A musical event is a token-like unit that represents a note action, timing information, or playing force in a music sequence. Treating music as events lets a GPT-style Transformer predict the next event from the preceding sequence, rather than generating a complete multidimensional music object at once. section 14.0 section 14.1
A performance-based representation reduces musical notes to four event types:
- Note-on starts a note and specifies its pitch. section 14.1
- Note-off ends a note and tells the instrument to stop playing it. section 14.1
- Time-shift records the amount of time between two musical events. section 14.1
- Velocity records how forcefully or quickly a note is played; higher values produce stronger, louder sound. section 14.1
From performance to model input
A performance can be represented as a sequence such as note-on, velocity, time-shift, and note-off events. Because each event type has multiple possible values, every unique event maps to an index. The resulting index sequence becomes input for a decoder-only music Transformer, which learns to predict the next event and can later convert generated events into a playable MIDI file. section 14.1 section 14.0
This approach differs from treating music as a 4D object: it follows performance events in time, making the task resemble text generation. Temperature can then regulate the creativity of generated sequences. section 14.1 section 14.5
