Chord Context and Harmonic Function in Tonal Music

christopher wm. white and ian quinn

This article investigates several questions of harmonic function using aggressively data-driven
approaches. We apply Hidden Markov Modeling—a technique used to identify contextual regulari-
ties within streams of data—to the Kostka-Payne, McGill Billboard, and Bach chorale corpora.

Downloaded from [Link] by SMT Member Access user on 12 August 2019

The resulting models question the generalizability of the traditional three-function model, illustrat-
ing the syntactic uniqueness of various corpora while also highlighting recurrent characteristics of
tonal repertories. Finally, this article offers some general observations, including questioning the
role that tonal hierarchy plays in theories of function and discussing the cultural politics inherent in
assuming the universality of one functional system.

Keywords: corpus analysis, function, tonality, Riemann, computational modeling, popular music,
harmony, Kostka and Payne, McGill Billboard, Bach chorale, machine learning, music cognition,
Hidden Markov models.

S
ince its articulation by Riemann (1893) in his chords they tend to precede and succeed. In doing so, we will be
Vereinfachte Harmonielehre of 1893, the concept of har- able to discuss certain statistical properties of the corpora
monic function has provided music theorists with re- themselves, thereafter relating these properties to larger dis-
markable explanatory power.1 Function theory explains the cursive and theoretical issues. (We further discuss some
basic intuition that certain sequences of chords are syntactical underpinnings of our corpus methods below, along with our
and others unexpected. What distinguishes function theory modeling techniques.)
from other syntactic theories of harmony (e.g., Stufentheorie) is We will investigate three corpora using these methods: the
that its small number of categories yield a very efficient system Kostka-Payne corpus, the McGill Billboard corpus, and the
of rules for harmonic progression. widely studied corpus of Bach chorales.2 After describing our
But while harmonic function remains a powerful tool findings, we will return to an examination of broader topics,
within contemporary music theory, with few exceptions, the including some generalizations that seem to hold across our
concept has remained relatively unchanged since Riemann. In datasets that problematize music theory’s standard
this article, we challenge commonly held assumptions regard- three-function model. We will demonstrate that certain char-
ing harmonic functions within music-theoretical discourse: acteristics generalize across these three repertoires, but these
first, while tonic and dominant are generally accepted as the similarities are outweighed by both stylistic differences be-
basic duality of any system of functions, the number, identity, tween the repertoires and the differences in representation
and behavior of additional functions is not necessarily clear; (e.g., Roman numerals versus lead-sheet symbols) that our
second, we question whether a chord’s functional identity is models use. Our findings also connect to larger theoretical
determined by its scale-degree content, its root, its local con- topics surrounding harmonic function: we will show evidence
text, or by some interaction between these factors. that surface events, such as passing tones and neighbor chords,
Our approach is fundamentally data-driven, creating exhibit syntactic regularities similar to consonant harmonies,
models based on large musical corpora. After reviewing the and argue that the perceived hierarchical superservience of cer-
basic definitional issues concerning harmonic function, we tain functions may be a simple matter of chord frequency.
will describe the technology used to model our data. This (Importantly, as we make clear in the ensuing prose, our goal
model will approach the idea of harmonic function from an is not to prove or solve any of these points but to provide novel
aggressively naı̈ve and skeptical standpoint: we will imple- observations in order to complicate and add subtleties to our
ment a model that groups chords into categories only by the understanding of harmonic function.)

CONTEXT, CONTENT, AND HARMONIC FUNCTION

1 The concept of grouping chords into equivalencies or hierarchies based on
musical parameters precedes the formal concept of “function.” Rameau
(1737) groups chords into tonic, subdominant, and dominant categories
We can begin to survey these issues by reviewing a frequent
based on root motions; Riepel (1755) and Momigny (1806) organize dia- music theoretical argument concerning the terms
tonic triads into a tonic-centered hierarchy; F!etis (1844) theorizes chords “subdominant” and “predominant.” The former term is often
containing certain scale degrees as being the impetus behind harmonic
tendencies; while Hauptmann (1853) generates tonal spaces using princi- 2 The Kostka-Payne corpus was used in Temperley (2009); the McGill
ples of dualism. Billboard corpus was used in Burgoyne (2012).

314
 chord context and function 315

associated with the IV chord, though when Rameau first used virtue of their formal position and their relationship to other
the term in his Nouveau Syst^eme, it referred to what we would chords rather than through any internal characteristics of the
now call ii65 .3 The latter term derives from the concept of domi- chords themselves. Thus, a given V chord might function as
nant preparation articulated by Allen Forte, under the influence the dominant in a phrase, not because it contains the leading
of Schenker.4 Some theorists conflate these usages: Deborah tone but because it resolves to the tonic and forms an authentic
Stein, for example, writes that “the subdominant functioned cadence.”9 Under this definition, functions gain their identity
either as preparation for the dominant or as a neighboring har- by their syntactic position, or the context in which the constit-
mony that prolonged the tonic chord.”5 Other theorists seek to uent chords occur.
tease these actions apart by arguing for two distinct functions. Much of contemporary music theoretical discourse, how-

Downloaded from [Link] by SMT Member Access user on 12 August 2019

Kevin Swinden makes the provocative claim that “harmonic ever, focuses not on context but on content: a chord’s function
function cannot be defined by pitch class alone,” arguing that a is determined by the scale-degree identity of its constituent
IV chord should only be called “subdominant” if it progresses scale degrees. The term function, of course, originates in
to I; otherwise, it is a “dominant preparation.”6 Charles Smith Hugo Riemann’s Vereinfachte Harmonielehre, which espouses
has gone so far as to argue that the plagal function (his term an undeniably content-based approach in which a chord’s
for the subdominant as distinct from predominant) should be function is determined by its relation to one of the three pri-
given the same status as the dominant function, with a pre- mary triads: I, IV, and V.10 Riemann’s motivation for a three-
plagal function analogous to the predominant.7 The Kostka- fold functional system stemmed from Hegelian dialectics
Payne textbook argues for a still more complex understanding combined with a dualist interest in the overtone (and under-
of IV’s functional meaning: “The IV chord is an interesting tone) series.11 But, as Eytan Agmon has argued, “the special
chord because it has three common functions. In some cases, status of root relationships by fifth is by no means a necessary
IV proceeds to a I chord. . . . More frequently, IV is linked assumption in the theory of harmonic function.”12 Pulling the
with ii; IV can substitute for ii (going directly to V or viio), or metaphysical rug out from under function theory, some
IV can be followed by ii (as in IV–ii–V).”8 The authors seem thinkers have refined their content-oriented definitions, fo-
to argue for a subdominant, a predominant, and a pre- cusing on chord prototypes rather than chord progressions, a
predominant function. direction Agmon himself pursues. Agmon situates the three
Example 1 schematizes a central aspect of the distinction primary triads in terms of prototype theory, where I, IV, and
between predominant and subdominant. On the left-hand V are central representatives of three categories maximally
side, we show a standard three-function model representing distant from one another in a voice-leading space (see Ex. 2).
the conflation of subdominant and predominant functions. He does not, however, question Riemann’s fundamental
In this model, the category “S/P” represents chords including assumptions about the number and identity of the harmonic
IV and other chords sharing scale degrees ^4 and ^6. The cate- functions.
gory is defined by its content: all IV chords belong in this Contemporary theorists seeking to extend the concept of
category. On the right-hand side, we disambiguate the third harmonic function beyond the Riemannian limits invoke the
function, showing the relationship between T and S as bidi- expectations, qualities, or hierarchical positions associated with
rectional, whereas a motion from T to P continues to D. specific scale degrees. Daniel Harrison retains a threefold func-
Here the categories are defined by their context: whether a tional system but considers individual scale degrees as the
particular IV chord belongs in the category S or P depends source of functional identity, with scale-degree assemblies
on what follows it. We will make a distinction between these rather than rooted tertian chords as the bearers of harmonic
two ways of defining harmonic functions: context-driven function.13 Fred Lerdahl similarly derives function from the
approaches are concerned with chords’ usage and, and con- placement of pitches and chords within a tonal hierarchy, with
tent-driven approaches are concerned with chords’ scale- tonic as the most superordinate, dominant the next, and so
degree constituents. on.14 To the three Riemannian functions, Lerdahl adds an
Context-driven theories define function in terms of syntax. additional four (departure, return, neighbor, and passing)
In the right-hand side of Example 1, the difference between S based on the distinct ways that subordinate chords can prolong
and P is the context in which the functions respectively occur.
According to Drew Nobile, “Chords gain their function by 9 Nobile (2014, 22).
10 Harrison (1994, 39–40); Kopp (2002, 5).
3 Lester (1992, 132). 11 “These ist die erste Tonika, Antithese die Unterdominante mit dem
4 Forte (1962). The Schenkerian concept of intermediate harmony includes Quartsextakkord der Tonika, Synthese die Oberdominante mit dem
a wider range of chords than we traditionally assign to the subdominant schliessenden Grundakkord der Tonika; thetisch ist die Tonika, anti-
and predominant categories. thetisch die Unter-, synthetisch die Oberdominante” (Riemann [1900–
5 Stein (1983, 156). 1901, 3]). For more on Riemann’s dialectics, see Harrison (1994),
6 Swinden (2005, 253). Klumpenhouwer (2007), and Dahlhaus (1968 [1990]).
7 Smith (1981). Tymoczko (2003) notes that the concept of plagal function 12 Agmon (1995, 199); see also Agmon (2013).
might find its roots in the Boulanger tradition. 13 Harrison (1994).
8 Kostka and Payne (2012, 114), emphasis added. 14 Lerdahl (2001).
316 MUSIC THEORY SPECTRUM 40 (2018)

