AudioSpoof 虚假音频检测数据集
收藏资源简介:
随着TTS(Text-to-Speech)技术的快速发展,当前语音克隆模型生成的声音已难以通过简单听觉判断真伪。然而,针对中文场景的音频伪造检测领域仍存在显著空白:1️⃣ 缺乏基于最新语音合成技术生成的伪造音频数据集(Audio Spoofing Dataset)2️⃣ 现有检测方法对零样本语音克隆攻击的防御能力不足。为此,我们基于 MagicData 中文普通话语料库,通过四大前沿开源TTS模型进行零样本语音克隆: NaturalSpeech3, CosyVoice, F5-TTS, Spark-TTS, 构建首个专注于中文场景的多模型伪造音频检测基准数据集。采用零样本克隆,可以获取较高质量的、较多人数的伪造音频数据集。
With the rapid development of Text-to-Speech (TTS) technology, voices generated by current voice cloning models can hardly be distinguished from genuine speech through casual auditory inspection. However, there remain significant gaps in the field of audio spoofing detection for Chinese scenarios: 1) There is a lack of spoofing audio datasets (Audio Spoofing Dataset) generated using state-of-the-art speech synthesis technologies; 2) Existing detection methods lack sufficient defense capabilities against zero-shot voice cloning attacks. To address these issues, we conducted zero-shot voice cloning with four cutting-edge open-source TTS models: NaturalSpeech3, CosyVoice, F5-TTS, and Spark-TTS, based on the Mandarin Chinese speech corpus from MagicData, and constructed the first benchmark dataset for multi-model audio spoofing detection focused on Chinese scenarios. Adopting zero-shot cloning enables us to obtain spoofing audio datasets with high quality and a large number of speakers.
AudioSpoof 虚假音频检测数据集概述
数据集背景
- 针对中文场景的音频伪造检测领域存在的空白:
- 缺乏基于最新语音合成技术生成的伪造音频数据集
- 现有检测方法对零样本语音克隆攻击的防御能力不足
数据来源
- 基于MagicData中文普通话语料库
- 使用四大前沿开源TTS模型进行零样本语音克隆:
- NaturalSpeech3
- CosyVoice
- F5-TTS
- Spark-TTS
数据集下载
- 托管平台:
- Hugging Face: https://huggingface.co/datasets/HuShou-ZMZN/audiofake
- zenodo: https://zenodo.org/records/15259855
数据构建方法
基础数据源
- 从原始数据集中分层抽样构建核心语料:
- 开发集:2人,每人随机选取10%录音
- 测试集:4人,同比例采样
- 训练集:20人,保持相同采样率
语音克隆流程
- 采用零样本克隆技术生成样本
数据结构
AudioSpoof/
├── metadata/
│ └── SPKINFO.txt
├── wav/
│ ├── dev/
│ ├── test/
│ ├── train/
│ ├── dev-naturalspeech3/
│ ├── test-naturalspeech3/
│ ├── train-naturalspeech3/
│ ├── dev-cosyvoice/
│ ├── test-cosyvoice/
│ ├── train-cosyvoice/
│ ├── dev-F5TTS/
│ ├── test-F5TTS/
│ ├── train-F5TTS/
│ └── dev-sparktts/
│ ├── test-sparktts/
│ └── train-sparktts/
└── text/
├── dev.txt
├── test.txt
└── train.txt
数据统计
| 子集 | 说话人数 | 真实音频数 | 伪造音频数(×4模型) | 总样本数 |
|---|---|---|---|---|
| dev | 2 | 118 | 472 | 590 |
| test | 4 | 180 | 720 | 900 |
| train | 20 | 1105 | 4,420 | 5,525 |
总计:26人 | 1,403真实 | 5,612伪造 | 7,015总样本
关键说明
- 每个克隆模型生成的三组目录(dev/test/train)保持原始音频目录结构
- 语音克隆目录命名统一采用
{子集}-{model_name}格式 - 克隆音频文件保留原始命名,仅通过目录路径区分不同模型的输出




