AI-driven classification and trend analysis of piano music genres using large language models

In this work, we propose a hybrid framework to classify piano genres and predict future genres based on symbolic features and large language models (LLMs). Genre overlap, multimodal data, and sparse metadata are cumbersome to traditional methods. SymD has been used to process symbolic data from MIDI files and textual metadata via GPT-4 Turbo. We trained predictions of 20,000 compositions on 20,000 LLM embeddings, which fusion features including note density, tempo variability, and harmonic structure to 94.0% accuracy and 0.93 F1. Matching historical data gave good alignment for temporal trend analysis. We also improve on existing methods and participate in overcoming metadata limitations. This study presents a new multimodal paradigm to analyse music, which can be applied in musicology, digital archiving, and recommendation systems. Real-time audio-based deployment and integration will occur in future work.