EEG-based cross-subject passive music pitch perception using deep learning models.

Area of Science:

• Neuroscience
• Music Perception
• Machine Learning

Background:

• Pitch is crucial for music perception and melody interpretation.
• Objective methods for decoding brain responses to musical pitch are needed.
• Electroencephalography (EEG) offers a non-invasive approach to measure neural activity.

Purpose of the Study:

• To objectively detect and decode brain responses to musical pitch perception using EEG.
• To develop and evaluate machine learning models for EEG-based pitch classification.
• To identify the optimal time window for brain decoding of pitch perception.

Main Methods:

• Collected EEG data from 34 subjects hearing violin sounds at specific pitches (G3, B6) using a passive Go/No-Go paradigm.
• Developed a lightweight modified EEGNet model for within-subject pitch classification.
• Employed a classifier ensemble (CE) method for cross-subject pitch classification based on within-subject models.

Main Results:

• The modified EEGNet achieved 77% average accuracy for within-subject pitch classification.
• The CE method reached 74% average accuracy for cross-subject pitch classification, significantly above chance (50%).
• The optimal time window for decoding pitch perception from EEG data was identified as 0.4 to 0.9 seconds post-stimulus onset.

Conclusions:

• The proposed EEG-based methods effectively and objectively assess musical pitch perception.
• The models demonstrate generalization ability, particularly the cross-subject classifier ensemble.
• These findings open avenues for objective evaluation of auditory perception in music.