MusicCaps
收藏资源简介:
MusicCaps数据集是一个用于音乐生成模型评估的基准数据集,包含5521个音乐描述提示,以及由不同音乐生成模型生成的音乐样本。数据集旨在支持对音乐生成模型进行透明、可重复和以人为中心的评估,以更好地反映人类审美判断。该数据集由伦敦玛丽女王大学数字音乐中心的研究团队创建,并通过比较实验评估了五个最先进的音乐生成模型,包括JASCO、Stable-Audio-Open、MusicGen、YuE和DiffRhythm,以了解不同评价方法的差异和偏差。研究结果表明,JASCO在内容有用性和制作质量方面表现出色,而DiffRhythm在制作复杂性方面表现突出。该数据集的发布旨在推动对生成模型进行更全面和系统的评估,以更好地反映人类偏好。
MusicCaps is a benchmark dataset for evaluating music generation models. It encompasses 5521 music description prompts and music samples generated by diverse music generation models. The dataset is developed to support transparent, reproducible, and human-centric evaluation of music generation models, thereby better reflecting human aesthetic judgments. This dataset was created by a research team from the Centre for Digital Music at Queen Mary University of London. Via comparative experiments, the team evaluated five state-of-the-art music generation models including JASCO, Stable-Audio-Open, MusicGen, YuE, and DiffRhythm, to investigate the discrepancies and biases across different evaluation methodologies. The research findings demonstrate that JASCO excels in both content usefulness and production quality, while DiffRhythm stands out in terms of production complexity. The release of the MusicCaps dataset is intended to facilitate more comprehensive and systematic evaluation of music generative models, ultimately better aligning with human preferences.

- 1From Aesthetics to Human Preferences: Comparative Perspectives of Evaluating Text-to-Music Systems伦敦玛丽女王大学数字音乐中心, 英国 · 2025年



