Participation Instructions

Note on Approval Timeline 

Approvals for Hugging Face and CodaBench are typically completed within one business day. Please submit both applications together and wait patiently for the approval. 

1. Register on Hugging Face to request dataset access

   Please first complete the registration form on the Hugging Face dataset page for AT-ADD. Access to the dataset is typically granted within one day after submission.

   - Track 1: https://huggingface.co/datasets/xieyuankun/AT-ADD-Track1

   - Track 2: https://huggingface.co/datasets/xieyuankun/AT-ADD-Track2

2. Register on Codabench

   Please register for the challenge on Codabench.  

   Note that each team is allowed to use only one Codabench account, and the email address must be consistent with the one used during Hugging Face registration.

   - Track 1: https://www.codabench.org/competitions/15477

   - Track 2: https://www.codabench.org/competitions/15481

3. Generate Predictions and Submit for Evaluation  

Participants are required to train their models using the officially released training and development sets, and then perform inference on the test audio. For each test utterance, the system must output a deterministic binary prediction, namely real or fake. After submission, the platform will automatically compute the scores and generate the leaderboard. 

Submissions should be uploaded as a `.zip` file (the file name can be arbitrary).

   - The zip file must contain a single file named:

     predict.csv

   - The required format of `predict.csv` is:

     name,predict

     ATADD_T1_Eval_000001.flac,fake

     ATADD_T1_Eval_000002.flac,real

     ...

   where:

   - `name` is the audio file name  

   - `predict` should be either `real` or `fake`