XMAD
收藏资源简介:
XMAD-Bench(跨领域多语言音频深度伪造基准)是一个大规模、平衡、多语言的音频反欺骗基准测试数据集,旨在评估模型在跨领域和跨语言场景下的音频深度伪造检测泛化能力。数据内容为二分类任务,区分真实的人类录音(bonafide)与合成或转换的伪造语音(spoof)。数据集总计包含368,085个音频样本,其中160,656个为真实样本,207,429个为伪造样本。数据覆盖7种语言:阿拉伯语(ar)、德语(de)、英语(en)、西班牙语(es)、罗马尼亚语(ro)、俄语(ru)和中文(zh)。音频数据来源于多个公开语音语料库,包括Common Voice、MASC、M-AILABS、VoxPopuli和AISHELL-3,这种多源构成旨在对模型的跨领域泛化能力形成压力测试。数据集中包含14个子集,其中9个子集同时包含真实和伪造两类样本,而5个来自M-AILABS和MASC的子集在此版本中仅包含伪造音频。每个数据样本包含以下字段:`path`(源相对路径)、`audio`(16 kHz单声道WAV格式的音频数据)、`label`(分类标签,bonafide 或 spoof)、`notes`(一个JSON字符串,包含话语ID、语言、语料库、说话人ID、攻击方法和数据划分等元信息)。该数据集适用于音频分类任务,特别是音频深度伪造检测和反欺骗研究。
XMAD-Bench (Cross-domain Multilingual Audio Deepfake Benchmark) is a large-scale, balanced, multilingual audio anti-spoofing benchmark dataset designed to evaluate the generalization capabilities of models in cross-domain and cross-language scenarios for audio deepfake detection. The data content is a binary classification task, distinguishing between genuine human recordings (bonafide) and synthetic or converted fake speech (spoof). The dataset contains a total of 368,085 audio samples, with 160,656 being genuine samples and 207,429 being fake samples. It covers 7 languages: Arabic (ar), German (de), English (en), Spanish (es), Romanian (ro), Russian (ru), and Chinese (zh). The audio data is sourced from multiple public speech corpora, including Common Voice, MASC, M-AILABS, VoxPopuli, and AISHELL-3, with this multi-source composition aimed at stress-testing models cross-domain generalization abilities. The dataset includes 14 subsets, of which 9 subsets contain both genuine and fake samples, while 5 subsets from M-AILABS and MASC contain only fake audio in this version. Each data sample includes the following fields: `path` (source relative path), `audio` (audio data in 16 kHz mono WAV format), `label` (classification label, bonafide or spoof), and `notes` (a JSON string containing metadata such as utterance ID, language, corpus, speaker ID, attack method, and data split). This dataset is suitable for audio classification tasks, particularly for audio deepfake detection and anti-spoofing research.
数据集概述:XMAD-Bench — 跨领域多语言音频深度伪造基准
许可证:CC BY-NC-SA 4.0(非商业用途,需署名且以相同方式共享)
语言:阿拉伯语 (ar)、德语 (de)、英语 (en)、西班牙语 (es)、罗马尼亚语 (ro)、俄语 (ru)、中文 (zh)
任务类别:音频分类(二分类任务:真实人声 vs 合成/转换语音)
数据集大小:约 368,085 条样本(100K < n < 1M)
标签:is_fake 字段(0 → 真实 real/,1 → 伪造 fake/)
数据构成
- 总样本数:368,085
- 真实(Bonafide):160,656
- 伪造(Spoof):207,429
- 子集数量:14 个(其中 9 个包含两类样本,5 个仅含伪造样本)
- 采样率:16 kHz 单声道 WAV
数据来源语料库
| 语言 | 来源语料库 |
|---|---|
| 阿拉伯语 (ar) | Common Voice, MASC |
| 德语 (de) | Common Voice, M-AILABS |
| 英语 (en) | Common Voice, M-AILABS |
| 西班牙语 (es) | Common Voice, M-AILABS |
| 罗马尼亚语 (ro) | Common Voice, VoxPopuli |
| 俄语 (ru) | Common Voice, M-AILABS |
| 中文 (zh) | AISHELL-3, Common Voice |
注意:M-AILABS(de/en/es/ru)和 MASC(ar)子集仅提供伪造音频,对应真实音频行已被移除(共 46,773 行缺失)。
数据模式(Schema)
| 列名 | 类型 | 描述 |
|---|---|---|
path |
string | 源相对路径,唯一标识(例如 en/commonvoice-en/fake/common_voice_en_100179.wav) |
audio |
Audio(16000) | 16 kHz 单声道 WAV 音频 |
label |
ClassLabel | "bonafide" (0) 或 "spoof" (1) |
notes |
string | JSON 格式,包含 utterance_id、language、corpus、speaker_id、attack、split |
notes 示例:
json
{"utterance_id": "en__commonvoice-en__fake__common_voice_en_100179", "language": "en", "corpus": "commonvoice-en", "speaker_id": "...", "attack": "xtts_v2", "split": "val"}
utterance_id 格式为 <lang>__<corpus>__<real|fake>__<filename-stem>,确保全局唯一。
论文与引用
- 论文:XMAD-Bench: Cross-Domain Multilingual Audio Deepfake Benchmark,arXiv 2506.00462
- 论文链接:https://arxiv.org/abs/2506.00462
- 引用格式: bibtex @article{xmadbench2025, title = {{XMAD-Bench: Cross-Domain Multilingual Audio Deepfake Benchmark}}, journal = {arXiv preprint arXiv:2506.00462}, year = {2025}, }
快速使用
python from datasets import load_dataset
ds = load_dataset("SpeechAntiSpoofingBenchmarks/XMAD", split="test") print(ds[0])
维护者
- 维护人:Kirill Borodin (SpeechAntiSpoofingBenchmarks)
- 邮箱:kborodin.research@gmail.com (旧邮箱已弃用)
- Telegram:@korallll_ai
评估与提交
评估说明和提交格式请参考数据集中的 submissions/README.md 文件。




