SONICS

该数据集是一个为端到端合成歌曲检测（SSD）而设计的新型数据集，包含了超过97,000首歌曲（总计4,751小时），其中超过49,000首歌曲来自各大流行平台。该数据集还包括了全面的评估，涵盖了有效集和测试集，这些集合包含了未见过的算法和新歌手，确保了数据之间不存在泄露。其规模达到了97,000首歌曲（4,751小时），任务是对合成歌曲进行检测。

This is a novel dataset designed specifically for end-to-end synthetic song detection (SSD). It comprises over 97,000 songs, with a total duration of 4,751 hours, and more than 49,000 of these songs are sourced from major mainstream music platforms. The dataset also includes comprehensive evaluation splits: the validation set and the test set, which consist of content generated by unseen song synthesis algorithms and from new artists, thereby eliminating any potential data leakage. With a scale of 97,000 songs (4,751 hours), its core task is synthetic song detection.

搜集汇总

背景与挑战

背景概述

SONICS是一个专注于检测AI生成歌曲的大规模数据集，包含超过97k首歌曲(4,751小时)，其中49k首为合成歌曲，来自Suno和Udio等平台。该数据集特别关注长时歌曲和音乐-歌词多样性，并提供了详细的元数据和模型性能指标，填补了现有合成歌曲检测数据集的空白。