Single-trial EEG-based classification reveals instrument-specific timbre perception via traditional machine learning...

Many users of hearing aids report challenges when listening to music. In the future, it may be possible to develop hearing aids that monitor brain activity in real-time and adapt their output to the volitions of the user. In music, this could mean selectively amplifying the sound of the instrument the listener wants to hear. The objective of this research is to determine whether traditional machine learning can be used to identify which instrument an individual is listening to based only on single-trial EEG. In this work, participants were presented with a series of brief tones that varied in timbre (Trombone, Clarinet, Cello, Piano and Pure Tone) while their ongoing EEG was recorded from 73 electrodes. To distinguish between EEG responses to the five different musical instruments, we investigated the use of four different classifiers - Linear Discriminant Analysis (LDA), Gradient Boosting (GB), Support Vector Machine (SVM) and k-NN, and four different sets of features - raw EEG, ERP-based features, harmonics-based features and regularity-based features. N1 and P2 components of the ERP were analyzed for differences between instruments. All four classifiers performed significantly above chance (i.e., approximately 20% for 5 classes) when trained using the raw EEG features (LDA: 34%, GB: 35%, SVM: 33%, k-NN: 26%). Precision, Recall and F1-scores closely mirrored overall accuracy. It may be possible to improve these results with more advanced classification algorithms or different transformations of features. Statistical analysis found the Cello to have contributed to the largest P2 amplitude and Pure Tone to the smallest, and for Cello to have contributed to the earliest N1 latency and Clarinet the latest.