Sound Research WIKINDX
Sound Research WIKINDX |
|
 |
|
Thompson, J. (2013). Neural decoding of subjective music listening experiences. Unpublished thesis Masters, Dartmouth College, New Hampshire. Added by: Mark Grimshaw-Aagaard (04/11/2014, 12:34) |
 |
| Resource type: Thesis/Dissertation BibTeX citation key: Thompson2013a Email resource to friend View all bibliographic details  |
Categories: General Keywords: Music, Neural decoding Creators: Thompson Publisher: Dartmouth College (New Hampshire) |
Views: 8/448
|
| Abstract |
|
How the human auditory system facilitates musical experience is not well understood. This thesis considers a computational investigation of the neural basis of musical experience alongside a discussion of physicality in experimental music, acoustic phenomenology, and computational methodologies. I argue that the study of the neural mechanisms underlying auditory experience is a key component in an interdisciplinary effort towards a phenomenological understanding of music and advocate for a phenomenological and computational approach to the design of psychological experiments. As a demonstration of this approach, I collected several hours of electroencephalography (EEG) data from only one subject: a highly trained musician and composer. The stimuli consisted of natural musical audio from a wide variety of styles, most of which was selected by the subject to be familiar and enjoyable. I investigated the extent to which various acoustic and musical features could be predicted from the continuous, single-trial EEG signal. To construct EEG features, independent component analysis was performed on the 64 channels from each session. The spatial maps of sufficiently stable and dipolar components were clustered and averaged. Finally, a short-time Fourier transform was applied to all accepted clusters and the spectrograms were concatenated into feature vectors. A linear multivariate regression model was trained to predict audio features from the EEG features. The model was evaluated in a leave-one-track-out cross validation and compared to a random baseline model which made predictions by sampling random frames from the training set. The root mean squared error of the regression model was significantly less than that of the random model for all audio features. Since there were no repetitions of the stimuli, these results cannot be explained by overfitting and must represent a general mapping between musical audio and the subject’s electrical brain activity. Future work will experiment with various non-linear regression techniques and a behavioral experiment will be conducted to investigate the musical experiences evoked by the reconstructed audio.
Added by: Mark Grimshaw-Aagaard Last edited by: Mark Grimshaw-Aagaard |