SpeechAntiSpoofingBenchmarks/ASVspoof2021_DF

ASVspoof 2021 DF

Benchmark-ready packaging of the DeepFake (DF) evaluation partition from ASVspoof 2021 for speech anti-spoofing and synthetic / deepfake voice detection.

Overview

This dataset contains the DF evaluation subset of the ASVspoof 2021 challenge. The task is binary classification: bonafide (genuine human speech) vs. spoof (synthetic, converted, or otherwise manipulated speech). The original dataset is available at https://www.asvspoof.org/index2021.html.

License & redistribution

This dataset is redistributed under the Open Data Commons Open Database License (ODbL). See LICENSE.txt for the full text.

The labels and evaluation protocol are unmodified. The audio was transcoded to canonical 16 kHz mono FLAC: a large fraction of the source FLAC files use an encoding the standard libsndfile/soundfile decoder cannot read (and therefore the Hugging Face datasets audio decoder cannot read), so each clip was decoded and re-encoded to a clean, universally decodable FLAC. The PCM samples are preserved bit-exactly (source → re-encode → decode round-trip difference of 0); the sampling rate (16 kHz) is unchanged.

Schema

Column	Type	Description
path	string	Stable archive-relative path (e.g. DF_E_2000011.flac), unique within dataset
audio	Audio(16000)	Audio waveform, 16 kHz mono
label	ClassLabel	"bonafide" (index 0) or "spoof" (index 1)
notes	string	JSON with utterance_id, speaker_id, subset, codec, source, attack_id, vocoder

notes example:

{"utterance_id": "DF_E_2000011", "speaker_id": "LA_0023", "subset": "progress", "codec": "nocodec", "source": "asvspoof", "attack_id": "A14", "vocoder": "traditional_vocoder"}

Quick Start

from datasets import load_dataset

ds = load_dataset("SpeechAntiSpoofingBenchmarks/ASVspoof2021_DF", split="test")
print(ds[0])
# {'path': 'DF_E_2000011.flac', 'audio': {'array': ..., 'sampling_rate': 16000},
#  'label': 1, 'notes': '{"utterance_id": "DF_E_2000011", ...}'}

Stats

Stat	Value
Total trials	611,829
Bonafide	22,617
Spoof	589,212

Source provenance

• Original dataset: https://www.asvspoof.org/index2021.html
• Evaluation protocol: protocols/ASVspoof2021.DF.cm.eval.trl.txt

Evaluation

For evaluation instructions and submission format, see submissions/README.md.

Citation

Original paper: https://arxiv.org/abs/2109.00537

arXiv version:

@misc{yamagishi2021asvspoof2021acceleratingprogress,
      title={ASVspoof 2021: accelerating progress in spoofed and deepfake speech detection},
      author={Junichi Yamagishi and Xin Wang and Massimiliano Todisco and Md Sahidullah and Jose Patino and Andreas Nautsch and Xuechen Liu and Kong Aik Lee and Tomi Kinnunen and Nicholas Evans and Héctor Delgado},
      year={2021},
      eprint={2109.00537},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2109.00537},
}

Peer-reviewed publication (ASVspoof 2021 Workshop):

@inproceedings{yamagishi21_asvspoof,
  title     = {{ASVspoof 2021: accelerating progress in spoofed and deepfake speech detection}},
  author    = {Junichi Yamagishi and Xin Wang and Massimiliano Todisco and Md Sahidullah and Jose Patino and Andreas Nautsch and Xuechen Liu and Kong Aik Lee and Tomi Kinnunen and Nicholas Evans and Héctor Delgado},
  year      = {2021},
  booktitle = {{2021 Edition of the Automatic Speaker Verification and Spoofing Countermeasures Challenge}},
  pages     = {47--54},
  doi       = {10.21437/ASVSPOOF.2021-8},
}

Maintainer

Contact: k.n.borodin@mtuci.ru