What it is
Open-source singing corpus with 2,300+ hours of audio and unified lyric-note-score annotations for score-native singing synthesis. 2,317 hours and 776k singing segments from 4 public sources, mixing synthetic and real songs with unified lyrics, MIDI pitch, note values, lyric-to-note alignment, tempo, and audio.
Key results
- 2,317 hours across 776k segments
- 4 sources: synthetic + real singing
- MIT license; complete repo is ~110.4 GB
Useful for training and evaluating score-conditioned singing voice synthesis with lyric-to-note alignment.