DALI

DALI数据集是由法国国家科学研究中心创建的一个大型多模态数据集，包含5358个音频轨道及其同步的声乐旋律音符和歌词，涵盖四个粒度级别。该数据集通过教师-学生机器学习范式自动创建，起始于卡拉OK游戏非专业用户的初步时间对齐歌词和音符的手动标注。数据集的创建过程中，通过网络检索音频候选，并使用深度卷积神经网络进行声乐检测，以匹配和更新时间对齐的歌词。DALI数据集旨在为声乐研究社区提供一个高质量的参考数据集，解决现有数据集规模小和质量不足的问题，适用于音乐信息检索和声乐分析等研究领域。

The DALI Dataset is a large-scale multimodal dataset developed by the French National Centre for Scientific Research (CNRS), which contains 5358 audio tracks alongside their synchronized vocal melody notes and lyrics, covering four granularity levels. This dataset was automatically constructed via a teacher-student machine learning paradigm, starting from manual annotations of temporally aligned lyrics and notes made by amateur users of karaoke games. During the dataset creation process, audio candidates were retrieved from the web, and deep convolutional neural networks were employed for vocal detection to match and update the temporally aligned lyrics. The DALI Dataset aims to provide a high-quality reference dataset for the vocal music research community, addressing the issues of small scale and insufficient quality of existing datasets, and is applicable to research fields including music information retrieval and vocal analysis.