Analysis of analysis: Using machine learning to evaluate the importance of music parameters for Schenkerian analysis

Abstract

While criteria for Schenkerian analysis have been much discussed, such discussions have generally not been informed by data. Kirlin (2014b) has begun to fill this vacuum with a corpus of textbook Schenkerian analyses encoded using data structures suggested by Yust (2006), and a machine learning algorithm based on this dataset that can produce analyses with a reasonable degree of accuracy. In this work, we examine what musical features (scale degree, harmony, metrical weight) are most significant in the performance of Kirlin's algorithm.

Voice-Leading Transformation and Generative Theories of Tonal Structure

“…This close correspondence to Schenkerian notation has made dynamic hierarchy useful for discovering statistical regularities in published Schenkerian analyses (see Kirlin and Jensen 2011, Kirlin 2014, 2016. For instance, Example 4d is ungrammatical because of the crossing slurs, which violate the principle of hierarchy of motions: there is a motion G-A, contained in both A-A and G-B , but there is no containment relation between these two larger spans.…”

Machine Learning and Music Generation

“…and Kirlin and Yust (2016) pay particular attention to the relative contribution of different features employed in representing music data. A challenging issue in computational music generation, especially if fully automated, is the determination of concrete musical events which however cohere with certain more abstract or underlying structures, such as metrical regularities (Ponce de León et al 2016) or both adjacent and distant repetition (Conklin 2016).…”

“…Machine learning is employed in generating events and event sequences, for example through model-based prediction (Giraldo and Ramirez 2016; Kosta et al 2016) or sampling from a statistical model (Conklin 2016), and in evaluating candidate event sequences (Ponce de León et al 2016) or ranking generated sequences (Conklin 2016) based on a statistical model. The contribution by Kirlin and Yust (2016) presents a step towards learning middleground and background regularities, which will further enable future inductive methods to achieve both stylistic and structural coherence of generated music pieces based on models learned from data.…”