Interaction features for prediction of perceptual segmentation : Effects of musicianship and experimental task
Hartmann, M., Lartillot, O., & Toiviainen, P. (2017). Interaction features for prediction of perceptual segmentation : Effects of musicianship and experimental task. Journal of New Music Research, 46 (2), 156-174. doi:10.1080/09298215.2016.1230137
Julkaistu sarjassaJournal of New Music Research
© 2016 Informa UK Limited, trading as Taylor & Francis Group. This is a final draft version of an article whose final and definitive form has been published by Taylor & Francis. Published in this repository with the kind permission of the publisher.
As music unfolds in time, structure is recognised and understood by listeners, regardless of their level of musical expertise. A number of studies have found spectral and tonal changes to quite successfully model boundaries between structural sections. However, the effects of musical expertise and experimental task on computational modelling of structure are not yet well understood. These issues need to be addressed to better understand how listeners perceive the structure of music and to improve automatic segmentation algorithms. In this study, computational prediction of segmentation by listeners was investigated for six musical stimuli via a real-time task and an annotation (non real-time) task. The proposed approach involved computation of novelty curve interaction features and a prediction model of perceptual segmentation boundary density. We found that, compared to non-musicians’, musicians’ segmentation yielded lower prediction rates, and involved more features for prediction, particularly more interaction features; also non-musicians required a larger time shift for optimal segmentation modelling. Prediction of the annotation task exhibited higher rates, and involved more musical features than for the real-time task; in addition, the real-time task required time shifting of the segmentation data for its optimal modelling. We also found that annotation task models that were weighted according to boundary strength ratings exhibited improvements in segmentation prediction rates and involved more interaction features. In sum, musical training and experimental task seem to have an impact on prediction rates and on musical features involved in novelty-based segmentation models. Musical training is associated with higher presence of schematic knowledge, attention to more dimensions of musical change and more levels of the structural hierarchy, and higher speed of musical structure processing. Real-time segmentation is linked with higher response delays, less levels of structural hierarchy attended and higher data noisiness than annotation segmentation. In addition, boundary strength weighting of density was associated with more emphasis given to stark musical changes and to clearer representation of a hierarchy involving high-dimensional musical changes. ...