AudioSet Download

This repository contains code for downloading the AudioSet dataset. The code is provided as-is, and is not officially supported by Google.

Requirements

• Python 3.9 (it may work with other versions, but it has not been tested)
• ffmpeg (for audio processing and duration validation)

Installation

# Install ffmpeg
sudo apt install ffmpeg
# Install audioset-download
pip install audioset-download

Usage

Standard AudioSet Download

The following code snippet downloads the unbalanced train set, and stores it in the test directory. It only downloads the files associated with the Speech and Afrobeat labels, and uses two parallel processes for downloading. If a file is associated to multiple labels, it will be stored only once, and associated to the first label in the list.

from audioset_download import Downloader
d = Downloader(root_path='test', labels=["Speech", "Afrobeat"], n_jobs=2, download_type='unbalanced_train', copy_and_replicate=False)
d.download(format='vorbis')

Strong AudioSet Download

The library also supports downloading the "strong" version of AudioSet, which provides more precise temporal annotations. This creates TSV files with path, machine label, and display label columns.

from audioset_download import Downloader
d = Downloader(root_path='strong_dataset', labels=["Speech", "Music"], n_jobs=2)
d.download_strong(root_path='strong_dataset', format='wav', download_sets=['train', 'eval'])

This will create:

• strong_dataset/train_audio/ - directory with training audio files
• strong_dataset/eval_audio/ - directory with evaluation audio files
• strong_dataset/train_strong.tsv - TSV file with format: path\tlabel_machine\tlabel_display
• strong_dataset/eval_strong.tsv - TSV file with format: path\tlabel_machine\tlabel_display

Implementation

The main class is audioset_download.Downloader. It is initialized using the following parameters:

• root_path: the path to the directory where the dataset will be downloaded.
• labels: a list of labels to download. If None, all labels will be downloaded.
• n_jobs: the number of parallel downloads. Default is 1.
• download_type: the type of download. It can be one of the following:
  • balanced_train: balanced train set.
  • unbalanced_train: unbalanced train set. This is the default
  • eval: evaluation set.
• copy_and_replicate: if True if a file is associated to multiple labels, it will be copied and replicated for each label. If False, it will be associated to the first label in the list. Default is True.

The methods of the class are:

• download(format='vorbis', quality=5): downloads the standard dataset.

  • The format can be one of the following (supported by yt-dlp --audio-format parameter):
    • vorbis: downloads the dataset in Ogg Vorbis format. This is the default.
    • wav: downloads the dataset in WAV format.
    • mp3: downloads the dataset in MP3 format.
    • m4a: downloads the dataset in M4A format.
    • flac: downloads the dataset in FLAC format.
    • opus: downloads the dataset in Opus format.
    • webm: downloads the dataset in WebM format.
    • ... and many more.
  • The quality can be an integer between 0 and 10. Default is 5.

• download_strong(root_path, format='vorbis', quality=5, download_sets=['train', 'eval']): downloads the strong version of AudioSet with precise temporal annotations.

  • root_path: directory where to save the strong dataset files
  • format: audio format (same options as standard download)
  • quality: audio quality (0-10)
  • download_sets: list of sets to download ('train', 'eval', or both)

• read_class_mapping(): reads the class mapping file. It is not used externally.

• download_file(...): downloads a single file. It is not used externally.

• get_audio_duration(file_path): validates audio file duration using ffprobe.

Features

• Duration validation: Automatically checks downloaded audio files for valid duration (> 0.0 seconds) and removes invalid files.
• Parallel downloads: Supports multiple parallel download jobs for faster processing.
• Flexible formats: Supports multiple audio formats through yt-dlp.
• Strong dataset support: Downloads AudioSet strong labels with precise temporal annotations.
• Label filtering: Download only specific labels instead of the entire dataset.