Spatial mixup: Directional loudness modification as data augmentation for sound event localization and detection

PyTorch implementation and demo of: Spatial mixup: Directional loudness modification as data augmentation for sound event localization and detection [arxiv]

Summary

This repo contains a PyTorch implementation of the Spatial Mixup [arxiv]] data augmentation technique for spatial audio. Spatial mixup: Directional loudness modification as data augmentation for sound event localization and detection. We include a Jupyter notebook that illustrates the process as well a small demo that trains a small Sound Event Localization and Detection network, using the DCASE 2021 task 3 dataset.

Reference

Spatial mixup: Directional loudness modification as data augmentation for sound event localization and detection, ICASSP 2022 [arxiv]]

-- Ricardo Falcon-Perez, Kazuki Shimada, Yuichiro Koyama, Shusuke Takahashi, Yuki Mitsufuji

TL;DR

• Spatial mixup modifies the directional loudness of the input signals.
• This is a soft spatial filter, than can be used as data augmentation for SELD tasks.
• The transform should not be too extreme, as the sound scene will be too different.

Requirements

We use [Spaudiopy ] to compute the spherical harmonics and other spatial audio operations.

conda create -n YOUR_ENV_NAME python=3.7
conda activate YOUR_ENV_NAME
pip install spaudiopy

Or you can create a conda environment from the provided file:

conda env create --file environment.yaml

How to use it

Refer to the file spatial_mixup.py which inlcudes the core processing.

Then the notebook demo_spatial_mixup.ipnybshows how to use it, and a few exmaples.

Citation

@INPROCEEDINGS{9747312,
  author={Falcón-Pérez, Ricardo and Shimada, Kazuki and Koyama, Yuichiro and Takahashi, Shusuke and Mitsufuji, Yuki},
  booktitle={ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  title={Spatial Mixup: Directional Loudness Modification as Data Augmentation for Sound Event Localization and Detection}, 
  year={2022},
  volume={},
  number={},
  pages={431-435},
  doi={10.1109/ICASSP43922.2022.9747312}}

License

Distributed under the MIT License. See LICENSE.txt for more information.

(back to top)

Acknowledgments

Thanks to Chris Hold for his spatial audio library and his comments and recomendations on how to use it.