Institute of Formal and Applied Linguistics
Charles University, Czech Republic
Faculty of Mathematics and Physics
Improvising accompaniment upon incomplete notation (chord symbols, melody or bass line), is a core skill for the majority of keyboard players across genres: jazz, popular music, or basso continuo in baroque music. These elements of intangible cultural heritage are under-represented in research and preservation, mostly because most of their value is not captured by tangible text. However, the digital domain offers probabilistic models appropriate for describing and analyzing improvisatory style. We will develop a machine learning model of improvised accompaniment, focusing on basso continuo performance.
We already collected a pilot dataset of 175 basso continuo MIDI recordings and created a system capable of aligning these recordings to the basso continuo musical score.
Building on this work, we will collect a dataset of 1000 MIDI recordings of basso continuo performances and create a machine learning model of basso continuo performance harmony and texture, suitable for analyzing collected recordings in the 1st and 2nd year. We will also be continuously looking at the development of (large) foundational models of symbolic music to see how they can be leveraged for the task.
In the 3rd year, we will work on style classification of recordings from the dataset (those not used for training) and test how the model can answer relevant style-related musicological questions. We want to develop methods in a way that they can be generalized to other styles of music.