15th European Conference on Artificial Intelligence
|July 21-26 2002 Lyon France|
Efstathios Stamatatos, Gerhard Widmer
Expressive music performance is the interpretation of a piece of music according to the artist's understanding of the structure (or 'meaning') of the piece. Every skilled performer intentionally modifies important parameters, such as tempo and loudness, in order to stress particular notes or passages. In this paper the problem of identifying the most likely music performer, given a set of piano performances of the same piece by a number of skilled candidate pianists, is addressed. We propose a set of features for representing the stylistic characteristics of a pianist. In addition to features that express deviations from the printed score in terms of timing, articulation, and dynamics, we introduce features that are based on the deviation of the pianist from the performance norm (i.e., average performance of a piece, derived from a different set of performers). We show that the norm deviation measures are more stable and reliable in comparison to score deviation measures. Moreover, features that exploit the melody lead phenomenon (i.e., notes in a chord that are not played simultaneously) are introduced. A database of piano performances of 22 pianists playing two pieces by F. Chopin in a special piano is used in the presented experiments. Due to the limitations of the training set size (i.e., only a few training examples per class are available) and the characteristics of the input features (i.e., the score deviation features are affected by slightly changed training sets) we propose a classification model that takes advantage of various techniques of constructing meta-classifiers: subsampling the training set, subsampling the input features, and a weighted majority scheme. We show that the proposed ensemble performs better than any of the constituent simple classifiers when training and test set are taken from different musical pieces and is able to cope efficiently with a very difficult task, even for human experts.
Keywords: Art and Music, Machine Learning
Citation: Efstathios Stamatatos, Gerhard Widmer: Music Performer Recognition Using an Ensemble of Simple Classifiers. In F. van Harmelen (ed.): ECAI2002, Proceedings of the 15th European Conference on Artificial Intelligence, IOS Press, Amsterdam, 2002, pp.335-339.