slakh2100
收藏资源简介:
Slakh2100(Synthesized Lakh)是一个大规模多轨音乐数据集,包含2,100首自动混音的曲目,每首曲目均带有分离的乐器音轨和对齐的MIDI文件。该数据集由西北大学的Manilow等人(2019年)创建,使用来自Lakh MIDI数据集的MIDI文件,通过专业级VST乐器渲染生成逼真的多轨音频。每首曲目包含每种乐器的独立音轨(如鼓、贝斯、吉他、钢琴、弦乐等),使Slakh成为音乐源分离研究的标准基准数据集。数据集采用无损FLAC压缩格式,总大小约为105GB。数据集分为训练集(1,500首)、验证集(375首)和测试集(225首),每首曲目包含完整的混音、独立音轨、MIDI文件和元数据文件。音频格式为44.1kHz立体声FLAC。数据集适用于音乐源分离、乐器识别和MIDI到音频的合成等任务。尽管使用了专业VST乐器,但音频仍为MIDI合成,缺乏真实录音的声学细节和表现力。数据集不包含人声音轨,且偏向于流行/摇滚音乐风格。数据集采用CC-BY 4.0许可,允许商业使用。
Slakh2100 (Synthesized Lakh) is a large-scale multi-track music dataset containing 2,100 automatically mixed tracks, each paired with separated instrumental audio tracks and aligned MIDI files. Developed by Manilow et al. (2019) from Northwestern University, this dataset leverages MIDI files sourced from the Lakh MIDI Dataset to render realistic multi-track audio using professional-grade VST instruments. Each track includes independent audio tracks for all featured instruments (e.g., drums, bass, guitar, piano, strings, etc.), positioning Slakh as a standard benchmark dataset for music source separation research. The dataset is stored in lossless FLAC compression format, with a total size of approximately 105 GB. It is partitioned into three subsets: a training set (1,500 tracks), a validation set (375 tracks), and a test set (225 tracks). Every track contains the full mixed audio, individual instrument tracks, MIDI files, and metadata files, with audio formatted as 44.1 kHz stereo FLAC. This dataset supports tasks including music source separation, instrument recognition, and MIDI-to-audio synthesis. However, despite employing professional VST instruments, all audio is MIDI-synthesized, lacking the acoustic details and expressive nuances of real-world recorded music. The dataset excludes vocal tracks and is biased towards pop/rock musical styles. It is licensed under CC-BY 4.0, permitting commercial usage.