Downloaded from [Link] by SMT Member Access user on 12 August 2019

example 1. Two probabilistic models of harmonic function

the behavior of groups based exclusively on that parameter;

second, it allows us to test our intuitions about chord content;
third, our method allows us to ask, “what about this corpus
invites us to hear syntactic categories?”—allowing us to begin
to connect our computational findings to our intuitions sur-
rounding harmonic function. To these ends, we propose a
model that takes into account context only, ignoring the scale-
degree content of chords. The approach is based on Hidden
Markov models (HMMs), a technique from machine learning
described in the next section. Our project is aggressively data-
driven: our results will model the contextual properties of large
amounts of musical data in a way that attempts to minimize
the effects of preexisting theoretical assumptions on the analy-
sis. Our goal in taking a radically data-driven approach is not
example 2. Agmon’s Circle of Functional Prototypes
to create a new theory from scratch or to discard existing theo-
ries, but to isolate one parameter of harmonic function (chord
relatively superordinate chords. Lerdahl’s Dep function, for ex- context) in its most simple form (two-chord progressions) in
ample, characterizes chords that, while lacking a leading tone, order to both learn about the syntactic properties of the
otherwise resemble back-relating dominants in having a tem- corpora in question and to interrogate our own ideas about
poral and prolongational dependency on a preceding tonic. In harmonic function.
one sense, this function abstracts away from the specific scale-
degree constituents of dominant harmony, but, as with each of SOME NOTES ON CORPUS METHODOLOGY
Lerdahl’s prolongational functions, its identity is determined
wholly by a prolongational structure that is itself dependent on Given that corpus methods are relatively novel within main-
a tonal hierarchy. What seems like a context-oriented function stream music theory, a few words on our motivations and
is, in fact, tethered to an elaborate and precisely determined methodologies are in order. Digital humanities scholars frame
notion of a hierarchical pitch space. What remains unex- their computational analyses in a number of ways, but, for our
plained, however, is how scale degrees are endowed with this purposes, we might divide approaches into two rough catego-
hierarchical power. ries, the backhoe and the microscope.15 “Backhoe” methods use
While most investigations of function start by intuiting that computers to do the sorts of things that humans do, just faster.
certain groups of chords progress to other groups of chords From this standpoint, the analyses that follow represent
and build functional categories from that standpoint, this study insights into chord function that one might come to given the
begins from the intuition that music simply has contextual reg- painstaking chord-to-chord annotations and repeatedly calcu-
ularity and then investigates one way in which chords might lating groupings based on those similarities. “Microscope”
group into those contextual categories. The power of this ap- methods suggest that their computational analyses illuminate
proach is threefold: first, it allows us to isolate one particular
corner of harmonic function—two-chord context—and isolate 15 These categories are roughly drawn from Wittig (1977).
 chord context and function 317

things that we could not see with purely human capacities. former and domesticates the latter.18 Other computational
From this standpoint, our analyses offer observations into the models of harmonic function will produce different results,
way that chord-to-chord progressions might group together, given their contrasting inner workings. In what follows, we
because they do so with a precision, insight, or impartiality im- will endeavor to make clear which features of our results are
possible to human perceptions. more interesting for music theorists and which are side effects
The ability to strictly isolate the contextual parameter in or- of our reliance on HMMs.
der to create broader contextual categories is arguably unique
to the domain of computation, and therefore more of a COMPOSITE HIDDEN MARKOV MODELS
“microscope” approach. But, regardless of how exactly our ap-

