Emotion classification and analysis of expressive performances in classical piano music

Abstract

The method of expressing emotions according to the performance expression is the most essential element of classical music, and several approaches have been conducted to analyze it. However, most of them are limited in that they attempted to analyze from the perspective of the listener, not the performer, and in a short melody composed arbitrarily. This study proposes an analysis that deviates from these limitations by analyzing the performance characteristics of emotional expression in real classical music from the perspective of the performer using machine learning and neural network. For use in research, classical piano performance data were collected and features that could explain the performance were defined from collected data. This research selects significant features by machine learning methodologies, and then proves that certain performance features effectively represent emotions. Also, the impact of each feature on emotional expression is analyzed to distinguish features that play a significant role in expression of a particular emotion. Furthermore, this research confirms that the emotions of the entire performance were classifiable by using machine learning methodologies and artificial neural network, and also tries multiple instance learning by piece sections. These processes show that the period-specific and compositional features of the piece by the composer can significantly affect emotion representation, and also confirms that noise from the compositional features of the song can affect emotion classification. These results show how each performance feature affects the actual performance and which of them is more important, suggesting that this research quantified the expressive features performers perceive and use empirically.