Word embedding models applied to classical music recover the circle of fifths in embedding space

Jaan Altosaar

Abstract:

We apply a word embedding model to a large
symbolic corpus of classical music to learn an
embedding space where chords are represented
by real-valued vectors. In early classical music,
the first two principal components of the embeddings of major triads form a circle. In music from
later composers, this circular topology is less evident. The order in which major triads are arranged on this structure corresponds to their order in the circle of fifths. The emergence and
perturbation of this structure is justified by reasoning about the probabilistic embedding model
and stylistic trends in the composition of classical music. We show how this technique is useful
for large-scale, quantitative stylistic analysis of
music, and musical document similarity in general, by using our learned embeddings and the
word-mover’s distance (Kusner et al., 2015) to
classify composers