Downloaded from [Link] by SMT Member Access user on 12 August 2019

proach situates within this continuum, it is important to keep Example 1 uses three kinds of symbols: Roman numerals sig-
in mind what the method does not do. Regardless of the nifying chords, letters signifying functions, and arrows signify-
formalisms, our goal is not to “discover” any sort of “truth” ing motions between functions. The logic of these models
underlying harmonic function—our models are neither involves an underlying progression of functions and a system
meant to be strictly predictive of human perception or anal- for realizing those functions as chords. A given functional pro-
ysis, nor to say what theorists or musicians should perceive gression can be realized in multiple ways: T–P–D–T, for ex-
or analyze, nor even to capture the compositional models ample, might be realized as I–IV–V–vi or as iii–vi–V–I. In the
used to produce these repertoires. Rather, we will investi- language of Hidden Markov Models, chords are observations
gate the notion of function by isolating one aspect of its generated by underlying hidden states, or functions.19 The basic
definition—context, which in its simplest form entails two- assumption of the HMM formalism is that a given observable
chord progressions—by producing a formal model using sequence is determined by two sets of probabilities: lexical
only those parameters.16 probabilities determine what observable form will be taken by
But, while our methods are not meant to strictly model ei- any given hidden state, and transition probabilities determine
ther the methods of composers or the behavior of listeners, the progression of hidden states. Example 3 illustrates the gen-
they do allow for speculative connections into these domains. erative idea of HMMs with its characteristic interaction of
After all, corpora were produced by composers and designed these two types of probability. The likelihood that one ob-
to be heard and understood by some group of listeners. As de- served symbol will follow another is determined only indirectly
scribed in White (2013), while not producing definitive evi- with hidden states governing both order (transition probabili-
dence, corpus studies allow us to discuss the properties that are ties) and observed form (lexical probabilities). A HMM is
observable in some composer’s style, as well as the kinds of sta- therefore comprised of four items 1) a set of k possible hidden
tistical regularities that allow listeners to both learn about and states, 2) a vocabulary of possible observed symbols, 3) a prob-
interpret music.17 ability distribution that determines how often each hidden
However, unlike other music-theoretic methodologies, the state proceeds to each other hidden state (transition probabili-
adoption of a computational approach compels researchers to ties), and 4) a probability distribution governing how likely
define precisely how their data will be collected, modeled, and each vocabulary item is to be produced by these hidden states
interpreted. A microscope will always produce two-dimen- (lexical probabilities).
sional images of the cells and microorganisms it observes, but The transition probabilities represented in Example 1 (the
that does not mean we believe that these images prove that arrows) predict that T–P–D–T is a much more likely func-
cells and microorganisms are two-dimensional creatures. In tional progression than D–P–T, and the lexical probabilities
the project documented in this article, our adoption of HMMs (the pie charts) make it unlikely that T–P–D–T would be real-
produces results that tell us both about contextual regularities ized as IV–iii–I–V.20 Again, the model does not directly pre-
in musical corpora and about the predilections of HMMs. It is dict what chords should follow a given chord. There is no rule
left to us to interpret the results in a way that foregrounds the that directly prohibits ii from moving to I, but the combina-
tion of transition and lexical probabilities make this succession
16 This somewhat differs from how statistical modeling is used in scientific highly unlikely. The HMM formalism does not depend on
inquiries of music (e.g., Krumhansl 1990, Temperley 2009, Burgoyne,
Wild, and Fujinaga 2013) but adheres to norms of digital humanities. 18 Mavromatis (2009, 2012) provide an excellent case study of this kind of
While in the former, computational models can be judged by their adher- interpretation.
ence to a ground truth (“correct answers”) of human annotations or 19 HMMs, while used frequently in linguistic studies, are relatively new to
responses, the latter observes some trend within a dataset in order to music theory research. Panayotis Mavromatis (2005) has experimented
deepen our understanding of the dataset itself. with HMMs to investigate meter in Palestrina and the syntax of Greek
17 White (2013). The latter claim is based in theories of statistical musical Orthodox chant, while several researchers (e.g., Bello and Pickens [2005])
learning in which humans learn musical norms by being exposed to the have used HMMs to induce harmonic content from sound waves.
statistical properties of some dataset (e.g., Saffran, Johnson, Aslin, and Raphael and Stoddard (2004) have used HMMs to fit pitch-class infor-
Newport 1999; Creel, Newport, and Aslin 2004; Saffran, Reeck, Niebuhr, mation into mode, key, and chord templates.
and Wilson 2005; Huron 2006; Loui, Wessel, and Hudson Kam 2010; 20 The absence of arrows between two states most often indicates that the
Loui 2012). transition is highly improbable rather than impossible.
318 MUSIC THEORY SPECTRUM 40 (2018)

Downloaded from [Link] by SMT Member Access user on 12 August 2019

example 3. The generative model of HMMs

prototype theory, though lexical probabilities can strongly favor A more detailed discussion of HMMs and examples of expec-
a particular realization of a given hidden state. For instance, in tation maximization are presented in Appendix A.21
the case of T, the most likely observed chord is I. (At this While the Baum-Welch algorithm runs largely unsuper-
point, it is important to remember that, even though we might vised, it requires an important decision to be made in advance:
use symbols that denote chord content to a human, these char- the number of hidden states (k). While there is no universally
acteristics are unavailable to the underlying algorithm. To the accepted method to determine the ideal number of states, we
computer, “I” and “vi” are arbitrary symbols, and it has no way introduce a novel approach in this essay. Our method exploits
to know that chords designated by these symbols share two a property of expectation-maximization algorithms, like
common tones. Rather, the algorithm is only aware of how Baum-Welch: in part because of the random initialization of
these random symbols are positioned in relation to one another the parameters, models estimated from the same training data
in the analyzed corpus. It is in this sense that our model is in repeated trials will not replicate each other consistently if
“purely contextual.”) the chosen value of k does not fit the data well. Example 4
If lexical and transition probabilities are known, the model depicts the procedure; we will begin our discussion in medias
can be used to analyze a series of observations, assigning the res, at what is labeled as the Cardinality Loop. For each value
most likely hidden state to each chord. This task is comparable of k under consideration, we use Baum-Welch to train a bank
to what we teach our undergraduate theory students to do of 300 HMMs on the same training data. We then use a stan-
when undertaking functional analysis: to reconcile an observed dard method (the Viterbi algorithm) to have each HMM de-
musical surface with an abstract knowledge of the rules of code a test data that was not included in the training data.
functional progression. For example, knowing whether to call Every observation (chord) in the test data is then associated
a given vi chord tonic or predominant depends both on what with a vector of 300 hidden-state labels, each assigned by one
the next chord is and what the acceptable functional progres- of the HMMs.
sions are. The novelty of our procedure lies in the next step, where
Techniques of machine learning make it possible to esti- we attempt to create a composite of the 300 HMMs. We use
mate transition and lexical probabilities directly from a given a different expectation-maximization algorithm, known as k-
series of observations. An iterative process called the Baum- medoids, to classify the set of observations in the test data into k
Welch algorithm finds an optimal fit between the hidden categories (where k is the number of hidden states in each
states and the observations. Baum-Welch begins with a ran- HMM) based on a simple dissimilarity measure equal to the
dom set of transition and lexical probabilities, and each itera- square of the number of HMMs that decoded the two observa-
tion of the algorithm consists of two steps known as tions in question into the same hidden state. The k-medoids al-
expectation and maximization. In the expectation step, these gorithm, as it is implemented in the R programming language,
(initially random) probabilities are used to decode or analyze the yields a quantity called silhouette width, a measure of the ease of
observation sequence, determining the most likely hidden state the algorithm’s task given the data. A low silhouette width indi-
for each observation given the current model. The model can cates overlapping and interpenetrating categories. The higher
then assign an overall probability to the sequence. Initially this the silhouette width, the clearer the boundaries between catego-
probability will be low, since the model’s parameters (the tran- ries are, and the higher the degree of consistency between the
sition and lexical probabilities) are set randomly. In the maxi- individual HMMs. A high silhouette width allows us to treat
mization step, the algorithm adjusts the parameters in an the categories learned by the k-medoids algorithm as hidden
attempt to improve the probability of the training data given states of a composite of all the HMMs.
the model. The expectation-maximization process is iterated
until the improvements reach a point of diminishing return. 21 See the online version of this article.
 chord context and function 319

Downloaded from [Link] by SMT Member Access user on 12 August 2019

example 4. Procedure for learning and cross-validating composite HMMs with optimal numbers of hidden states

We prefer values of k for which the HMMs replicate each some fraction of the dataset and then assess how well the
other more consistently (indicated a higher silhouette width) model conforms to the remaining fraction. If one iterates this
than for neighboring values of k. A higher silhouette width indi- process with different divisions and the same results hold, one
cates more consistent solutions within the hundreds of models can be relatively certain that the model is capturing some gen-
using that k value: if that value produces a peak compared to the eralizable and robust property (rather an artifact of random
surrounding values, that means k hidden states organize the variation). The following analyses vary the types of validation
space better than both k-1 functions, and k þ 1 functions. In used: since the Kosta-Payne corpus is derived from a textbook
this study, we prefer values of k exhibiting the most dramatic that addresses issues of harmonic function, we compare our
peaks. All silhouette-width figures are shown in Appendix A.22 results to the authors’ own statements on this topic. Given
There are two approaches to validating an HMM. One way that we mean to be skeptical of how notions of harmonic func-
is to use it to decode some test data and compare the results to tion manifest on musical surfaces, the Bach chorale and popu-
a ground truth derived from human analyses of the same data. lar music corpora are both sufficiently large to engage in the
If the model reproduces the insights of the human analyst, the latter cross-validation tactics.
correlation suggests that the model is producing some salient Such a model of harmonic function takes only the context
result. We are, however, less interested in how a modern lis- of chords—and not their content—into consideration. If the
tener might analyze a corpus than we are in identifying consis- model puts two chords into the same functional category, it is
tent and reproducible properties of corpora. To this end, we on the basis of their behavior alone and not any shared
also use cross validation, in which we create a model using only scale-degree content. In what follows, we apply this modeling
procedure to several corpora, beginning with a relatively
22 See online version of this article. straightforward dataset: the Kostka-Payne corpus.
 320 MUSIC THEORY SPECTRUM 40 (2018)

