The paper discusses the use of modern deep probabilistic generative models, specifically a Diffusion Model, to compute the likelihood of a musical input sequence. The model learns complex non-linear features directly from a training set. The study found that these models are able to more accurately represent music surprisal for human listeners. The results showed a negative quadratic relationship between musical surprisal values and subject ‘liking’ ratings. This research presents a step towards developing advanced models of music expectation and subjective likability.
Publication date: 5 Oct 2023
Project Page: https://arxiv.org/abs/2310.03500v1
Paper: https://arxiv.org/pdf/2310.03500
