benjamin-paine's picture
Update README.md
fab1ff1 verified
metadata
dataset_info:
  features:
    - name: audio
      dtype: audio
    - name: username
      dtype: string
    - name: freesound_id
      dtype: uint64
    - name: license
      dtype:
        class_label:
          names:
            '0': CC0-1.0
            '1': CC-BY-4.0
            '2': CC-BY-3.0
            '3': CC-BY-NC-3.0
            '4': CC-BY-NC-4.0
            '5': CC-Sampling+
    - name: attribution_required
      dtype:
        class_label:
          names:
            '0': 'No'
            '1': 'Yes'
    - name: commercial_use
      dtype:
        class_label:
          names:
            '0': 'No'
            '1': 'Yes'
  splits:
    - name: train
      num_bytes: 4824879667
      num_examples: 50000
  download_size: 4638574544
  dataset_size: 4824879667
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
license: cc-by-4.0
task_categories:
  - audio-to-audio
  - audio-classification
tags:
  - freesound
  - freesound.org
  - laion
  - laion-audio
pretty_name: FreeSound.org LAION-640k Dataset (16 KHz, Small)

About this Repository

This repository is the training split of the complete FreeSound LAION 640k dataset, limited only to licenses that permit commercial works, resampled to 16khz using torchaudio.transforms.Resample.

This is ideal for use cases where a variety of audio is desired but fidelity and labels are unnecessary, such as background audio for augmenting other datasets.

Dataset Versions

  • The full dataset contains 403,146 unique sounds totaling 37.5 GB.
  • The large dataset contains 200,000 unique sounds totaling 18.7 GB.
  • The medium dataset contains 100,000 unique sounds totaling 9.29 GB.
  • You are looking at the small dataset which contains 50,000 unique sounds totaling 4.64 GB.
  • The tiny dataset contains 20,000 unique sounds totaling 1.84 GB.

Sampling Method

To generate the smaller datasets, the following method was applied:

  • Using the complete dataset's tag metadata, generate a feature vector for each sample.
  • Cluster the feature vectors using k-means sampling.
  • Sample from each cluster in a round-robin fasion until the desired dataset size is reached.

What about download links?

Links were ommitted for the sake of size, as they can be constructed from the data already present. To reconstruct a link, use the following format:

https://freesound.org/people/{username}/sound/{id}

About this Dataset

LAION-Audio-630K is a large-scale audio-text dataset consisting of 633,526 pairs with the total duration of 4,325.39 hours. It contains audios of human activities, natural sounds and audio effects, consisting of 8 data sources (see the data source table below) from publicly available websites. We collect these datasets by downloading audios and relevant text descriptions. Based on our current knowledge, LAION-Audio-630K is the largest audio-text dataset publicly available and a magnitude larger than previous audio-text datasets (by 2022-11-05).

LAION-AI, github.com

Acknowledgment

The whole collection process as well as all usage of the LAION-Audio-630K are conducted by Germany non-profit pure research organization LAION. All contributors and collectors of the dataset are considered as open source contributors affiliated to LAION. These community contributors (Discord ids) include but not limited to: @marianna13#7139, @Chr0my#0173, @PiEquals4#1909, @Yuchen Hui#8574, @Antoniooooo#4758, @IYWO#9072, krishna#1648, @dicknascarsixtynine#3885, and @turian#1607. We would like to appreciate all of them for their efforts on the LAION-Audio-630k dataset.

License

  • LAION dataset metadata is released under The MIT License.
  • Audio is released under one of four licenses:

Please read the entirety of these licenses before deciding if you can use the audio for your project.