RIRplay: A Replay Stereo Corpus for Voice Biometrics Anti-Spoofing

This is the Zenodo repository for the v0.5 RIRplay dataset. RIRplay is a novel simulated corpus for Physical Access (PA) spoofing detection, created to support research on replay attack countermeasures for voice biometric systems. Unlike existing replay corpora, which are often limited in size, diversity, or realism, RIRplay reproduces the full replay attack pipeline through simulation, including realistic room acoustics, varied playback and recording conditions, and non-linear device effects. The database has been developed to address the lack of representative training data for replay detection and to improve the robustness and generalization of anti-spoofing systems against replay attacks in unseen scenarios. Results on benchmarks including ASVspoof 2021 and ReMASC show that models trained with RIRplay clearly outperform those trained on previous corpora, such as ASVspoof 2019, confirming its value as a scalable resource for the development of replay attack detection systems.

More details about the generation process and the conducted experiments for version v0.0 can be found in the published IEEE Transactions on Information Forensics and Security paper:

J. C. Sanchez-Valera, A. M. Peinado, J. M. Martin-Doñas, A. Gomez-Alanis, A. M. Gomez, and M. Todisco, "RIRplay: Generation of a replay stereo corpus for voice biometrics anti-spoofing," IEEE Transactions on Information Forensics and Security, 2026.

We would greatly appreciate it if you cite the aforementioned journal article, rather than only the Zenodo repository, when using this dataset or comparing your results with ours:

@ARTICLE{sanchezvalera26,
author={Sanchez-Valera, Jose C. and Peinado, Antonio M. and Martin-Doñas, Juan M. and Gomez-Alanis, Alejandro and Gomez, Angel M. and Todisco, Massimiliano},
journal={IEEE Transactions on Information Forensics and Security},
title={RIRplay: Generation of a Replay Stereo Corpus for Voice Biometrics Anti-Spoofing},
year={2026},
volume={21},
number={},
pages={4106-4118},
doi={10.1109/TIFS.2026.3684815}
}

Version v0.5 release notes (This version)

Version v0.5 fixes clipping and short-duration issues detected in some audio files after an exhaustive audio analysis. This new release normalizes spoofed audio signals to unit amplitude before applying the loudspeaker non-linear processing. In addition, it adopts a new strategy regarding VAD trimming: if the default silence-detection threshold (0.5) excessively removes regions that actually contain speech, the audio is kept as originally generated.

This version also includes the files containing all metadata (RIRplay-PA-*_150426_ASASVI_V0.5_metadata.txt) used to acoustically simulate the different generation parameters for bona fide and spoofed samples, as well as the correspondence information required to leverage the stereo feature of the dataset. 

For further details about this release, including updated benchmark results and usage notes, please download and consult the file README_RIRplay.pdf included in this repository.

Please note that this RIRplay variant refers to the extended version reported in the article

Users interested in using a reduced version of the dataset may simply select half of the samples listed in the metadata files of each subset. In that case, the resulting number of audio files becomes directly comparable to the size of the ASVspoof 2019 dataset, enabling fairer dataset-scale comparisons in training and evaluation.

Acknowledgments and funding

The associated publication is part of the project PID2022-138711OB-I00 funded by MICIU/AEI/10.13039/501100011033 and by ERDF/EU. This dataset follows the Data Management Plan of the ASASVI project (Signal and Neural Processing against Spoofing Attacks and Deepfakes for Secure Voice Interaction, PID2022-138711OB-I00). Its ASASVI dataset identifier is RIRplay-PA-*_150426_ASASVI_V0.5, and the corresponding DMP metadata record is included as RIRplay-PA_150426_ASASVI_V0.5.xml.

Please check LICENSE.txt and README_RIRplay.pdf files before downloading the database.