CHiME-5
Warning: The CHiME-5 dataset has been superseded by the newer CHiME-6 release which fixes audio alignment issues that existed in the original data. We recommend that you use CHiME-6 instead of CHiME-5 for any new work.
As of 1st Jan 2024, the CHiME-5 dataset has been re-issued under a standard CC BY-SA 4.0 license, and is free for both academic and commercial purposes.
As of November 2026, the CHiME-5 dataset has been re-issused under a standard CC BY 4.0 license (it was previously CC BY-SA 4.0), and is free for both academic and commercial purposes.
The data is available on HuggingFace at https://huggingface.co/datasets/chimechallenge/chime5
You will find the following six gzipped tar files:
- CHiME5_train.tar.gz [91 GB] - the training set audio
- CHiME5_dev.tar.gz [10 GB] - the development set audio
- CHiME5_eval.tar.gz [13 GB] - the evaluation set audio
- CHiME5_transcriptions.tar.gz [10 MB] - complete set of transcriptions
- CHiME5_floorplans.tar.gz [1.4 MB] - approximate floorplans for the 20 living spaces.
Note, for publications that use the data please cite the following paper:
- Barker, J., Watanabe, S., Vincent, E., Trmal, J. (2018) The Fifth ‘CHiME’ Speech Separation and Recognition Challenge: Dataset, Task and Baselines. Proc. Interspeech 2018, 1561-1565, doi: 10.21437/Interspeech.2018-1768
You can use the follow bib entry:
@inproceedings{barker18_interspeech,
author={Jon Barker and Shinji Watanabe and Emmanuel Vincent and Jan Trmal},
title={{The Fifth 'CHiME' Speech Separation and Recognition Challenge: Dataset, Task and Baselines}},
year=2018,
booktitle={Proc. Interspeech 2018},
pages={1561--1565},
doi={10.21437/Interspeech.2018-1768}
}