MODELING THE KOSTKA-PAYNE CORPUS WITH FOUR FUNCTIONS distributed. In particular, to make sense of IV–I progressions,
the model places IV chords into the “dominant” category, es-
The Kostka-Payne corpus comprises the analytic annotations sentially conflating S and D functions under Example 1(b)
to the musical examples provided in the instructors’ edition of into a single “pretonic” category. Chords with a root of ^6 have
the eponymous harmony textbook. This corpus records the also been removed from the tonic category in this version:
root of each chord in both absolute (pitch) and relative (scale while this simplifies the lexical probabilities by placing all such
degree) terms, along with metadata about the composer, title, chords into the predominant category, it makes deceptive
and mode (major or minor) of each excerpt. Relative-pitch cadences very unlikely events. Because of these difficulties, this
chord roots are represented as integers modulo 12, correspond- model also assigns relatively low probabilities to the observa-

Downloaded from [Link] by SMT Member Access user on 12 August 2019

ing to chromatic scale degrees: I ¼ 0, [II (or ]I) ¼ 1, II ¼ 2, tion sequences. Other three-state models are similarly compro-
etc. These representations are equivalent to Roman numeral mised, though in different ways, and the composite three-state
analyses that specify neither inversions nor the existence of model is incoherent, suggesting that this value of k does not fit
nonchord tones or chordal sevenths, with the qualification that the contextual regularities of the Kostka-Payne corpus.
no enharmonic distinctions are made between, say, [II and When k ¼ 4, the Baum-Welch algorithm produces a family
vii# /ii, or [VI and vii# /vi. Chord quality is also not included in of models that are more consistent with each other. Example 7
these annotations: V/ii and vi, for instance, would receive the shows the composite solution, again using arrows for the tran-
same annotation since they share the same root. Temperley sition probabilities and pie graphs of Arabic numerals for the
provides two versions of the corpus that differ in their treat- lexical probabilities. Moving clockwise from the upper left,
ment of cadential 64 chords: we use the version in which these these hidden states function as tonic, pre-predominant, pre-
chords are assigned a root of ^5. Example 5 shows the transi- dominant, and dominant/pretonic.24 (In what follows, we will
tions between the chords within the major-mode segment of refer to these as T, P–, P, and D/T–.) The most frequent
the corpus, with the most frequent transitions in bold (transi- transitions occur between T and D/T–, with more transitions
tions are from rows to columns). from D/T– to T than vice versa. T goes to several functions,
The Kostka-Payne corpus provides an ideal starting point but T is only progressed to via the D/T–function. The remain-
for this project, since we can compare the authors’ own discus- ing two functions offer two other routes to prolong the T–D/
sion of each chord’s contextual role with our results. We, T––T progression, either first visiting the P function before
therefore, began our HMM analysis of this corpus with a going to the dominant, or prepending the P– into that pro-
series of three-state models, to see whether the HMM would gression. Notably, this four-function diagram reflects Kostka
sort chords into the conventional tonic, dominant, and sub/ and Payne’s analysis (discussed above) of the “three common
predominant functions. The clusters at k ¼ 3 produced a low functions” of a IV chord: proceeding to I, proceeding to V, or
silhouette width, indicating a failure of the individual HMMs proceeding to ii, which will in turn proceed to V. This diagram
to converge collectively on a single composite model. Example also illustrates the authors’ preference for falling fifths and
6 shows two sample solutions produced by the EM algorithm. thirds within chord syntax.25 The most probable chord pro-
Transition probabilities are represented by the weight of the gression involving all four functions, given these lexical and
arrows, and lexical probabilities are represented by pie charts transition probabilities, would be I–vi–ii–V–I, a prototypical
corresponding to the hidden states. Due to the corpus’s ambi- falling fifth/third progression.
guity regarding chord quality, chord roots are represented as Importantly, these results conform to the authors’ descrip-
scale degrees in Arabic numerals rather than Roman numerals. tion of chord function. We quoted above a crucial passage
The size of each pie chart represents the relative probability of concerning the IV chord, suggesting three functions for that
each hidden state. Specific probability tables can be found in sonority: in the composite four-function model, IV chords do
Appendix B.23 indeed have three possible roles. Furthermore, our HMM’s
Many aspects of the model in Example 6(a) resemble the behavior also reflects many aspects of Kostka-Payne’s chord-
idealized model in Example 1. Moving clockwise from the up- function diagram, reproduced below in Example 8. Just as in
per left, the three states correspond to tonic, predominant, and the four-function HMM, the diagram shows a “dominant”
dominant functions, and the chord roots output by each hid- category with V and viio progressing to I, a “predominant” cat-
den state are those predicted by traditional three-function the- egory with ii and IV progressing to the dominant chords.
ory. This model, however, is not representative of the family of Unlike our model, Kostka and Payne group iii and vi chords
models produced by the Baum-Welch algorithm. Since its into their own categories; but like our model, these chords
transition probabilities are distributed relatively evenly—no progress between one another and progress to the predominant
pathway is unidirectional and no arrow is particularly thick— category. While the iii and vi chords may have their own inde-
the model’s probability judgments will be somewhat low. pendent transitional tendencies (iii, for instance transitions to
Example 6(b), on the other hand, contains several unidirec- vi most frequently, while vi transitions most often to ii), the
tional arrows, but its lexical probabilities are more evenly
24 “Pre-predominant” is adopted from Doll (2007).
23 See online version of this article. 25 Kostka and Payne (2012, 104–5).
 chord context and function 321

Downloaded from [Link] by SMT Member Access user on 12 August 2019

example 5. Transitions between chords in the major-mode segment of the Kostka-Payne corpus

example 6. Two representative three-state models: (a) and (b)

overall interactions of these behaviors create the contextual cat-

egories captured in this model.26
generally proceed to a ii chord in actual composition; however, the HMM
26 This feature of the resulting models makes them often insufficient to gen- would seem to indicate that this chord could proceed to any other chord
erate actual chord progressions. For example, a viio/ii chord should within the P- or P functions.
 322 MUSIC THEORY SPECTRUM 40 (2018)

Downloaded from [Link] by SMT Member Access user on 12 August 2019

example 9. McHose’s Four-Chord Classifications

disambiguate these categories, nor produce consistent results.)

(See the discussion of Examples A4–6 in Appendix A.)
Our analysis of the Kostka-Payne corpus, then, has
highlighted two issues: the surprising failure of the Baum-
Welch algorithm to learn the standard three-function model
from the corpus, and the inability of a purely contextual model
to segregate chords in some ways valued by traditional function
example 7. The composite solution for four states theory, typified by “dominant” IV chords. In what follows, we
expand our investigation into pop and rock music, a corpus in
which the IV chord has an even wider array of uses. This
move allows us not only to access one of the largest publicly
available corpora of Roman numerals, but also to engage di-
rectly with the in-depth, critical, and contentious scholarly
engagements with harmonic function that can be found in the
literature on pop-music analysis.

