Audio Augmentation using AudioSet

• AudioSet

  • Google's audio dataset that manually annotated audio events
  • ontology : https://github.com/audioset/ontology

• Goal

  • Augment with various sound situation for speech related tasks.

• Report on research case

  • If you wanna study on specific dataset, it can be not effects for getting better result.
    • In source separation case, it gets higher losses with audioset augmentation.
  • But it gets better result on test cases.

How to Use

• Installation
  • install ffmpeg version 4, and this package

$ apt install -y software-properties-common
$ add-apt-repository ppa:jonathonf/ffmpeg-4
$ apt update
$ apt install -y ffmpeg
$ pip install -e .

• Download
  • Audioset give us separated meta information that label balanced or not.
  • default : balanced

$ python audioset_augmentor/download.py [--file_path='assets/balanced_train_segments.csv' --savedir='.data' --n_jobs=4 --delay=0.05]

• Preprocess Audio
  • Process adjust volume, sample rate, file type on audio files.

$ python audioset_augmentor/preprocess.py [--master_dir, --out_dir, --meta_path='assets/balanced_train_segments.csv' --out_sr=22050 --min_size=1000000(file checker) --n_jobs=4]

• After all, you should set master_dir on assets/default.json for using augment function.

License

• 'Audioset' license is announced on https://research.google.com/audioset/download.html

  • File List

    1. assets/balanced_train_segments.csv
    2. assets/unbalanced_train_segments.csv
    3. assets/ontology.json

The dataset is made available by Google Inc. under a Creative Commons Attribution 4.0 International (CC BY 4.0) license, while the ontology is available under a Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license.

• Other sources are under MIT License