df_eval

Public evaluation-only speech deepfake detection dataset, organized like Common Voice language configs: each config is a standard eval protocol (ASVspoof, ADD, In-the-Wild, …) with embedded audio.

Companion code: github.com/Shuo-H/df_eval

Configs are added as uploads complete. Declared configs below match currently available parquet shards on the Hub.

Load

from datasets import load_dataset

ds = load_dataset("shuohann/df_eval", name="sonar", split="test")
print(ds[0]["audio"], ds[0]["label"], ds[0]["file_id"])

Schema

column	type	description
audio	Audio	waveform
label	ClassLabel(bonafide, spoof)	ground-truth
file_id	string	basename of original file
source	string	config name
path	string	original protocol path (metadata)

Split: test only.

Planned configs

Aligned with Speech DF Arena protocol files:

• asvspoof_2019, asvspoof_2021_la, asvspoof_2021_df, asvspoof_2024
• add_2022_track_1, add_2022_track_3, add_2023_round_1, add_2023_round_2
• codecfake, dfadd, fake_or_real, in_the_wild, librisevoc, sonar

License / redistribution notice

This Hub dataset repackages evaluation audio from multiple third-party corpora. Each source has its own license (often research / non-commercial). Upstream sources include:

• ASVspoof 2019
• ASVspoof 2021 LA/DF
• ASVspoof 2024
• ADD 2022 / 2023
• FakeOrReal
• Codecfake
• DFADD
• LibriSeVoc
• SONAR
• In The Wild

You are responsible for complying with each original license. Prefer citing and downloading from the original authors when required. The uploader does not claim ownership of the underlying audio.

Citation

Please cite the original dataset papers for any config you use, and Speech DF Arena if you use their protocol definitions.