This repository provides a singing annotation dataset that records vocal information in pop songs. It mainly labels pitch, duration, lyric, onset, and offset of each musical note. Meanwhile, all the information is recorded in a hierarchical structure.
Two kind of annotations are offered: enhanced LRC and MIDI. The enhanced LRC annotations are recommended because the singing information is recorded in a hierarchical structure.
The enhanced LRC files are in the "enhanced_lrc" folder. Each line in an enhanced LRC file records the vocal information of a music phrase. Each line is in the format:
[phrase time tag]<onset time tag>lyric pitch duration{offset time tag}<onset time tag>lyric pitch duration{offset time tag}...<onset time tag>lyric pitch duration{offset time tag}
"read_enhanced_lyric.py" can be used to read the annotations.
The annotation MIDIs are also provided in the "midi" folder.
The midi labels are initialized by music notation and LRC files. The corrected musical notation and LRC files are in folder "notation" and "lrc".
"initialize_label.py" can be used to create coarse labels.
All the labels are calibrated via a manually process. Annotators correct the time tags in the LRC files to calibrate the whole song.
Meanwhile, we provide a method that directly converting the enhanced LRC files to MIDI files. "elrc2midi.py" can be used to execute this process.
The youtube links of all the raw audio are recorded in "youtubeLinks.txt".
"download.py" can be used to get the raw audio.