholds.txt.

Included Files

isolated_events/: 16,357 mono WAV files, organized as train, val, and eval splits by class.
metadata.csv: root-level metadata for all isolated-event WAV files.
isolated_events/{train,val,eval}/metadata.csv: split-local metadata used by the Hub dataset viewer.
meta/lbl_map.json: 51-class label map from the upstream release.
noises/: background noise files used for mixture generation.
mixtures_jams.tar.gz: original JAMS annotation archive for the mixture subsets.
license/LICENSE-DATASET: upstream CC-BY-4.0 license text.
silence_thresholds.txt: upstream silence filtering thresholds.

The very large pre-rendered mixture audio tarballs and pretrained model archive from the Zenodo record are not expanded in this mirror. The uploaded mixtures_jams.tar.gz, noises/, and isolated_events/ assets are the compact, reproducible ingredients for the mixture soundscapes.

Statistics

Item	Value
Isolated event files	16,357
Classes	51
Train files	12,465
Validation files	1,716
Test files	2,176
Total isolated-event duration	18.45 hours
Audio format	WAV, PCM 16-bit
Sample rate	22,050 Hz
Channels	Mono

Mixture JAMS annotations included:

Subset	JAMS files
train	60,000
val	10,000
eval	10,000
train_30k	30,000
train_40k	40,000
train_50k	50,000
train_p2	10,000
train_p3	10,000
train_p4	10,000
train_p5	10,000

Source And Checksums

Source record: Zenodo 4593514

Verified upstream MD5 checksums:

isolated_events.tar.gz: c68db0b56d5cf036ee6811eece327a09
meta.tar.gz: 404fbe07e9c429139a9ac2b55ee8931d
license.tar.gz: 945b211c67babb7eae5c4c0e7cb18794
noises.tar.gz: 326eb670b9ebb94dd9e305375b04692b
mixtures_jams.tar.gz: 137e24af27e2950046fc545a0008e40c
silence_thresholds.txt: 6824cb74b7973036e3c991334b6c047b

Citation

@article{yadav2021gise51,
  title = {GISE-51: A scalable isolated sound events dataset},
  author = {Yadav, Sarthak and Foster, Mary Ellen},
  journal = {arXiv preprint arXiv:2103.12306},
  year = {2021}
}

GISE-51 is based on FSD50K; please also cite FSD50K when using this dataset.