AI Insight
This study presents a novel approach for training audio autoencoders by combining noise-augmented encoding with perceptually motivated loss functions, resulting in representations that align with human perceptual hierarchies. The researchers demonstrate that their method captures perceptually important musical information in coarser representational structures compared to conventional training approaches. When applied to latent diffusion models, this technique achieves superior performance in predicting pitch surprisal in music and forecasting EEG brain responses during music listening, outperforming existing methods.
Why it matters
This work advances the field of computational musicology and neuroscience by creating machine learning models that better mirror human perceptual processing of music. The improved ability to predict brain responses and musical expectations could enhance applications in music recommendation systems, therapeutic interventions using music, and brain-computer interfaces.
Understand the Science
arXiv:2511.05350v3 Announce Type: replace-cross
Abstract: We argue that training autoencoders to reconstruct inputs from noised versions of their encodings, when combined with perceptually motivated losses, yields encodings that are structured according to a perceptual hierarchy. We demonstrate the emergence of this hierarchy by showing that, after training an audio autoencoder in this manner, perceptually salient information is captured in coarser representation structures than with conventional training. Furthermore, we show that such perceptual hierarchies improve latent diffusion decoding in the context of estimating pitch surprisal in music and predicting EEG-brain responses to music listening. In both cases, our results surpass those of previous methods. Pretrained weights are available on github.com/CPJKU/pa-audioic.
Source: Perceptually Aligning Representations of Music via Noise-Augmented Autoencoders