Exploring the Possibilities of Automated Annotation of Classical Music with Abrupt Tempo Changes

Abstract

In this paper, we introduce options for automatic measure detection based on synchronization, beat detection, and downbeat detection strategy. We evaluate proposed methods on two motifs from the dataset of Leos Janacek's string quartet music. We use specific user-driven metrics to capture annotation efficiency simulating a scenario where a musicologist has to use the output of an automated system to create ground-truth annotations on given recordings. In the case of the first motif, synchronization outperformed other methods by detecting most of the measure positions correctly. This procedure was also the most suitable for the second motif—it achieved a low number of correct detections, but the vast majority of transferred time positions belonged within the outer tolerance window. Therefore, in most cases, only shifting operations were needed resulting in higher annotation efficiency. Results suggest that the state-of-the-art downbeat tracking is not yet efficient for expressive music.