EARS-WHAM and EARS-Reverb generation scripts

This repository contains generation scripts for the EARS-WHAM and EARS-Reverb benchmarks.

Changelog:

• [04/25/2025] Version 2 (v2) released.
• [06/14/2024] Version 1 (v1) released.

Environment

Please make sure you have installed the required packages.

You can run the following command to install them:

python -m pip install -r requirements.txt

Generate EARS-WHAM_v2

To generate EARS-WHAM, first download the original EARS and WHAM! datasets with the following command, where <data_dir> is the directory where the datasets will be downloaded:

bash download_ears_wham.sh <data_dir>

Then you can generate the EARS-WHAM dataset with the following command:

python generate_ears_wham.py --data_dir <data_dir>

Optionally, you can generate a 16 kHz version by adding the --16k flag:

python generate_ears_wham.py --data_dir <data_dir> --16k

Improvements in EARS-WHAM_v2:

• Uses the recommended train, valid, test split from WHAM for the noise files.
• Filters low-frequency noise in clean files with a high-pass filter.
• Utilizes energy thresholds to avoid empty speech files.

Generate EARS-Reverb_v2

To generate EARS-Reverb, first download the original EARS and RIR datasets with the following command, where <data_dir> is the directory where the datasets will be downloaded:

bash download_ears_reverb.sh <data_dir>

Then you can generate the EARS-Reverb dataset with the following command:

python generate_ears_reverb.py --data_dir <data_dir>

Optionally, you can generate a 16 kHz version by adding the --16k flag:

python generate_ears_wham.py --data_dir <data_dir> --16k

Improvements in EARS-Reverb_v2:

• Defines train, valid, and test splits for RIRs by splitting among datasets.
• Filters low-frequency noise in clean files with a high-pass filter.
• Utilizes energy thresholds to avoid empty speech files.

License

The code and dataset are released under CC-NC 4.0 International license.

References

If you use the dataset or any derivative of it, please cite our research paper:

@inproceedings{richter2024ears,
  title={{EARS}: An Anechoic Fullband Speech Dataset Benchmarked for Speech Enhancement and Dereverberation},
  author={Richter, Julius and Wu, Yi-Chiao and Krenn, Steven and Welker, Simon and Lay, Bunlong and Watanabe, Shinjii and Richard, Alexander and Gerkmann, Timo},
  booktitle={ISCA Interspeech},
  year={2024}
}

For audio samples, visit the project page.