Evolving Performance Models by Performance Similarity: Beyond Note-to-note Transformations

This paper focuses on expressive music performance modeling. We induce a population of score-driven performance models using a database of annotated performances extracted from saxophone acoustic recordings of jazz standards. In addition to note-to-note timing transformations that are invariably introduced in human renditions, more extensive alterations that lead to insertions and deletions of notes are usual in jazz performance. In spite of this, inductive approaches usually treat these latter alterations as artifacts. As a first step, we integrate part of the alterations occurring in jazz performances in an evolutionary regression tree model based on strongly typed genetic programming (STGP). This is made possible (i) by creating a new regression data type that includes a range of melodic alterations and (ii) by using a similarity measurement based on an edit-distance fit to human performance similarity judgments. Finally, we present the results of both learning and generalization experiments using a set of standards from the Real Book.