Automatic Distillation of Musical Structures: Learning the Grammar of Music

This paper explores the possibility of extracting the structures underlying music sequences in an unsupervised manner using similar methods as those used for analyzing natural language. We apply ADIOS, a grammar induction algorithm shown to perform well on natural language corpora, to a corpus of Israeli folk songs, exploring several possible textual representations of music. We show that one of these representations allows the algorithm to infer rich and complex structure, which is shown to be 'harmonically-meaningful'. It is our hope that the fact that the same algorithm can infer both natural language grammar and musical structure will shed light on the similarity between the two domains.