HARPSET: A Dataset of Harp Performances for Automatic Musical Performance Analysis

HARPSET is an open-access dataset designed for the automatic analysis of musical performance in the context of the harp. The dataset provides a one-to-one correspondence between symbolic music scores in MusicXML format and audio recordings of harp performances, enabling detailed studies of score–performance relationships. HARPSET includes both reference performances and performances with deliberately introduced, controlled deviations, allowing the analysis of interpretative variability. For each musical piece, up to four performance variants are available: original performances without errors (O), performances with note errors (N), performances with rhythmic deviations (R), and performances combining both note and rhythmic deviations (RN). The dataset contains high-quality stereo WAV audio recordings, symbolic MusicXML scores, structured metadata indexing score–audio pairs, and annotation files describing pitch-related deviations. In addition, auxiliary scripts are provided to validate the dataset structure, generate metadata, and reproduce derived annotation files, ensuring full reproducibility. HARPSET has been developed within the framework of an academic research project and is intended for applications in Music Information Retrieval (MIR), automatic performance assessment, score–performance alignment, error detection, and music education research.