Strategies towards the Automatic Annotation of Classical Piano Music

Analysis and description of musical expression is a substantial field within musicology. However, manual annotation of large corpora of music, a prerequisite for describing and comparing different artists’ styles, is very laborintensive. Therefore, computer systems are needed that can annotate recordings of different performances automatically, requiring only minimal corrections by the user. In this paper, we apply Dynamic Time Warping for audioto-score alignment to extract the onset times of all individual notes within an audio recording, and compare two strategies for improving accuracy. The first one is based on increasing the temporal resolution of the features used. To cope with constraints in terms of computational costs, we apply a divide and conquer pattern. The second strategy is introducing a post-processing step in which the onset time of each individual note is revised. The advantage of this method is that, in contrast to default algorithms, arpeggios and asynchronies can be resolved as well.