MODELING THE MCGILL BILLBOARD CORPUS WITH EIGHT

example 8. Kostka and Payne’s chord-progression chart
FUNCTIONS

But this model also reflects the outlook of other work link-
ing harmonic function to corpus analysis. Allen Irvine Scholars disagree about some very basic definitions about har-
McHose, an early advocate of a corpus-based approach to monic function in popular music, ranging from what consti-
tonal harmony, uses a model strikingly similar to our four- tutes dominant function to whether traditional harmonic
state model in his 1947 textbook, The Contrapuntal-Harmonic function has any relevance to pop-music at all. On the one
Technique of the 18th Century. Undertaking an exhaustive hand, scholars like Allan Moore, Ken Stephenson, David
quantification of the root motions in “Bach, Graun, Handel, Temperley, and Trevor deClercq argue for a fundamental dif-
and Other Contemporaries,” McHose groups chords into four ference between popular music and common-practice har-
“classifications,” shown in Example 9.27 The classifications mony.28 Temperley and DeClercq, in a corpus study of rock
schematize a series of falling fifths into a final tonic, creating harmony, emphasize many aspects of this difference: rock har-
five equivalence classes reminiscent of the four-function model mony does not have strong unidirectional tendencies (e.g.,
under Example 7. V progresses to IV as much as IV progresses to V), and, in
The most peculiar property of our model, which sets it many cases, IV (rather than V) functions as the primary non-
apart from McHose and nearly every other textbook, is that of tonic triad. On the other hand, several analysts have attempted
“dominant” IV chords. The model conflates dominant and to theorize pop/rock harmonic syntax as an extension of
subdominant functions, creating a “pretonic” hidden state. common-practice norms. Nicole Biamonte and Chris Doll, for
This conflation is a consequence of the model’s exclusive focus instance, argue for including modal harmonies into functional
on two-chord context: without knowledge of scale-degree con- models, with [VII functioning as dominant (Doll’s “rogue
tent or common tones, the model unifies the dominant and dominant”) or as IV/IV (Biamonte’s “Double Plagal” progres-
subdominant functions, recognizing that they tend to occur sion).29 Going even further, Drew Nobile entirely dissociates
in the same contexts. (Notably, k ¼ 5 solutions neither traditional harmonic functions from the scale-degree content

28 Moore (1992); Stephenson (2002); deClercq and Temperley (2011).

27 McHose (1947, 4–9). 29 Biamonte (2010), Doll (2007).
 chord context and function 323

of chords.30 In Nobile’s formalization, almost any chord can The main circuit contains two primary poles, T and S, let-
function as a tonic, dominant, or predominant: “a chord’s ters chosen because of the similarity between these states and
function is given more by formal considerations—i.e., what the traditional tonic and subdominant functions. T is most
role it plays within the form—than by its internal structure or frequently represented by a I chord, and S most often produces
any specific voice-leading motion.”31 Nobile allows for pre- IV chords. These two chords are the most frequent in the cor-
dominant V chords, dominant IV chords, and so on. pus (23.5% and 21.7%, respectively). Furthermore, 22.3% of
all transitions in the corpus are between T and S. In this sense,
MATERIALS AND METHODS FOR THE MCGILL BILLBOARD CORPUS S is what we call the antitonic function: like the D/T– function
in the Kostka-Payne model, S provides the most frequent

Downloaded from [Link] by SMT Member Access user on 12 August 2019

For this study we performed an HMM analysis on the McGill transition into and out of tonic.37 The third most frequent
Billboard Project dataset, one of the largest corpora of Roman chord, V, accounts for 15.5% of the corpus, and is divided be-
numeral progressions publicly available.32 This dataset includes tween the remaining two functions in the main circuit: T1
harmonic annotations for 649 randomly chosen songs appear- and S1. The defining characteristic of T1 is that it falls be-
ing on Billboard’s “Hot 100” list between 1958 and 1991. The tween T and S; the defining characteristic of S1 is that it falls
annotations were made by human analysts and included infor- between S and T. In this sense, T1 is pre-antitonic/post-tonic
mation about root and chord quality. The entire corpus con- while S1 is post-antitonic/pretonic. While T1 can manifest a
sists of 270,366 chords. We removed annotations that number of chords other than V ([VII, vi, etc.), S1 tends to
duplicated the immediately previous annotation, leaving produce only chords closely related to V. The distinction be-
64,591 chords. The corpus’s key annotations were used to as- tween T1 and S1 is thus a distinction of both syntactic posi-
sign scale degrees: given that mode is not designated in this tion and of chord membership. The fact that there are four
corpus’s annotations and the major/minor duality is not as functions in the central circuit reflects the fact that most re-
clear in popular music as in classical music,33 we elected to use peated chord progressions in this corpus are two (T–S), three
all pieces regardless of mode. Stretches of chords in a single (T–T1–S or T–S–S1), or four (T–T1–S–S1) chords long.
key were used as the observations for the HMMs. Example 11 displays three excerpts from the model’s analy-
To ensure the statistical robustness of our models, we fol- sis of Meat Loaf’s “Paradise by the Dashboard Light.” The
lowed a cross-validation procedure: we divided the corpus into first excerpt, in D major, contains three overlapped instances
five equal parts, training each HMM on four parts (the train- of a I–IV–V–I progression, analyzed as cycles through the left
ing set) and using it to analyze the remaining part (the test set). side of the circuit (T–S–S1–T). The second excerpt, also in D
For each value of k between 2 and 30 (inclusive) and each test major, uses a different ordering of the same chords (I–V–IV–
set, we produced 300 models.34 As with our procedure for the I), traversing the other half of the circuit (T–T1–S–T): the V
Kostka-Payne corpus, we optimized k by identifying the most chord represents the S1 state in the first excerpt and the T1
consistent group of models with the added constraint of seek- state in the second. Although the contents of the two chords
ing agreement among the five test sets. The most consistent are identical, they have different syntactic relationships to the
solution involved eight states, shown in Example 10. The spe- antitonic IV chord, with S1 as the post-antitonic function,
cific probability tables can be found in Appendix B.35 and T1 as a pre-antitonic function. The second excerpt ends
with a T–T1–S–S1 progression in which the pre-antitonic
RESULTS T1 is instantiated with different content, a [VII chord; the
third excerpt shows the same functional progression (now in C
Example 10 shows a composite eight-state model of the major) with the V chord again occupying the pre-antitonic T1
McGill Billboard corpus. The diagram shows a main circuit of position.
functions labeled T, T1, S, and S1, with two peripheral pairs, Example 12 shows a chord progression that more closely
P/Q and X/W. The main circuit accounts for the substantial resembles common-practice syntax: I–vi–II–V–I. The model
majority (67.2%) of chord transitions; X/W accounts for interprets this progression as a clockwise traversal of the T–
22.7%, and P/Q accounts for only 2.3%. The remaining T1–S–S1–T circuit, with S now producing a II chord.
7.87% of transitions are accounted for by the improbable (but (While IV is by far the most frequent S-functioning chord,
possible) moves between the three circuits.36
three-function model using their “bottleneck” method. Their method for
30 Nobile (2013). discovering what they call “optimal k-categorization schemes” assumes
31 Nobile (2013, 35). that all instances of a given chord belong to a single category. What dis-
32 Burgoyne (2012) and Burgoyne, Wild, and Fujinaga (2013). tinguishes an HMM from a basic Markov process is precisely the possibil-
33 Everett (2004). ity that two instances of the same symbol might belong to two different
34 This number was chosen after several pilot studies found 300 models to categories. Given that their model’s underlying logic differs from ours and
balance robustness of the results with the computing time required. that the authors are committed to favoring simpler models, we believe
35 See online version of this article. that our findings complement theirs.
36 This contrasts with the findings of Jacoby, Tishby, and Tymoczko (2015), 37 This reflects deClercq and Temperley’s findings that IV is the most prom-
who find that popular music can be ideally described using the traditional inent nontonic chord within the harmonic language of popular music.
 324 MUSIC THEORY SPECTRUM 40 (2018)

