Towards Automatic Measure-Wise Feature Extraction Pipeline for Music Performance Analysis

The task of obtaining ground-truth annotations is of fundamental importance for music performance analysis. Measure positions could be used to navigate throughout the piece, indicate the tempo changes, or help with structure segmentation. In this paper, we introduce an automatic measure-wise feature extraction pipeline. We first annotate one interpretation of the string quartet music and use an audio synchronization strategy to transfer measure positions to all other recordings. We extract features related to tempo, dynamics, and timbre. We compute average values in each measure and propose measure-wise feature matrices. This procedure could be used for any number of recordings as long as at least one reference annotation is available. Finally, we create a binary label for each interpretation based on the Czech origin of performers as an experiment and evaluate the measure-wise tempo distribution.