Toward Actionable Computational Models of Piano Fingering

Determining an optimal sequence of fingerings to use to perform a piece of music is a fundamental skill for accomplished pianists. This work contributes to efforts to model the piano fingering problem computationally and to produce computational systems to help advancing pianists arrive at actionable fingering solutions more quickly than is possible with printed scores. It begins by clarifying the range of sub-problems that exist in the domain and attempts to address the dearth of data available for training and evaluation purposes. A novel software module is described that can be embedded in web pages to make it easy for pianists to finger electronic scores manually. It proceeds to describe a novel hardware/software solution that can capture fingering choices automatically from actual performances. With this tooling, several new data sets for use in the domain are collected and leveraged to train machine learning models. Specifically, conditional random fields are applied to the piano fingering problem for the first time. The merits of these systems are compared to recently published hidden Markov and deep learning models. Given the acceptance of “multiple ground truths” in the domain, a computer system promising truly actionable piano fingering advice must ultimately function like a recommender system. An evaluation method for piano fingering systems based on expected reciprocal rank, a method suggested for web search, is therefore presented here. Finally, next steps are charted for the domain, urging smaller steps in which individual pianists are modeled using data collected automatically and evaluated with simpler metrics. When a diverse set of pianists have been well modeled, the components will be in hand to develop useful recommender systems in the piano fingering domain.