Reproduce “EEG2Mel: Reconstructing Sound from Brain Responses to Music” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Adolfo G. Ramirez-Aristizabal, Chris Kello
Researchers at the University of California Merced developed a deep learning method using a Convolutional Neural Network to reconstruct music directly from human EEG responses. The approach achieved an 85% success rate in a perceptual task where human listeners identified the original songs from the brain-derived audio, demonstrating improved fidelity over previous methods.
Key insights
- Using Power Spectral Density as EEG input and mel-spectrograms as music targets provides the most effective representation for preserving semantic information in music reconstruction.
- A relatively straightforward CNN regressor can effectively reconstruct perceptually identifiable music from naturalistic, long-duration brain responses without complex multi-view inputs or extensive feature engineering.
- Despite some quantitative differences in lower frequency reconstruction, the overall acoustic information recovered is sufficient for human listeners to reliably identify the original musical pieces.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.