AI models that create music often represent notes as a stream, ignoring the regular beats and bars that give music its rhythm. As a result, the generated music can feel off-beat or loose.
To fix this, the team introduced REMI (revamped MIDI-derived events), a new way to represent digital music (MIDI stands for Musical Instrument Digital Interface). REMI explicitly includes events for bar starts, beat positions, and chord changes, so the model knows the musical structure.
Using REMI, their Pop Music Transformer generated piano pieces with much better rhythmic regularity than previous top models. This shows that adding basic music knowledge, like beats and chords, helps AI compose more naturally and expressively.