A musical event is a token-like unit that represents a note action or timing information in a performance-based music sequence. Converting music into these events lets a GPT-style Transformer predict the next event from everything that came before it. section 14.1 section 14.0
Reduce performance data to four event types:
- Note-on marks the start of a note and specifies its pitch. section 14.1
- Note-off marks the end of a note. section 14.1
- Time-shift records the elapsed time between two musical events. section 14.1
- Velocity records how forcefully or loudly a note is played. section 14.1
Each event type can take multiple values. Assign every unique event a distinct index, turning a performance into a sequence of indexes rather than the multidimensional representation used by an image-like music model. section 14.1 section 14.5
From performance to generation
A music sequence becomes an ordered stream of note-on, velocity, time-shift, and note-off events, then an indexed token sequence. A decoder-only music Transformer learns patterns and relationships in such sequences and predicts the next musical event; generated events can be converted back into a playable MIDI file. section 14.0 section 14.5
