On the Relationship Between Entropy and Meaning in Music: An Exploration with Recurrent Neural Networks

Meyer (1956) postulated that meaning in music is directly related to entropy–that high entropy (uncertainty) engenders greater subjective tension, which is correlated with more meaningful musical events. Current statistical models of music are often limited to music with a single melodic line, impeding wider investigation of Meyer’s hypothesis. I describe a recurrent neural network model which produces estimates of instantaneous entropy for music with multiple parts and use it to analyze a Haydn string quartet. Features found by traditional analysis to be related to tension are shown to have characteristic signatures in the model’s entropy measures. Thus, an information-based approach to musical analysis can elaborate on traditional understanding of music and can shed light on the more general cognitive phenomenon of musical meaning.