Abstract
We explore the potential of a popular distributional semantics vector spacemodel, word2vec, for capturing meaningful relationships in ecological (complexpolyphonic) music. More precisely, the skip-gram version of word2vec is used tomodel slices of music from a large corpus spanning eight musical genres. Inthis newly learned vector space, a metric based on cosine distance is able todistinguish between functional chord relationships, as well as harmonicassociations in the music. Evidence, based on cosine distance betweenchord-pair vectors, suggests that an implicit circle-of-fifths exists in thevector space. In addition, a comparison between pieces in different keysreveals that key relationships are represented in word2vec space. These resultssuggest that the newly learned embedded vector representation does in factcapture tonal and harmonic characteristics of music, without receiving explicitinformation about the musical content of the constituent slices. In order toinvestigate whether proximity in the discovered space of embeddings isindicative of `semantically-related' slices, we explore a music generationtask, by automatically replacing existing slices from a given piece of musicwith new slices. We propose an algorithm to find substitute slices based onspatial proximity and the pitch class distribution inferred in the chosensubspace. The results indicate that the size of the subspace used has asignificant effect on whether slices belonging to the same key are selected. Insum, the proposed word2vec model is able to learn music-vector embeddings thatcapture meaningful tonal and harmonic relationships in music, thereby providinga useful tool for exploring musical properties and comparisons across pieces,as a potential input representation for deep learning models, and as a musicgeneration device.