Musical Instrument Timber-evoked EEG Classification Based on Deep Learning

Timber is a major attribute of musical sound perception and an important aspect of the hearing evaluation. With the development of brain science and neuroscience, how to use electroencephalogram (EEG) signals to identify musical instrument timbers raises concerns. In this paper, we studied the classification of EEG signals evoked by musical instrument timber based on deep learning methods. First, an oddball paradigm was designed to collect the auditory EEG signals when musical instrument sounds (piano, violin and trumpet) were played to a normal hearing (NH) subject. Then the spatial and temporal characteristics of the EEG data were analysed and the event-related potentials (ERPs) could be found after about 400 ms. Based on these features, four deep learning models were used to classify the EEG data. The experimental results showed that the two neural network models with spatiotemporal decoding structures had better performance than other two models, and their classification accuracies were all up 80% above. It is concluded that the design of musical instrument sounds-evoked EEG experimental paradigm is reasonable, and the proposed deep learning methods can be used objectively to evaluate the hearing of the subject for musical instrument timber.