Anonymousv22222/MuseBench-part1

MuseBench是一个用于区分人类创作音乐与AI生成音乐的样本数据集，用于人类vs.AI音乐分类的基准测试。该数据集将真实人创作音乐与来自10种不同文本到音乐/音乐生成系统的AI生成对应音乐配对，并对AI端进行8种扰动/编解码器变体增强，以支持鲁棒性基准测试。数据集还包括分布外（OOD）人类音乐分割和AI-人类共同创作分割。具体数据来源包括：人类音乐（分布内）来自JamendoMaxCaps数据集（约9,945条轨道），人类音乐（分布外）来自Free Music Archive 2008-2010（约8,673条轨道），AI生成音乐来自8种生成器（如MusicGen、MAGNeT、AudioLDM、JASCO、Magenta、Mustango、Riffusion、Stable Audio Open）以及两种独立生成器（Suno和Udio）。此外，所有AI轨道还提供了8种音频扰动变体（如AAC重新编码、MP3转码、均衡化滤波等），用于检测器鲁棒性评估。数据集结构按目录组织，便于对齐人类与AI配对轨道。数据集旨在用于训练和评估人类vs.AI音乐检测器、基准测试多生成器鲁棒性、鲁棒性测试（对抗编解码重新编码、均衡化、响度、音高/时间拉伸、加性噪声和PGD对抗扰动）、评估AI-人类共同创作音乐的“灰色区域”检测器、研究标题条件生成质量以及通过OOD-FMA分割分析分布偏移。数据集整体基于CC BY-SA 3.0许可证发布，但各组件音频遵循其原始许可证（如Jamendo音频使用Creative Commons许可证，AI生成音频遵循各生成器模型许可证）。MuseBench is a sample dataset for distinguishing human vs. AI-generated music, used for benchmarking human-vs-AI music classification. It pairs real human-made music with AI-generated counterparts from 10 different text-to-music/music-generation systems, and augments the AI side with 8 perturbation/codec variants for robustness benchmarking. The dataset also includes an out-of-distribution (OOD) human-music split and an AI-human co-created split. Specific data sources include: human music (in-distribution) from the JamendoMaxCaps dataset (~9,945 tracks), human music (out-of-distribution) from the Free Music Archive 2008-2010 (~8,673 tracks), AI-generated music from 8 generators (e.g., MusicGen, MAGNeT, AudioLDM, JASCO, Magenta, Mustango, Riffusion, Stable Audio Open) and two standalone generators (Suno and Udio). Additionally, all AI tracks are provided with 8 audio perturbation variants (e.g., AAC re-encoding, MP3 transcoding, equalization filtering) for detector robustness evaluation. The dataset is structured by directories to facilitate alignment of human-AI paired tracks. It is intended for training and evaluating human-vs-AI music detectors, benchmarking robustness across multiple generators, robustness testing (against codec re-encoding, EQ, loudness, pitch/time stretching, additive noise, and PGD adversarial perturbation), evaluating detectors on gray-zone AI-human co-created music, studying caption-conditioned generation quality, and analyzing distribution shift via the OOD-FMA split. The dataset as a whole is released under CC BY-SA 3.0, but individual audio components retain their original licenses (e.g., Jamendo audio uses Creative Commons licenses, AI-generated audio follows each generators model license).