Generating a musical accompaniment from a given melody is challenging since AI must produce matching notes from only a few starting notes. Standard models often need a complete sequence to work properly.
Anticipation creates an interleaved structure that places each provided note close to the notes it is meant to accompany. This lets a single model learn to predict the next note while following the given melody.
The model matched the quality of standard music generators for generating from scratch, while also learning to fill in missing notes or add accompaniments. Human listeners found its accompaniments similar to human-composed music over 20-second clips. This controllability allows musicians to guide the AI rather than leaving it fully automatic.