DECRO (eval)

Benchmark-ready packaging of the evaluation partition of the DECRO (DEepfake CROss-lingual) dataset — a cross-lingual (English + Chinese) speech anti-spoofing / synthetic-voice detection benchmark from Transferring Audio Deepfake Detection Capability across Languages (TheWebConf / WWW 2023).

Overview

DECRO pairs an English and a Chinese subset whose spoofed speech is generated with the same synthesis algorithms, so it isolates the effect of language on deepfake detection. The task is binary classification: bonafide (genuine human speech) vs. spoof (TTS / VC synthetic speech). This packaging contains the eval split of both language subsets, combined into a single test set. The original dataset is at https://github.com/petrichorwq/DECRO-dataset.

License & redistribution

Redistributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license — the upstream DECRO license. See LICENSE.txt. CC BY 4.0 permits redistribution and derivative works with attribution. Labels and the evaluation protocol are unmodified; the audio was decoded and re-encoded to canonical 16 kHz mono FLAC (the source mixes sampling rates — 16 / 22.05 / 24 / 44.1 kHz — and a few FLOAT-subtype WAVs, so a uniform 16 kHz re-encode is required for evaluation).

Schema

Column	Type	Description
path	string	<utterance_id>.flac, unique
audio	Audio(16000)	16 kHz mono FLAC
label	ClassLabel	"bonafide" (0) / "spoof" (1)
notes	string	JSON: utterance_id, language, speaker_id, system_id

utterance_id is the language-prefixed filename stem (en_<stem> / ch_<stem>); raw stems are not globally unique across the two subsets, so the prefix is the stable join key. system_id is the spoof algorithm (e.g. hifigan, vits, baidu_en) or, for bonafide, the source corpus (e.g. asv19, aishell1).

notes example:

{"utterance_id": "en_1-4993-40677-0048", "language": "en", "speaker_id": "1", "system_id": "baidu_en"}

Quick Start

from datasets import load_dataset

ds = load_dataset("SpeechAntiSpoofingBenchmarks/DECRO", split="test")
print(ds[0])

Stats

Stat	Value
Total trials	37,314
Bonafide	10,415
Spoof	26,899
English (en_eval)	19,190 (4,306 bonafide / 14,884 spoof)
Chinese (ch_eval)	18,124 (6,109 bonafide / 12,015 spoof)

Source provenance

• Original repository: https://github.com/petrichorwq/DECRO-dataset
• Protocols: en_eval.txt, ch_eval.txt (format: SPEAKER_ID AUDIO_FILE_NAME - SYSTEM_ID KEY)
• Bonafide: ASVspoof2019 LA (English); Aidatatang/Aishell/freeST/MagicData (Chinese). Spoof: WaveFake, FAD, and TTS/VC systems (Tacotron, FastSpeech2, VITS, StarGANv2-VC, NVC-Net, HiFiGAN, MB-MelGAN, PWG, Baidu, Xunfei).

Evaluation

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

Citation

@inproceedings{ba2023transferring,
  title     = {Transferring Audio Deepfake Detection Capability across Languages},
  author    = {Ba, Zhongjie and Wen, Qing and Cheng, Peng and Wang, Yuwei and Lin, Feng and Lu, Li and Liu, Zhenguang},
  booktitle = {Proceedings of the ACM Web Conference 2023 (WWW '23)},
  year      = {2023},
  doi       = {10.1145/3543507.3583222},
}

Maintainer

Contact: k.n.borodin@mtuci.ru