Can Audio Reveal Music Performance Difficulty? Insights from the Piano Syllabus Dataset

• DOI: 10.48550/arxiv.2403.03947
• Publication Date: 2024-03-06
• Authors: Pedro Ramoneda; Min-Hee Lee; Dasaem Jeong; Jose J. Valero-Mas; Xavier Serra
• Keywords: Syllabus

Automatically estimating the performance difficulty of a music piece represents key process in education to create tailored curricula according individual needs students. Given its relevance, Music Information Retrieval (MIR) field depicts some proof-of-concept works addressing this task that mainly focuses on high-level abstractions such as machine-readable scores or sheet images. In regard, potential directly analyzing audio recordings has been generally neglected, which prevents students from exploring diverse pieces may not have formal symbolic-level transcription. This work pioneers automatic estimation with two precise contributions: (i) first audio-based dataset -- namely, Piano Syllabus (PSyllabus) featuring 7,901 piano across 11 levels 1,233 composers; and (ii) recognition framework capable managing different input representations both unimodal multimodal manners derived perform task. The comprehensive experimentation comprising pre-training schemes, modalities, multi-task scenarios prove validity proposal establishes PSyllabus reference for MIR field. well developed code trained models are publicly shared promote further research