Downloaded from [Link] by SMT Member Access user on 12 August 2019

example 10. Eight-state model of the McGill–Billboard Corpus

the model recognizes that II occurs within the same context— chord—or, more precisely, appears in a place the model recog-
between chords of the T1 and S1 functions.38) The progres- nizes that V chords often appear.40
sion itself bears a resemblance to the four-state model of the The peripheral functional pairs, Q/P and X/W, arise from
Kostka-Payne corpus: four functions accommodate a series of portions of the corpus that differ in harmonic vocabulary from
falling fifths. However, the fact that this progression expands a the major-mode, triadic music we have examined so far. The
central T–S axis rather than a dominant–tonic pair is signifi- Q/P pair tends to include chords involving sevenths and
cant. It shows the crucial distinction between the pop/rock ninths (perhaps indicating harmonic languages more influ-
and the Kostka-Payne functional paradigms: while the enced by jazz). Example 14 shows the model’s analysis of the
dominant function acts both as pretonic and antitonic in Little River Band’s “We Two.” The alternation between I9
common-practice music, these positions are distinct in the and ii7 chords shuttles back and forth between the Q and P
pop/rock corpus, with S as antitonic and either S or S1 acting functions. To a human trained in music theory, the scale-
as pretonic. degree content of these chords would indicate equivalencies
This paradigm then allows for cases such as Example 13. between I9 and I chords and ii7 and ii (or IV); with such
This song exhibits the same functional progression, but in a equivalencies in hand, we might imagine this passage as going
realization foreign to common-practice norms. The post- between the T and S functions. However, with no such equiv-
antitonic S1 function becomes [VII rather than V.39 The alencies available in our strictly context-oriented approach, the
four-function circuit interprets both progressions as exemplars model recognizes seventh and ninth chords as completely dif-
of a more general case: the expansion of the primary T–S pair ferent objects from their triadic counterparts. Since fewer
with T1 and S1 functions. Unlike the T1-functioning [VII chords in this corpus have sevenths and ninths, and these non-
chord within “Paradise by the Dashboard Light,” the [VII in triadic chords tend to progress to other nontriadic chords, the
“I Want You to Want Me” seems to substitute for a V model relegates them to their own peripheral syntax, where it
simply groups chords into two categories that follow one an-
other. The tonic-antitonic polarity does not appear to hold on
the periphery of the model: note, for example, that chords
38 These categorizations reflect the general tendencies of groups of chords— with roots of I, IV, and V appear in both P and Q.
the broader the grouping, the less the model captures the behaviors of in-
The other significant lexical minority in the corpus consists
dividual chord types. The grouping of IV and II is an example of this:
taken on its own, the II chord primarily moves to chords in the S1 func-
of minor-mode chords, which the model relegates to the X
tion, while the IV chord is more equally disposed to move toward T and and W functions. Example 15 shows an excerpt from “Funky
S1. But regardless of these differences, the model recognizes the chords’ Nassau,” a 1971 song by The Beginning of the End that is
similarity and groups them together into one function.
39 Note that [VII is a very improbable, but still possible, chord to be output 40 The former functions akin to Biamonte’s “double plagal” progression, and
by the S1 function, coming under “others” in Example 10. the latter as Biamonte and Doll’s “dominant” [VII chords.
 chord context and function 325

Downloaded from [Link] by SMT Member Access user on 12 August 2019

example 11. Meat Loaf, “ Paradise by the Dashboard Light” (1977)

example 12. Tee Set, “ Ma Belle Amie” (1969)

one of a small number of minor-key songs in the corpus. system with characteristics very different than that of the
Here, X and W appear to function as tonic and antitonic cate- common-practice. The most obvious difference is the identity
gories, respectively. In general, however, these two functions of the antitonic: while V functions as the most frequent non-
do not map onto tonic/antitonic categories; rather, as is the tonic chord in the common-practice Kostka-Payne corpus, IV
case for P and Q, each state simply comprises chords that tend fills that role in the pop McGill-Billboard corpus. This is a di-
to progress to chords in the other state. For instance, major IV rect consequence of the impact that chord frequency has on
and minor iv tend to appear in different categories, with iv the functional categories the HMM learns from the corpus.
appearing opposite the tonic and IV appearing in the same The most frequent chords in the Kostka-Payne corpus are I,
category as i but opposite [VII. V, and ii, each of which becomes the nucleus of a function. In
Several interesting topics arise from these examples, partic- the McGill-Billboard corpus, the most frequent chords are I,
ularly when comparing them to our earlier Kostka-Payne func- IV, and V. In both cases, the second most frequent chord takes
tions. Notably, because the pop-music model’s parameters on the antitonic role, while the behavior of the remaining
were learned directly from the corpus rather than adapted from chords defines the dynamics surrounding the main tonic/anti-
preconceived three-function model, it yields a functional tonic polarity, exerting influence relative to their frequency. In
326 MUSIC THEORY SPECTRUM 40 (2018)

example 13. Cheap Trick, “ I Want You to Want Me” (1975)

Downloaded from [Link] by SMT Member Access user on 12 August 2019

example 14. Little River Band, “ We Two” (1983)

example 15. The Beginning of the End, “ Funky Nassau” (1971)

the Kostka-Payne corpus, ii’s frequent preparation of the dom- practices, resulting in suboptimal peripheral functions. These
inant creates the predominant category, while in the McGill- cases reveal a significant weakness of the HMM approach: its
Billboard corpus, V’s role both before and after IV produce the underlying assumption is that a single set of probabilistic rules
T1 and S1 functions. governs the entire training set. In this case, different segments
This distribution of V over two functional categories of the corpus seem to be governed by different syntactic princi-
brings us to the second major difference between the two cor- ples, and the resulting model is tailored only to that segment
pora: where common-practice syntax contains a “dominant” with the greatest representation. Minority segments are rele-
category that acts as both antitonic and pretonic, these two gated to the periphery. We hypothesize that training separate
roles are differentiated in pop syntax, where the pretonic posi- HMMs on individual subdivisions would yield more robust
tion can be occupied either by an S chord, with antitonic and syntactic models for minor-key music, jazz-influenced music,
pretonic functions coinciding, or by an S1 chord that acts as and other distinct styles. In what follows, we examine these
pretonic and usually follows an antitonic S. issues in a completely uniform corpus, that of the Bach
Finally, our modeling method uses only contextual infor- Chorales.
mation with no knowledge of scale-degree overlap or voice-
leading similarity. The fact that purely contextual data can MODELING THE BACH CHORALE CORPUS: thirteen FUNCTIONS,
produce a workable functional model is notable in and of itself. AND SYNTACTIC DISSONANCE
It is not at all obvious that a content-blind method would be
able to reproduce workable categories at all. For instance, it Consider Examples 16(a) and (b), two excerpts from BWV
was noted that the method models the Kostka-Payne corpus 146 and BWV 402, transposed to C major for ease of compar-
with functional categories very similar to how the authors ison. The first excerpt will give a listener familiar with Bach’s
themselves describe tonal dynamics; similarly, the pop/rock chorale style momentary pause: something about the initial
results reproduce observations made by several theorists of chord progression seems unusual. In fact, in our corpus (de-
pop-music function. scribed below), only 21 of Bach’s 2,130 V7 chords are immedi-
This investigation did, however, show the difficulties that ately preceded by vi chords. In contrast, V7 is immediately
arise when analyzing a corpus containing multiple styles and preceded by IV 177 times: Example 16(b) therefore seems
 chord context and function 327

somewhat more idiomatic. Similarly, the tonic expansion in then determined by comparing the eight windows containing
Example 16(c) seems extremely idiomatic, with the chord that chord (excluding ambiguous windows). The key of the
marked with a X, {^1, ^2, ^4, ^5}, prolonging tonic. Indeed, this window having the highest confidence value was taken as the
sonority occurs between two I chords 206 out of the 392 times key of the chord. Stretches of chords in a single key were used
this sonority appears. In contrast, {^1, ^2, ^4, ^5} provides a con- as the observations for the HMMs.
trapuntal prolongation of a ii7–V7 progression only five times, Unlike the constrained vocabularies of the earlier examples
one instance of which (from BWV 244) is shown in (the Kostka-Payne corpus uses 12 chords, the McGill-
Example 16(d). Billboard uses 68), this method produced a vocabulary of 329
Notably, these constraints in Bach’s choices are not en- distinct salami-slice types. Relative to the corpus size, this vo-

