ArA-DF-2026

ArA-DF-2026 is an Arabic speech deepfake detection dataset for binary audio classification.

• 0: spoofed or synthetic speech
• 1: bona fide speech

The public train and dev splits include labels. All public development-test and final-test splits are unlabeled and are scored externally through CodaBench.

All released audio is 16 kHz mono PCM audio packaged as lossless FLAC inside WebDataset TAR shards.

Splits

Split	Rows	Hours	Labels	Public naming
train	22,500	59.56	yes	filenames include dialect-family and gender tokens
dev	21,000	52.69	yes	filenames include dialect-family and gender tokens
track-1_development_test	16,023	40.94	no	anonymized filenames
track-1_test	144,210	366.16	no	anonymized filenames
track-2_development_test	14,193	36.51	no	anonymized filenames
track-2_test	127,746	329.03	no	anonymized filenames

Train/dev filenames use <dialect-family>_<gender>_<split>_<7-digit-number>. Test filenames use test_<7-digit-number> and intentionally do not expose dialect-family, gender, labels, or source information.

Each public test set is provided as a separate Hugging Face config. Within those configs, the Hugging Face split is named test for compatibility with the datasets library; the challenge split name is preserved in the metadata split column and in the file paths.

Challenge Tracks

Track 1 evaluates deepfake detection under dialect-generalization conditions. The development-test split is for experimentation during the development phase; the final-test split is for official evaluation.

Track 2 evaluates deepfake detection under held-out acoustic and channel robustness conditions. The development-test split is for experimentation during the development phase; the final-test split is for official evaluation.

Development phase: June 15 to July 20, with up to 100 submissions per team and at most 10 submissions per day. The best development submission is shown on the public leaderboard and does not require a final commitment.

Evaluation phase: July 20 to July 25, with up to 3 system submissions per team on the blind final-test splits. The best EER across the 3 submissions is the official score.

CodaBench competitions:

• Track 1: https://www.codabench.org/competitions/17138/
• Track 2: https://www.codabench.org/competitions/17139/

Baseline model: https://huggingface.co/ArabicSpeech/ArA-DF-Baseline

Submission Format

Submit one CSV per CodaBench phase/track:

audio_id,logit
test_0000000,-0.81234567
test_0000001,1.43678901

audio_id must match the public test ID. logit must be a real-valued score where higher values indicate stronger confidence that the utterance is bona fide. This matches the dataset label map: 1 = bona fide speech and 0 = spoofed or synthetic speech. Sample submission files contain random placeholder logits and are only format templates.

Loading Metadata

from datasets import load_dataset

train_dev = load_dataset("ArabicSpeech/ArA-DF-2026")
track1_dev = load_dataset("ArabicSpeech/ArA-DF-2026", "track-1_development_test")["test"]
track1_test = load_dataset("ArabicSpeech/ArA-DF-2026", "track-1_test")["test"]
track2_dev = load_dataset("ArabicSpeech/ArA-DF-2026", "track-2_development_test")["test"]
track2_test = load_dataset("ArabicSpeech/ArA-DF-2026", "track-2_test")["test"]

Loading Audio Shards

The audio is packaged as WebDataset shards under data/{split}/. Each sample has a FLAC file and a JSON sidecar with the same stem.

from datasets import load_dataset

data_files = {
    "train": "hf://datasets/ArabicSpeech/ArA-DF-2026/data/train/*.tar",
    "dev": "hf://datasets/ArabicSpeech/ArA-DF-2026/data/dev/*.tar",
    "track-1_development_test": "hf://datasets/ArabicSpeech/ArA-DF-2026/data/track-1_development_test/*.tar",
    "track-1_test": "hf://datasets/ArabicSpeech/ArA-DF-2026/data/track-1_test/*.tar",
    "track-2_development_test": "hf://datasets/ArabicSpeech/ArA-DF-2026/data/track-2_development_test/*.tar",
    "track-2_test": "hf://datasets/ArabicSpeech/ArA-DF-2026/data/track-2_test/*.tar",
}
ds = load_dataset("webdataset", data_files=data_files, streaming=True)
sample = next(iter(ds["train"]))