MMGenre is a diagnostic benchmark for studying how singing voice synthesis
(SVS) systems behave across diverse musical genres. This repository provides
the benchmark's symbolic music scores, genre taxonomy, loading utilities, and
analysis examples. It contains no audio.
## Dataset Statistics
MMGenre contains **3,152** score segments across **10** major genres and **26**
subgenres. The names and counts below are derived directly from the score IDs.
| Genre | Segments | Subgenres and segment counts |
|---|---:|---|
| blues | 247 | `delta-blues` (128), `jump-blues` (119) |
| classical | 266 | `art-song` (85), `opera` (181) |
| country | 317 | `bluegrass-country` (134), `classic-country` (183) |
| electronic | 287 | `house` (48), `techno` (155), `trance` (84) |
| jazz | 264 | `bebop-jazz` (115), `swing-jazz` (149) |
| pop | 566 | `ballad-pop` (255), `folk-pop` (203), `synthpop` (108) |
| rap | 246 | `east-coast-gangsta-rap` (92), `old-skool-rap-pioneers` (64), `reggaeton-latin-rap` (90) |
| rnb | 407 | `motown-soul` (119), `neo-soul` (95), `new-jack-swing` (193) |
| rock | 309 | `alternative-rock` (125), `heavy-metal` (59), `punk-rock` (125) |
| world | 243 | `africa` (53), `arab` (105), `india` (85) |
The machine-readable hierarchy is available at
[`data/genre_taxonomy.json`](data/genre_taxonomy.json).