Downloaded from [Link] by SMT Member Access user on 12 August 2019

tirely accounted for by the rules of counterpoint. Both IV cabulary was too large for the Baum-Welch procedure to pro-
and vi lead easily to V7, and the common tones within duce sufficiently consistent results. Initial experimentation
1, ^2, ^4, ^5} make it a natural interpolation between ii7 and
{^ with the corpus suggested three simplifying steps to reduce the
V7. If Bach’s chorales contain many such style-specific regu- number of distinct slice types within the observation sequen-
larities, then subjecting the music’s surface events—e.g., not ces. First, we considered only major-mode stretches of music
only the underlying triads, but dissonances, applied chords, (totaling 22,569 slices), to avoid the problem of “lexical minor-
and nontriadic structures—to an HMM analysis will pro- ities” discussed above in connection with the McGill-Billboard
duce a syntactic model with functional categories compara- corpus.44 Second, we ignored dyads and singletons that imme-
ble to those modeled above, but now tailored specifically to diately preceded or followed proper supersets of themselves: a
Bach’s chorales. 3–^
^ 5 dyad following a complete I chord would be ignored.
This step removed 5.6% of the salami slices. Finally, low-
MATERIALS AND METHODS OF THE BACH CHORALE CORPUS probability slices were removed from the observation sequence,
since removing these chords dramatically improved the
Before implementing the HMM, we needed to determine HMMs’ performance in pilot studies. Slices with a frequency
what kind of computer-readable representations of the Bach below the 40th percentile were removed, which amounted to
chorales were to be processed and how to divide these repre- only 2.3% of the total corpus.45
sentations into observations. We began with the chorale cor- The HMM procedure was otherwise identical to that used
pus distributed with the music21 software.41 In order to with the McGill-Billboard corpus.
remain as close as possible to the musical surface, we parsed
the surface not into Roman numerals or chords but salami RESULTS
slices: each time the pitch-class content of the chorale changes
by adding or subtracting one or more notes, a new salami slice The values of k yielding the most consistent analyses of the
is identified, consisting of a snapshot of all sounding pitch- Bach chorales were 3 and 13. Example 17 schematizes the
classes.42 Salami slices are represented as unordered sets of three-function solution (specific lexical and transition tables
scale degrees relative to the local key, which was determined can be found in Appendix B), which closely resembles the
using a preexisting key-finding process.43 The result was 35,139 three Riemannian functions. I, V, and IV are the primary
salami slices, of which 1,079 were excluded for tonal ambiguity. chords for the three hidden states, and, just as we might ex-
To assign a key to these files (and to identify modulations), pect, vi and iii stand in for tonic in some situations, while viio
we subjected the corpus to a preprocessing step. The process, sometimes acts as dominant, and several chords—ii, ii7, or
described in White (2013), used a moving window of eight vi—act as predominant. As in traditional tonal theory, the pre-
chords; the pitch-class content of each window was analyzed ferred order of functions is T–P–D–T. Example 18 shows
by a key-profile analysis implemented in music21. The two how the model reads a I–ii7–V4–3–I cadence as just such a
most likely keys for each window were determined together functional progression.
with the algorithm’s confidence in both keys (on a scale of -1 While the model conforms to many of our intuitions
to 1) and, if the confidence value for those two keys differed about harmonic function, it has some counterintuitive fea-
by less than a threshold amount (0.1), the window was deter- tures that can be attributed to the strictly context-based anal-
mined to have ambiguous tonality. The key of each chord was ysis of salami slices. In Example 18, for instance, the D
function does not arrive until the last beat of the first full
41 Cuthbert and Ariza (2011).
42 Quinn (2010).
measure. This is because the model recognizes that V chords
43 White and Quinn (2016). Note that this representation does not distin- with a suspended fourth always precede V chords and are,
guish between chords with different voicings, doublings, or inversions: therefore, categorized in the P function. Example 19(a) and
only scale-degree content matters. For example, the first two beats of
Example 16(a) contain three salami slices: assuming the key of C major, 44 This excluded the 33.7% of the corpus in the minor mode and removed
the first quarter note’s slice would be the set <0, 4, 9>, then the next two 29.4% of the chord types.
eighth-note pulses would be the sets <2, 5, 7, 11> and <2, 5, 9>, 45 This procedure is frequently used in data sets with “long tails.” See Quinn
respectively. and Mavromatis (2011) for such a discussion.
328 MUSIC THEORY SPECTRUM 40 (2018)

Downloaded from [Link] by SMT Member Access user on 12 August 2019

example 16. (a) BWV 146b, (b) BWV 402, (c) BWV 245, (d) BWV 244

example 17. The three-function model

(b) illustrate further peculiarities of the model. In Example The three-function solution, then, produces a model with
19(a), apparent fourth-inversion I9 chords—verticalities that traditional tonal relationships along with some nontraditional
pass between two tonic triads—function as D, since these instantiations of those relationships. The asymmetry of the
sorts of chords always progress to T chords. Notice that even transitions between functions is familiar: the P–D transition is
though the second of these I9 chords does not proceed to I, it unidirectional, while T–P and T–D transitions are basically
does go to vi, a chord that may function as T. Similarly, the bidirectional, with a disposition for T–P–D–T motion.46
penultimate verticality of Example 19(b), {^1, ^4, ^6, ^7}, always However, the model sometimes applies these functions and
proceeds to I or vi chords and is, therefore, analyzed as D.
Both excerpts also include “predominant” V chords: the
model recognizes that V often precedes chords that them-
46 This results in a very messy predominant function. While 70% of tonic
selves precede T chords. In both examples, the V chord is chords are I triads and more than 35% of dominant chords are V triads,
followed by the addition of ^4 to the texture, creating the viio the most frequent predominant chord, IV, only instantiates that function
and V7 slices, respectively. around 16% of the time.
 chord context and function 329

