A score in a relational space
In a traditional score, notes are represented on a staff. Their position indicates their pitch, thereby describing the melody as a sequence of sounds.
Artificial-intelligence models use a different representation. Each element—a word in a text or a note in a melody—is associated with a vector whose position depends on its relationships with other elements. Notes that occur in similar musical contexts therefore occupy nearby positions in a latent space learned from a large number of scores.
This representation is not limited to relationships between notes alone. It also retains certain fundamental musical properties: for example, a note’s pitch can be recovered from its position in this space. The model’s internal geometry therefore carries both relational and referential information.
In the visualisation above, each note is represented by a three-dimensional vector. The score is no longer read solely on a staff: it becomes a trajectory through space, revealing an invisible geometry of the melody.
The video above shows this trajectory for an excerpt from Frédéric Chopin’s Étude Op. 10 No. 3 (“Tristesse”), using a public-domain recording made available by Musopen. It is the same excerpt used for the volumetric and surface representations, as well as in the Musical Garden.
This geometry is not merely a visualisation: it is an internal representation of the music, from which some of its properties can be recovered and new sculptural forms can be generated.