Access is granted automatically after you provide the requested information and accept the agreement. Use is limited to non-commercial academic research and education.

By requesting access, you confirm that you have read and accept the AT-ADD Dataset Access Agreement in this repository, including the non-commercial research, third-party licensing, conditional academic redistribution, takedown, privacy, and responsible-use conditions.

AT-ADD Track 1

This repository hosts Track 1 of the AT-ADD All-Type Audio Deepfake Detection Challenge. It contains the released audio splits and privacy-preserving sample-level metadata for non-commercial academic research and education.

Access

This is a gated dataset. Sign in to Hugging Face, review the access agreement, complete the short access form, and click Agree and access dataset. Access is granted automatically after acceptance.

Direct repository access is personal, and access credentials must not be shared. Redistribution is permitted only for non-commercial academic research or education under the conditions in the AT-ADD Dataset Access Agreement, including carrying the agreement with each copy and respecting the original licenses and terms of third-party source material.

Repository contents

The audio archives are provided at the repository root:

AT-ADD_T1_train.tar
AT-ADD_T1_dev.tar
AT-ADD_T1_eval.tar
AT-ADD_T1_progress.tar

Privacy-preserving metadata is provided under meta_label/:

meta_label/t1_train_meta.csv — 49,575 records
meta_label/t1_dev_meta.csv — 49,734 records
meta_label/t1_eval_meta.csv — 146,346 records

Each CSV row corresponds to one audio clip and uses UTF-8 with BOM. Track 1 metadata fields are:

name, label, lang, source, source_id, speaker_id, generation_task, generator, transformation

- means unavailable or not applicable. See meta_label/README.md for a directory-level description.

Metadata notes

lang gives the sample language. Regional and dialect tags are preserved; every released Track 1 record has a populated language tag.
source identifies the dataset-level source. Undisclosed internally collected sources are represented as Others.
For fake samples, source_id and speaker_id are suppressed as -, except for EchoFake-sourced samples where available official identifiers are retained for traceability.
Real speakers from undisclosed internal sources use consistent pseudonymous identifiers such as spk1 through spk341.
generation_task describes the generation paradigm, including Voice cloning, TTS, Voice conversion, Vocoder, SVC, and Replay where applicable.
generator follows the released AT-ADD generator labels.
transformation records secondary processing; - means that no transformation is recorded.
Internal raw paths and the private mapping from original identities to pseudonymous identifiers are not included.

Responsible use and citation

The dataset may be used only for non-commercial academic research and education. Users must respect third-party source licenses, privacy conditions, attribution requirements, and redistribution restrictions. If a rights holder or source provider raises a complaint or requests removal, the organizers may restrict or remove the affected content at any time; users who are notified must stop using and redistributing it. Publications or public results using this dataset should cite the official AT-ADD paper and acknowledge the dataset.

Questions, rights-holder concerns, or deletion requests may be submitted through this repository's Community tab.

Downloads last month: 99

Total file size:

25 GB