DECRO
收藏资源简介:
DECRO(eval)是DECRO(DEepfake CROss-lingual)数据集的评估分区,经过重新封装,可直接用于基准测试。该数据集是一个跨语言(英语和中文)的语音反欺骗或合成语音检测基准,源自WWW 2023会议论文《Transferring Audio Deepfake Detection Capability across Languages》。其核心设计是配对英语和中文子集,其中欺骗语音使用相同的合成算法生成,从而可以隔离语言对深度伪造检测的影响。数据集的任务是二分类:区分真实语音和欺骗语音(由TTS或语音转换系统生成的合成语音)。此版本包含了两个语言子集的评估集,合并为一个统一的测试集。数据总样本量为37,314条,其中真实语音10,415条,欺骗语音26,899条。具体而言,英语子集(en_eval)包含19,190条样本(4,306条真实,14,884条欺骗),中文子集(ch_eval)包含18,124条样本(6,109条真实,12,015条欺骗)。每个数据样本包含以下字段:path(唯一音频文件名,格式为<utterance_id>.flac)、audio(统一编码为16 kHz单声道FLAC格式的音频数据)、label(分类标签,bonafide表示真实,spoof表示欺骗)以及notes(一个JSON字符串,包含原始话语ID、语言、说话人ID和系统ID)。其中,系统ID标识了生成欺骗语音的具体算法(例如HiFiGAN、VITS)或真实语音的来源语料库。真实语音来源于ASVspoof2019 LA(英语)以及Aidatatang、Aishell等多个中文数据集;欺骗语音则来源于WaveFake、FAD以及多种TTS/VC系统(如Tacotron、FastSpeech2、StarGANv2-VC等)。数据集遵循Creative Commons Attribution 4.0 International (CC BY 4.0)许可协议。
DECRO (eval) is the evaluation partition of the DECRO (DEepfake CROss-lingual) dataset, repackaged for direct use in benchmarking. It serves as a cross-lingual (English and Chinese) speech anti-spoofing/synthetic speech detection benchmark, originating from the WWW 2023 conference paper Transferring Audio Deepfake Detection Capability across Languages. Its core design pairs English and Chinese subsets, where spoofed speech is generated using the same synthesis algorithms, enabling isolation of language impact on deepfake detection. The datasets task is binary classification: distinguishing between genuine (bonafide) and spoofed speech (synthetic speech generated by TTS/voice conversion systems). This version includes evaluation sets from both language subsets, merged into a unified test set. The total sample size is 37,314, with 10,415 bonafide and 26,899 spoofed samples. Specifically, the English subset (en_eval) contains 19,190 samples (4,306 bonafide, 14,884 spoofed), and the Chinese subset (ch_eval) contains 18,124 samples (6,109 bonafide, 12,015 spoofed). Each sample includes fields such as path (unique audio filename in <utterance_id>.flac format), audio (audio data uniformly encoded as 16 kHz mono FLAC), label (classification label, bonafide for genuine, spoof for spoofed), and notes (a JSON string containing original utterance ID, language, speaker ID, and system ID). The system ID identifies the specific algorithm used to generate spoofed speech (e.g., HiFiGAN, VITS) or the source corpus for bonafide speech. Bonafide speech is sourced from ASVspoof2019 LA (English) and multiple Chinese datasets like Aidatatang and Aishell; spoofed speech comes from WaveFake, FAD, and various TTS/VC systems (e.g., Tacotron, FastSpeech2, StarGANv2-VC). The dataset is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.
数据集简介
- 名称:DECRO (eval)
- 任务:音频分类(语音反欺骗/合成语音检测),二分类任务:真实语音(bonafide) vs 伪造语音(spoof)
- 语言:英语 + 中文(跨语言)
- 许可证:Creative Commons Attribution 4.0 International (CC BY 4.0)
- 论文来源:Transferring Audio Deepfake Detection Capability across Languages (TheWebConf / WWW 2023)
- 原始数据仓库:https://github.com/petrichorwq/DECRO-dataset
数据规模与分布
| 统计项 | 数值 |
|---|---|
| 总样本数 | 37,314 |
| 真实语音(Bonafide) | 10,415 |
| 伪造语音(Spoof) | 26,899 |
| 英语子集(en_eval) | 19,190(真实 4,306 / 伪造 14,884) |
| 中文子集(ch_eval) | 18,124(真实 6,109 / 伪造 12,015) |
数据模式(Schema)
| 列名 | 类型 | 说明 |
|---|---|---|
path |
string |
<utterance_id>.flac,唯一标识 |
audio |
Audio(16000) |
16 kHz 单声道 FLAC 格式 |
label |
ClassLabel |
"bonafide" (0) / "spoof" (1) |
notes |
string |
JSON 格式,包含 utterance_id, language, speaker_id, system_id |
- 每条样本的
utterance_id带有语言前缀(en_<stem>或ch_<stem>),确保跨子集唯一。 system_id表示伪造算法(如hifigan,vits,baidu_en)或真实语音的来源语料库(如asv19,aishell1)。
数据来源与处理
- 真实语音来源:
- 英语:ASVspoof2019 LA
- 中文:Aidatatang、Aishell、freeST、MagicData
- 伪造语音来源:WaveFake、FAD 以及多种 TTS/VC 系统,包括 Tacotron、FastSpeech2、VITS、StarGANv2-VC、NVC-Net、HiFiGAN、MB-MelGAN、PWG、Baidu、Xunfei。
- 音频处理:原始音频采样率不一(16 / 22.05 / 24 / 44.1 kHz),统一重编码为 16 kHz 单声道 FLAC。
- 许可说明:基于原数据集 CC BY 4.0 许可重新分发,标签与评估协议未修改。
快速使用
python from datasets import load_dataset ds = load_dataset("SpeechAntiSpoofingBenchmarks/DECRO", split="test") print(ds[0])
评估与引用
- 评估说明与提交格式:参见数据集中
submissions/README.md。 - 引用: bibtex @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}, }
维护者
- 联系邮箱:k.n.borodin@mtuci.ru




