CD-ADD
收藏资源简介:
CD-ADD(跨领域音频深度伪造检测)是一个专为语音反欺骗和合成语音检测任务设计的基准数据集。该数据集将真实的人类语音与五种先进的零样本文本转语音系统(OpenVoice、VALL-E、WhisperSpeech、YourTTS、Seamless)生成的深度伪造语音进行配对。数据来源于LibriTTS的test-clean子集以及CD-ADD发布的所有TED演讲内容。核心任务为二分类:区分真实语音(bonafide)与伪造语音(spoof)。数据集包含20,786个音频样本,其中3,661个为真实样本,17,125个为伪造样本,总时长约58.4小时。每个样本包含以下字段:音频路径(path)、16 kHz单声道音频数据(audio)、分类标签(label,0为真实,1为伪造)以及包含唯一话语ID、数据来源(libritts/ted)、生成系统、说话人ID和可用文本转录等元数据的JSON字符串(notes)。数据集采用Creative Commons Attribution 4.0 International (CC BY 4.0) 许可证,音频以无损FLAC格式编码。主要评估指标为等错误率(EER)。
CD-ADD (Cross-Domain Audio Deepfake Detection) is a benchmark dataset designed for speech anti-spoofing and synthetic speech detection tasks. It pairs genuine human speech with deepfake speech generated by five advanced zero-shot text-to-speech systems (OpenVoice, VALL-E, WhisperSpeech, YourTTS, Seamless). The data sources include the test-clean subset of LibriTTS and all TED talk content released by CD-ADD. The core task is binary classification: distinguishing between bonafide (real) speech and spoof (fake) speech. The dataset contains 20,786 audio samples, with 3,661 being genuine and 17,125 being fake, totaling approximately 58.4 hours. Each sample includes the following fields: audio path (path), 16 kHz mono-channel audio data (audio), classification label (label, where 0 is genuine and 1 is fake), and a JSON string (notes) containing metadata such as unique utterance ID, data source (libritts/ted), generation system, speaker ID, and available text transcription. The dataset is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license, with audio encoded in lossless FLAC format. The primary evaluation metric is the Equal Error Rate (EER).
数据集名称
CD-ADD(Cross-Domain Audio Deepfake Detection)
许可协议
Creative Commons Attribution 4.0 International (CC BY 4.0)
语言
英语
任务类别
音频分类(二分类:真实语音 vs. 伪造语音)
数据集规模
样本总数:20,786(训练/测试未明确,提供 test 拆分)
- bonafide(真实):3,661 条
- spoof(伪造):17,125 条
- 总时长:约 58.4 小时
数据来源与构成
- 源自 CD-ADD 发布版本,包含 LibriTTS
test-clean子集及所有 TED Talks 内容。 - 每个真实语音片段对应至多 5 个由零样本文本转语音(TTS)系统生成的伪造样本。
- 伪造系统包括:OpenVoice、VALL-E、WhisperSpeech、YourTTS、Seamless。
- 所有音频均为原始 16 kHz 单声道信号,无损重新编码为 FLAC(比特精确)。
数据字段
| 字段 | 类型 | 描述 |
|---|---|---|
| path | 字符串 | 数据集内唯一路径(例如 dataset_LibriTTS/test-clean/2300/131720/2300_131720_000002_000001/valle.wav) |
| audio | 音频(16kHz 单声道) | 嵌入的 16 kHz 单声道音频 |
| label | 类别标签(bonafide, spoof) | 0 = bonafide,1 = spoof |
| notes | 字符串(JSON 格式) | 包含唯一 utterance_id、来源(libritts 或 ted)、系统(real/openvoice/valle/whisperSpeech/yourTTS/seamless)、说话者 ID 以及可用时的文本转录 |
配置与加载
- 配置名称:
default - 数据文件:
data/test-*.parquet - 加载方式: python from datasets import load_dataset ds = load_dataset("SpeechAntiSpoofingBenchmarks/CD-ADD", split="test")
评估指标
主要指标为等错误率(EER,%),值越低越好。具体评估配置参见 eval.yaml 和 submissions/README.md。
引用
原始论文:https://arxiv.org/abs/2404.04904 bibtex @article{li2024crossdomain, title = {Cross-Domain Audio Deepfake Detection: Dataset and Analysis}, author = {Li, Yuang and Zhang, Min and Ren, Mengxin and Ma, Miaomiao and Wei, Daimeng and Yang, Hao}, journal = {arXiv preprint arXiv:2404.04904}, year = {2024} }
维护方
SpeechAntiSpoofingBenchmarks(联系邮箱:k.n.borodin@mtuci.ru)




