gulaerchen.github.io

MIR-SingerSeparation Dataset & Auto-selection

Dataset Description

The dataset consists of 3 types (Please refer to paper to introduce various categories.) of singer separation datasets, each track 10 seconds long, segments from 476 English and 500 Chinese songs, and male/female vocalist ratio for English songs was 269:207, while that for Chinese songs was 223:277. The tracks are all 8kHz Mono 16-bit audio files in .wav format.

Versions

1.0.0 (default): No release notes.

System Demo

Singer Separation

Download datasets

• EN-D
• CH-D
• EN-S

Download size

• EN-D: 17.0 GB
• CH-D: 8.25 GB
• EN-S: 11.6 GB

Dataset size

• EN-D: 22 GB
• CH-D: 12 GB
• EN-S: 15 GB

Splits

Auto Selection

• Pitch Data

  python auto_selection.py 
  

Citation

Please cite the paper to use the dataset.

@misc{chen2021singer,
    title={Singer separation for karaoke content generation},
    author={Hsuan-Yu Chen and Xuanjun Chen and Jyh-Shing Roger Jang},
    year={2021},
    eprint={2110.06707},
    archivePrefix={arXiv},
    primaryClass={cs.SD}
}