ASVspoof logo 

 

ASVspoof 2021

POST CHALLENGE RELEASES

We are pleased to announce the release of the keys and detailed metadata for the ASVspoof 2021 logical access (LA), physical access (PA) and deepfake (DF) databases used for the ASVspoof 2021 challenge.

Each database comprises real bonafide (human) and spoofed speech generated with diverse spoofing attack algorithms. The databases are available under an Open Data Commons Attribution Licence and can be downloaded from the Zenodo repository.

SPEECH DATA

LA subset: https://zenodo.org/record/4837263

PA subset: https://zenodo.org/record/4834716

DF subset: https://zenodo.org/record/4835108

EVALUATION KEYS AND METADATA

LA keys and metadata: https://www.asvspoof.org/asvspoof2021/LA-keys-full.tar.gz

PA keys and metadata: https://www.asvspoof.org/asvspoof2021/PA-keys-full.tar.gz

DF keys and metadata: https://www.asvspoof.org/asvspoof2021/DF-keys-full.tar.gz

CHALLENGE BASELINES & SCRIPTS TO COMPUTE EER and min t-DCF

https://github.com/asvspoof-challenge/2021

POST-CHALLENGE ANALYSIS PAPER

An overview article on the ASVspoof 2021 challenge has been submitted to IEEE/ACM-T-ASLP. It contains a detailed analysis of the results, limitations and potential future directions and is available from arXiv: https://arxiv.org/abs/2210.02437

 

TIMELINE OF THE CHALLENGE

ASVspoof 2021 results summary paper

1st December

Keys and metadata are now available: LA, PA and DF.

16th September

ASVspoof 2021 workshop proceedings are now available on the ISCA archive!

3rd September

ASVspoof 2021 workshop: technical program and registration now available!

24th June

paper submission platform is open. Click here to submit your paper!

1st June

CodaLab leaderboards online

 

source code for 4 reproducible baselines (LFCC-GMM, CQCC-GMM, LFCC-LCNN, RawNet2) are available from

 

challenge metric (tandem detection cost function) reference implementations

 

 28th May

 databases release via Zenodo

 

 ASVspoof 2021 evaluation plan v0.4

 

ANNOUNCEMENT

We are delighted to announce the 4th edition of the Automatic Speaker Verification Spoofing and Countermeasures. ASVspoof 2021 will be organised as a satellite event in conjunction with Interspeech 2021


ASVspoof 2021 will comprise three separate sub-challenges:

  • a logical access sub-challenge extending the 2019 challenge with a focus on robustness to channel variation;

  • a physical access sub-challenge similar to the 2019 setup, but with recordings made in real physical environments;

  • speech deepfake detection sub-challenge (no ASV).

While no new training and no new development data will be released, new evaluation data will be released for each of the three conditions.

Participants will be encouraged to prepare their own training and development data by using ASVspoof 2019 training and development subsets and data augmentation techniques, e.g. to simulate channel variability. The challenge evaluation plan will be released in the first quarter of 2021, with a submission deadline in the second quarter.

 

List of organisers (alphabetical)

Héctor Delgado, Nuance Communications Inc., Spain
Nicholas Evans, EURECOM, France
Tomi Kinnunen, University of Eastern Finland, Finland
Kong Aik Lee, Institute for Infocomm Research, Singapore
Xuechen Liu, University of Eastern Finland, Finland
Andreas Nautsch, EURECOM, France
Jose Patino, EURECOM, France
Md Sahidullah, Inria, France
Massimiliano Todisco, EURECOM, France
Xin Wang, National Institute of Informatics, Japan
Junichi Yamagishi, National Institute of Informatics, Japan

 

Funding agencies 

Academy of Finland Agence Nationale de la Recherche Japan Science and Technology Agency Agency for Science, Technology and Research