Samples

Loading example…

Generated mixture

48 kHz Opus

0:00 / 0:30

MIDI playback

GM preview

Crop notes rendered with FluidR3 GM, with a fixed output gain per crop. Raw velocities are preserved. Held notes are retriggered at the start. Controller effects and room IR are omitted.

0:00 / 0:30

MIDI samples load when you press play.

Piano roll

Loading notes…

Left-edge markers indicate held notes. Diamonds indicate drums. Hover for raw program, pitch and velocity.

Source & generation details

GigaMIDI examples cover five arrangement groups × three density levels. Its Score MIDI / Performance MIDI labels are timing estimates. All SymphonyNet MIDI is classified as score. Original timing estimates remain in the details and downloadable metadata.

Original GigaMIDI and SymphonyNet MIDI is filtered and split, cropped for 30 active seconds with pre-roll, rendered with FluidSynth and random SoundFonts and IR reverb, then saved as aligned audio and annotations.

Sources. A subset of GigaMIDI v2.0.0 (Lee et al., 2025; CC BY-NC 4.0) and all eligible SymphonyNet pieces. Each source is assigned one 30-second crop. Original MIDI programs, velocities and controllers drive the mixture.

Rendering. One of five SoundFonts per crop, plus AIR, ReverbFX or Samplicity room IRs. Wet mix is sampled per song from 15–35%; FIR filters are limited to four seconds. Only mixtures are generated.

Completed datasets · one 30-second crop per source MIDI
DatasetTrainValTestTotal
GigaMIDI Full481,383300300481,983
GigaMIDI Small40,00030030040,600
SymphonyNet45,53030030046,130

Full includes Small's 40,600 crops and shares its validation/test sets. Programs use raw zero-based GM numbers.

Dataset distribution

Open full-size figure ↗
Dataset distributions for note count, instrument count, score and performance MIDI types, arrangement and density groups, GM instrument families, and the seven most and least common programs.

Counts cover all saved crop metadata. GigaMIDI types use timing estimates; SymphonyNet is all score. Program rankings count songs containing each program across the 526,913 Full + SymphonyNet training crops.