Song Describer Dataset (SDD)

Song Describer Dataset (SDD) 是一个由女王玛丽大学伦敦分校、巴塞罗那庞培法布拉大学音乐技术组和环球音乐集团国际有限公司合作创建的高质量音频-标题对数据集。该数据集包含1106条人类编写的自然语言描述，对应706个音乐录音，所有内容均公开可访问并采用创意共享许可证。SDD数据集特别适用于评估音乐与语言模型的性能，如音乐标题生成、文本到音乐的生成以及音乐-语言检索等任务。数据集的创建过程涉及通过在线注释平台收集描述，并由数据集创建者手动审核以确保质量。SDD数据集的应用领域主要集中在音乐与语言的交叉研究，旨在解决现有数据集在公共可访问性和一致性评估实践方面的不足。

Song Describer Dataset (SDD) is a high-quality audio-title paired dataset co-developed by Queen Mary University of London, the Music Technology Group at Universitat Pompeu Fabra Barcelona, and Universal Music Group International Limited. This dataset includes 1106 human-written natural language descriptions corresponding to 706 music recordings, all of which are publicly accessible and licensed under Creative Commons. The SDD dataset is particularly well-suited for evaluating the performance of music and language models, covering tasks such as music caption generation, text-to-music generation, and music-language retrieval. The creation process of SDD involves collecting descriptions via online annotation platforms, followed by manual quality checks conducted by the dataset developers. The primary application scope of SDD lies in interdisciplinary research at the intersection of music and language, with the aim of addressing the limitations of existing datasets in terms of public accessibility and standardized evaluation practices.