Understanding how the brain transforms low-level acoustic information into abstract musical representations remains a central challenge in music neuroscience. We present a dynamic representational similarity analysis (dRSA) framework that relates hierarchical representations learned by a generative deep neural network (DNN) to whole-brain fMRI responses during naturalistic music listening. Brain–DNN correspondence was characterised by two complementary measures: Prominence, reflecting the overall strength of representational alignment, and Abstraction, reflecting the tendency for alignment to increase towards deeper DNN layers. Prominence was greatest in auditory and sensorimotor regions, whereas Abstraction peaked in hippocampal, parahippocampal, basal ganglia, and prefrontal regions. Moreover, representational abstraction increased systematically along both auditory processing pathways, indicating hierarchical transformations from sensory encoding to increasingly abstract representations. These findings suggest that generative DNNs provide a useful systems-level model of music representation in the brain and illustrate how dRSA can reveal distributed representational organisation beyond local feature-based mappings.
