LAION · REPRODUCIBLE AUDIO RESEARCH
Singing Voice Data Lab
From a song to a lead voice, melody notes and MIDI tokens.
ROSVOT + smoothed pitch splits + gap filling
Listen to the voice and its piano melody, then compare methods. The default adds stable Procedural pitch segments to ROSVOT timing, smooths short fragments and fills missing notes.
These are medians from the broader 75-recording exploratory benchmark, against shared automatic RMVPE pitch, not human-verified accuracy. This public player contains 25 licensed LeVo2 examples. No listening winner is established.
Loading the public clip index…
Notes and continuous pitch
Inspect complete notes, MIDI tokens and downloads
| Pitch | Name | Onset (s) | Offset (s) | Duration (s) |
|---|
Complete reversible MIDI tokens · 10-ms grid
How to read the metrics and choose a method
Coverage asks whether any note exists during reference-voiced time; a wrong note counts. Within 50 cents requires one unambiguous note within half a semitone, with missing predictions earning no credit. Pitch difference is measured only on jointly covered, unambiguous frames and excludes missing material. Read all three together.
ROSVOT and Procedural share RMVPE. The Procedural baseline is built from the scoring pitch track, so agreement is biased. Procedural alone covers less singing time and scores more strongly on the frame-F1 diagnostic. There is no universal quality winner.
The selected split rule needs a stable 90-ms Procedural pitch run, absorbs sub-50-ms fragments within each ROSVOT parent note, and adds uncovered Procedural fragments of at least 90 ms. It creates more boundaries: listen for unwanted fragmentation. The 69-rule search used a 15/60 internal selection/reporting split on previously examined data, not independent external validation.