Surprise Estimation in Music

Description

The presentation will focus on Bjare's work on latent autoregressive diffusion models for computationally estimating experienced "expectedness" and "surprise" (surprisal) in music listening, presented at Neurips 2025 - AI for Music Workshop. We revisit an established connection between music appreciation during listening and the extent to which humans or autoregressive models can predict musical continuations. We visit Music2Latent, an open-source, computationally efficient audio codec used as the audio representation on which surprisal is modeled. We review GPT-style autoregressive diffusion models and show how they are suitable for surprisal estimation. We discuss the prediction of EEG responses to music listening based on our surprise estimates.

Period	Jan 2026
Event title	71st Deep Learning Meetup: Agentic RAG / Music Surprise Estimation: Surprise Estimation in Music
Event type	Conference
Degree of Recognition	National