Datasets:

bond005
/

audioset-nonspeech

dataset_info:
  features:
    - name: audio
      dtype: audio
    - name: label
      dtype: string
  splits:
    - name: train
      num_bytes: 2972583582.24
      num_examples: 9418
    - name: validation
      num_bytes: 335149610.534
      num_examples: 1059
    - name: test
      num_bytes: 1114463859.732
      num_examples: 3517
  download_size: 4328763714
  dataset_size: 4422197052.506
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
license: cc-by-4.0
task_categories:
  - audio-classification
language:
  - en

Audioset-Nonspeech

Audioset-Nonspeech is a processed version of the well-known agkphysics/AudioSet dataset. The processing was performed to remove all audio recordings that may contain clearly distinguishable human speech, leaving only non-speech audio recordings. The resulting Audioset-Nonspeech dataset can be used not only for audio event classification but also for augmentation (mixing with a specified signal-to-noise ratio) of speech recordings when training speech recognition models.

Description

For each example, the following is provided:

Link to the audio data (audio)
List of class labels, separated by semicolons (label)

Usage example

from datasets import load_dataset

ds = load_dataset("bond005/audioset-nonspeech")