UMD-600MB: Refined MIDI Dataset for Symbolic Music Generation
收藏NIAID Data Ecosystem2026-05-02 收录
下载链接:
https://zenodo.org/record/14560335
下载链接
链接失效反馈官方服务:
资源简介:
UMD-600MB
The Universal MIDI Dataset 600MB (UMD-600MB) is a proprietary collection of 149,230 MIDI files curated for research and development within our organization. This collection is a subset sampled from a larger dataset developed for pretraining symbolic music models.
The field of symbolic music generation is constrained by limited data compared to language models. Publicly available datasets, such as the Lakh MIDI Dataset, offer large collections of MIDI files sourced from the web. While the sheer volume of musical data might appear beneficial, the actual amount of valuable data is less than anticipated, as many songs contain less desirable melodies with erratic and repetitive events.
The UMD-600MB employs an attention-based approach to produce more desirable outputs by focusing on human-reviewed training examples of single-track melodies, chord progressions, leads, bass, and arpeggios, each averaging 8 bars in duration. The dataset has been continuously refined since 2022 to ensure consistent quality and tempo alignment for further development in AI music generation projects. Additionally, the dataset is normalized by setting the timing information to 120 BPM with a tick resolution (PPQ) of 96.
Melody Styles
A major portion of the dataset is composed of newly produced private data to represent modern musical styles.
Pop: 1970s to 2020s Pop music
EDM: Trance, House, Hardstyle, Dance, Synthwave, Retro, Arcade
Jazz: Bebop, Ballad, Latin-Jazz, Bossa-Jazz, Blues, Ragtime, World
Rock: Pop, Alternative, Folk
Soul: Classic, Modern, Neo-Soul, Funk, Latin-Soul
Urban: Pop, Hip-Hop, Lo-Fi, Trap, R&B, Afrobeat
World: Latin, Bossa Nova, European
Other: Film, Cinematic, Game music and piano references
Actual MIDI files are unlabeled for unsupervised training.
Dataset Access
Please note that this is a closed-source dataset with very limited access. Considerations for access include proposals for data augmentation, chord extraction and other enhancement methods, whether through scripts, algorithmic techniques, manual editing in a DAW or additional processing methods.
For inquiries about this dataset, please email us.
通用MIDI数据集600MB(Universal MIDI Dataset 600MB,缩写UMD-600MB)是我司为内部研发工作甄选整理的专属馆藏数据集,共收录149,230个乐器数字接口(MIDI)文件。该数据集源自为预训练符号音乐模型所构建的大型数据集的采样子集。
与语言模型领域相比,符号音乐生成方向受限于可用数据规模不足。当前公开可用的数据集(如Lakh MIDI数据集)提供了大量从网络抓取的MIDI文件合集。尽管海量音乐数据看似具备应用优势,但实际可用的优质数据却远低于预期——许多曲目存在旋律欠佳、节奏紊乱且重复冗余的问题。
UMD-600MB采用基于注意力机制的筛选方法,聚焦经人工审核的单轨旋律、和弦进行、主奏声部、低音声部与琶音片段作为优质训练样本,以此生成更符合要求的输出结果。上述单轨样本平均时长为8小节。本数据集自2022年起持续优化,以保障数据质量统一与节拍对齐,适配AI音乐生成项目的后续研发需求。此外,数据集已完成标准化处理:将时序信息统一校准为120拍每分钟(BPM, Beats Per Minute),节拍分辨率(PPQ, Pulses Per Quarter Note)设为96。
旋律风格
本数据集的核心部分由全新制作的私有数据构成,用以覆盖各类现代音乐风格:
- 流行音乐(Pop):涵盖1970年代至2020年代的流行乐作品
- 电子舞曲(EDM):包含出神舞曲(Trance)、浩室音乐(House)、硬派电子(Hardstyle)、舞曲(Dance)、合成器浪潮(Synthwave)、复古风(Retro)与街机风(Arcade)等子风格
- 爵士乐(Jazz):涵盖比波普(Bebop)、抒情爵士(Ballad)、拉丁爵士(Latin-Jazz)、波萨爵士(Bossa-Jazz)、蓝调(Blues)、拉格泰姆(Ragtime)与世界爵士(World Jazz)
- 摇滚乐(Rock):包含流行摇滚、另类摇滚与民谣摇滚
- 灵魂乐(Soul):涵盖经典灵魂乐、现代灵魂乐、新灵魂乐(Neo-Soul)、放克(Funk)与拉丁灵魂乐(Latin-Soul)
- 都市音乐(Urban):包含流行乐、嘻哈(Hip-Hop)、低保真(Lo-Fi)、陷阱音乐(Trap)、节奏布鲁斯(R&B)与非洲节拍(Afrobeat)
- 世界音乐(World):涵盖拉丁音乐、波萨诺瓦(Bossa Nova)与欧洲本土音乐
- 其他风格:包含影视原声、电影配乐、游戏配乐与钢琴改编参考曲目
本数据集的MIDI文件均未标注,适用于无监督训练场景。
数据集获取方式
请注意,本数据集为闭源数据集,获取权限极为有限。申请获取权限需提交相关研究提案,例如数据增强、和弦提取及其他优化方法的研究方案,可通过脚本编写、算法技术、数字音频工作站(DAW, Digital Audio Workstation)手动编辑或其他后续处理手段开展相关研究。
若需咨询本数据集相关事宜,请发送邮件联系我们。
创建时间:
2024-12-27



