This thesis investigates computational approaches to understanding and evaluating piano performance, focusing on three interrelated tasks: note-level MIDI velocity estimation, symbolic Dynamics Marking (DM) prediction, and automatic feedback tagging and scoring of student performances. While motivated by pedagogical applications, the primary aim is to advance computational modeling of piano performance to support analysis and intuitive visualization.

The first part reviews prior work in Music Information Retrieval on piano performance analysis, emphasizing expressive dimensions such as dynamics and timing, and identifying limitations in existing models. The second part focuses on note-level MIDI velocity estimation using both audio recordings and note-level frame information. Several deep learning approaches are proposed, including FiLM-conditioned CNN–GRU models, diffusion-based networks, and attention-based U-Net architectures, which demonstrate improved accuracy, robustness, and generalization compared to previous works. The third part introduces a novel framework for symbolic DM prediction by modeling dynamics markings as a contextual sequence learning task. A score-aware tokenization scheme and a RoBERTa-based masked language model are used to infer expressive markings from aggregated velocity profiles. Finally, a prototype system for automatic feedback tagging and scoring is presented, detecting common performance issues and assigning interpretable performance-level scores. 

Together, these components form a modular pipeline for piano performance understanding, contributing new methods to Music Information Retrieval with applications in music pedagogy, performance visualization, and automated feedback generation.

Video: https://www.youtube.com/watch?v=pw86x3U9QrI