Neural dynamics of music listening reflect predictions embedded in deep generative networks

Poster Session E - Monday, March 9, 2026, 2:30 – 4:30 pm PDT, Fairview/Kitsilano Ballrooms

Arun Asthagiri¹, Psyche Loui¹; ¹Northeastern University

Predictive coding theory posits that the brain summarizes features across multiple timescales by comparing bottom-up sensory signals against top-down prediction signals. While recent approaches in language processing investigate predictions in the brain by comparing neural activity to embeddings in deep neural networks, less is known about how these predictions operate over continuous acoustic signals, and even less is known about how they play a role in music. Here, we test whether and how neural activity reflects continuous acoustic predictions during music listening by comparing human EEG activity to representations within MusicGen (Copet et al., 2023), a multi-stage transformer model trained on autoregressive acoustic prediction. Across two naturalistic music listening datasets collected in our lab (n=26; n=38), we tested whether including MusicGen’s hidden state representations in ridge regression improved predictive accuracy on withheld EEG data over and above acoustics (lagged samples of cochlear-filtered audio). Parametric and non-parametric tests revealed significant improvements in encoding accuracy when including hidden state activity over acoustics (p<.005). After filtering neural and hidden state signals over a range of frequencies, we found that corresponding timescales in the delta, theta, and gamma bands contributed to strongest encoding improvements, suggesting nested timescales of prediction in the brain. To interpret MusicGen’s representations, we performed frequency-based analyses of its hidden state activity over time, which revealed that nested oscillations emerge within the model at multiples of the musical beat purely from predictive constraints. Together, our results reveal how the time-varying aspects of musical prediction underlie the neural dynamics of music listening.

Topic Area: PERCEPTION & ACTION: Audition