GlobalDISCO
收藏资源简介:
GlobalDISCO是一个大规模的生成音乐数据集,包含来自世界各地的音乐传统,旨在探索音乐生成模型中的潜在偏见,并解决生成音乐领域缺乏大型、多文化和多语言数据集的问题。该数据集由来自79个国家的9.3万首真实音乐和73万首生成音乐组成,涵盖了147种语言和991种音乐风格。音乐曲目是由四种最先进的商业音乐生成模型生成的,包括Udio、Suno、Mureka和Riffusion。数据集的构建过程涉及从MusicBrainz和Wikipedia收集艺术家信息,匹配LAION-DISCO-12M中的参考曲目,并根据这些信息构建艺术家档案。音乐风格描述和合成歌词是根据艺术家档案生成的,然后使用音乐生成模型生成音乐。该数据集被设计用于评估音乐生成模型在地理区域和音乐风格上的偏差和多样性,并支持研究界在音乐生成中识别和解决偏见,促进未来模型开发中的更大全球多样性。
GlobalDISCO is a large-scale generative music dataset encompassing musical traditions from across the globe, designed to explore latent biases in music generation models and address the shortage of large-scale, multi-cultural and multi-lingual datasets in the music generation field. The dataset comprises 93,000 authentic music tracks and 730,000 generated music tracks originating from 79 countries, covering 147 languages and 991 musical genres. The musical tracks were generated by four state-of-the-art commercial music generation models, namely Udio, Suno, Mureka and Riffusion. The dataset construction pipeline involves collecting artist information from MusicBrainz and Wikipedia, matching reference tracks within the LAION-DISCO-12M dataset, and building artist profiles based on the collected data. Musical style descriptions and synthetic lyrics are generated based on the artist profiles, which are then used to produce music via the aforementioned models. This dataset is designed to evaluate the biases and diversity of music generation models across geographic regions and musical styles, support the research community in identifying and mitigating biases in music generation, and foster greater global diversity in future model development.




