Detection of Note Onsets From EEG While Listening to Music

This paper proposes an approach to predicting the onsets of notes in music from electroencephalogram (EEG) signals. Participants listened to 45 kinds of single-tone melodies produced by piano sounds set on the same tempo. Training labels (onset or not-onset) were given by 100 ms using the scores of the melodies. An EEG while listening music was divided into segments with a window width of 500 ms and an overlap of 100 ms. Then, we solve the classification problems using logistic regression (LR) or support vector machine (SVM). We report that five out of fourteen participants' areas under the curve (AUC) indicated more than 0.7. Furthermore, when the predicted onset sequence was used to predict the musical stimulus being listened to, the maximum classification accuracy was 91.7%. These results suggest that each note can be decoded from brain response. The proposed approach can measure brain responses to each note or adapted for brain-computer interface (BCI) using natural music.