EmoFake_test
收藏资源简介:
EmoFake Test是一个专门用于语音反欺骗和音频深度伪造检测的基准测试数据集,专注于情感语音的深度伪造检测任务。该数据集包含真实的情感语音样本(bonafide)以及经过情感转换处理的欺骗性语音样本(spoof),旨在评估模型在区分真实情感表达与伪造情感语音方面的能力。数据集总规模为17,500个样本,其中真实样本3,500个,欺骗样本14,000个。每个样本包含四个字段:音频文件名(path)、16kHz单声道的音频波形数据(audio)、二分类标签(label,指示样本为真实或欺骗)以及JSON格式的详细元数据(notes),其中记录了话语ID、说话人ID、情感类别、生成方法和子集划分等信息。数据集采用CC BY 4.0许可证,基于相关研究论文(arXiv:2211.05363)构建,适用于语音反欺骗模型评估、音频深度伪造检测算法开发以及情感语音合成安全性研究等场景。
EmoFake Test is a purpose-built benchmark dataset for speech anti-spoofing and audio deepfake detection, with a core focus on the task of emotional speech deepfake detection. This dataset encompasses both bona fide emotional speech samples and spoofed speech samples generated via emotional conversion, designed to evaluate models' capability to differentiate between genuine emotional expressions and forged emotional speech. The total size of the dataset is 17,500 samples, consisting of 3,500 bona fide samples and 14,000 spoofed samples. Each sample includes four fields: audio file path (path), 16kHz monaural audio waveform data (audio), binary classification label (label, which indicates whether the sample is bona fide or spoofed), and detailed JSON-formatted metadata (notes) that records information such as utterance ID, speaker ID, emotion category, generation method, and subset partitioning. The dataset is licensed under CC BY 4.0, constructed based on the relevant research paper (arXiv:2211.05363), and is applicable to scenarios such as speech anti-spoofing model evaluation, audio deepfake detection algorithm development, and safety research on emotional speech synthesis.




