Not All Roads Lead to Rome: Pitch Representation and Model Architecture for Automatic Harmonic Analysis

Abstract

Automatic harmonic analysis has been an enduring focus of the MIR community, and has enjoyed a particularly vigorous revival of interest in the machine-learning age. We focus here on the specific case of Roman numeral analysis which, by virtue of requiring key/functional information in addition to chords, may be viewed as an acutely challenging use case. We report on three main developments. First, we provide a new meta-corpus bringing together all existing Roman numeral analysis datasets; this offers greater scale and diversity, not only of the music represented, but also of human analytical viewpoints. Second, we examine best practices in the encoding of pitch, time, and harmony for machine learning tasks. The main contribution here is the introduction of full pitch spelling to such a system, an absolute must for the comprehensive study of musical harmony. Third, we devised and tested several neural network architectures and compared their relative accuracy. In the best-performing of these models, convolutional layers gather the local information needed to analyse the chord at a given moment while a recurrent part learns longer-range harmonic progressions. Altogether, our best representation and architecture produce a small but significant improvement on overall accuracy while simultaneously integrating full pitch spelling. This enables the system to retain important information from the musical sources and provide more meaningful predictions for any new input.

On Local Keys, Modulations, and Tonicizations

“…Machine learning approaches, especially using neural networks, have recently gained popularity in MIR research, including key estimation. Independently, Chen et al [8,9] and Micchi et al [31] designed models that estimate local keys as well as roman numeral analysis annotations. Tonicization information is implied by the roman numeral analysis annotations.…”

“…Catteau [6] 10 Manually-built chord sequences Chai [7] 10 Various (Classical) Chen [8,9], Micchi [31] 23 Beethoven Chew [11] 2 Bach Feisthauer [17] 38 Mozart (Classical) Izmirli [21] 17 Pop songs 152 Naxos set (Classical) 17 Kostka-Payne (Classical) Mearns [30] 12 Bach Chorales Micchi [31] 27 TAVERN (Classical) 70 ABC (Beethoven) 72…”

“…. This model was introduced by Micchi et al [31] and utilizes an LSTM network to provide the harmonic analysis of a musical excerpt, which is described through six features: key, degree 1, degree 2, quality, inversion, and root. We evaluate the key predictions of the model when it is used without any postprocessing.…”

“…We would like to thank Micchi et al for facilitating us with the predictions of their algorithm [31] using our dataset. Similarly, we would like to thank the anonymous reviewers of a previous version of this manuscript, who provided precious feedback for this version.…”

Automatic Chord Labelling: A Figured Bass Approach

“…In this paper, we propose an innovative, rulebased model for automatic chord labelling that considers both the musical surface and figured bass. Compared to existing methods that only considered the musical surface [5,6,9,13,15,18,19], the advantages of our approach are: Figure 1: The first measures of BWV 33.06 "Allein zu dir, Herr Jesu Christ" from our Bach Chorale Figured Bass (BCFB) dataset. FBAs and chord labels are shown below the bass line.…”