Downloaded from [Link] by SMT Member Access user on 12 August 2019

example 18. BWV 10.7, mm. 12–14

example 19. (a) BWV 127.5, m. 1, (b) BWV 10.7, m. 11

transitions to slices that we would not ordinarily consider as Example 24 shows the models in analytical action. The
function-bearing chords. three-state analysis assigns hidden states in a way that conforms
A distinct advantage of the thirteen-function solution (see to our intuitions, moving through two complete P–D–T cycles.
Ex. 20) is that it adds new functional categories that accom- Functions are sometimes instantiated by more than one chord,
modate such slices. This model retains the basic tonic-centric and the tonic of the first cycle is expanded by a passing domi-
circular flow we have seen in many of our HMM solutions, nant. The thirteen-state analysis distinguishes between the non-
but adds parallel pathways and detours. Example 21 illustrates tonic functions of the first measure and those of the second.
some idiomatic progressions of these pathways. The top of The initial measure uses the weak predominant (p) and domi-
Example 20 shows T (tonic), a function which contains mostly nant (d) functions, and both p and T are prolonged by their
I triads along with several vi and iii chords. Tx (tonic expansion) corresponding expanding functions. The cadential part of the
is composed of slices resulting from passing and neighbor mo- phrase uses the strong predominant (P) and dominant (D)
tion between two T slices, including the nontriadic {^ 1, ^ 4, ^
2, ^ 5} functions. The phrase, on the whole, moves first through the in-
of Examples 16(c) and 21(a). Example 21(b) shows the pri- ner loop of Example 20, and then through the outer loop.
mary pathway of “strong” functions around the outside of the Example 25 investigates several other functions within the
diagram, so named because of their phrase-ending cadential thirteen-state model. The phrase as a whole begins in E major
function. Here, T first moves to P (strong predominant), to and leads to what turns out to be a fleeting tonicization of the
D—the strong dominant—and then to D1, the late dominants, relative minor. Our model’s thirteen-state analysis shows three
comprised mostly of V7). The cadential progression I–ii–V8–7 functions (T, p, and R), each prolonged by their expanding
would traverse this outer pathway, as would the V-chord sus- functions. The tonicization of the submediant is treated as a
pension of Example 21(b). within-key phenomenon.
The inner cycles show progressions that tend to precede ca- Example 26 highlights the remaining functions available to
dential progressions, or the “weak” functions. Example 21(c) the thirteen-state system. The first measure involves a tonic
shows such a progression. First, the I42 chord functions as a late
tonic T1, a passing function that progresses from T to p, the
middle-of-phrase weak predominant. The px function (predom- 47 There are two further important characteristics concerning the x/ii func-
tion. First, in the thirteen-state model, this function involves both ii and
inant expansion) then prolongs the p function with its passing
its applied chords, while a fourteen-state model divides these two catego-
chord. Finally, the model involves three tonicization detours. ries (just like R and Rx in the thirteen-state model). This functional divi-
Example 22 shows how V, vi, and ii can receive their own ton- sion is the primary difference between thirteen- and fourteen-state
icization functions, and Example 23 summarizes the thirteen models. Second, it is notable that the models distinguish between ii as a
functions.47 tonal area and ii as a predominant chord.
 330 MUSIC THEORY SPECTRUM 40 (2018)

Downloaded from [Link] by SMT Member Access user on 12 August 2019

example 20. The thirteen-state model

example 21. Some representative instantiations of the thirteen functions: (a), (b), and (c)

expansion using first a late tonic followed by a p-functioning hand, the fact that the model does not recognize bass notes might account
iv7 chord.48 The second measure involves a dominant for the model’s apparent difficulty with I chords preceding cadences. But,
if this progression happened often enough, we would imagine that the
HMM would categorize those I chords that follow ii and IV chords and
48 Note that the following passing I chord is labeled as a T: this is because precede V7 chords as D or P functions. However, Bach simply does not
the model labels all I chords as T due to the overwhelming frequency of I use enough cadential 64 s for the model to incorporate this behavior. Only
chords occurring in the early tonic context. Another notable issue of the 30 of Bach’s 455 I–V progressions are in the proper inversion to be con-
model struggling to analyze a I chord involves cadential 64 chords. On one sidered cadential candidates.
 chord context and function 331

Downloaded from [Link] by SMT Member Access user on 12 August 2019

example 22. Some further representative instantiations of the thirteen functions: (a) and (b)

example 23. A summary of the thirteen functions

 332 MUSIC THEORY SPECTRUM 40 (2018)

Downloaded from [Link] by SMT Member Access user on 12 August 2019

example 24. BWV 115.6, mm. 3–4

example 25. BWV 124.6, mm. 8–9

expansion and late dominant function, and the final eighth and by the HMM method itself. The expectation-maximization
note of that measure shows the “early tonic” capacity of the Tx procedure for training the HMMs produces a model that
function, intervening between the cadential V7 (D1) and the assigns a high probability to observation sequences in the train-
cadential tonic (T). The second half of the third measure ing set, and this preference favors models that make strong pre-
briefly tonicizes the minor supertonic, something the thirteen- dictions about the corpus’s most frequent chords. High lexical
state model labels as a patch of sx states.49 This model treats probabilities associated with common chords and high transi-
tonicizations of ii differently from tonicizations of V and vi. tion probabilities associated with common hidden-state progres-
Each of the latter two Stufen has separate states for the tonic- sions will result in high probability estimates for observation
izer and the tonicized, but tonicizations of ii are rare enough sequences. In other words, because a hidden state that is equally
that the model only learns a single state for all chords involved likely to produce any of a handful of observable chords will re-
in tonicizations. sult in low probability estimates for any of those chords, the
Baum-Welch algorithm prefers high lexical and transition prob-
COMPARING AND CONTRASTING THE MODELS: THREE
abilities associated with the most frequent hidden states. This
GENERALIZATIONS
corresponds to a minimization of lexical diversity in the most
common hidden states: note that the larger pie charts in the
The differences between the models makes it difficult to gener- examples tend to have fewer, larger slices.
alize about context-oriented harmonic function, and the fact However, several characteristics arise neither because of the
that we present only three corpora makes generalizations about properties of HMMs nor the Baum-Welch algorithm, but be-
musical syntax even more suspect. Although our discussion so cause of the properties of the corpus. These recurrences are no-
far has primarily emphasized the differences between the con- table not only because they arise within all the corpora
textual regularities of each corpus, our models have some com- investigated, but because there is no reason to expect our
monalities that are shaped both by the properties of the corpora methods to produce models with these characteristics.
First, each model’s top two most frequently occurring
49 Recall that sx also involves ii itself. In the fourteen-state model, sx is di- chords create two functions that strongly associate with one
vided into two states, one of which is solely ii, while the other comprises another in some way (e.g., the I and V chords in the Kostka-
applied chords that move to ii. Payne and Bach corpora, and the I and IV chords in the
 chord context and function 333

Downloaded from [Link] by SMT Member Access user on 12 August 2019

example 26. BWV 157.5, mm. 5–8

McGill Billboard corpus). In each case, the most frequent pie slices show the proportional frequency with which each
chord, I, creates a recognizable tonic function and, as such, scale degree occurs in each function: pies with fewer, larger sli-
creates an important pillar in the functional system. The sec- ces are dominated by fewer scale degrees. The pies are ordered
ond most frequent chord, then, provides a secondary pillar, from left to right by their ascending common-tone scores, indi-
and its relationship to tonic creates one of the most defining cated on the y axis. The score is a measure of the degree to
transitions of the model. We will refer to this general property which chords in the function share scale degrees. It is not a
as the tonic/antitonic dichotomy: each corpus has two hidden raw count of common tones, but a scaled and normalized mea-
states built around two most frequent chords in the corpus. sure: a score of zero corresponds to the average number of
Second, each corpus has a unidirectional relationship be- common tones between any two chords randomly chosen from
tween a state or states that move into tonic but to which tonic the corpus. A positive score for a function means that chords
progresses less frequently. We could imagine this as the within that function have a greater than average number of
“cadential progression”; however, not wanting to import the common tones; a negative score means a lower than average
meanings inherent in that term, we will adopt the term number of common tones. The score is scaled in terms of stan-
pretonic/tonic relationship. In the Kostka-Payne and Bach cor- dard deviations of the distribution over the corpus as a whole
pora, this dynamic involves the dominant states progressing (what statisticians call a z-score).50
to the tonic states, and in the popular music corpus the S1– Nearly every function in each corpus has a positive
T progression instantiates this relationship. Notably, while common-tone score, indicating that a small number of scale
these dynamics overlap with the tonic/antitonic dichotomy degrees dominates each function. In the Kostka-Payne corpus,
for the Bach and Kostka-Payne corpora (the dominant-tonic for instance, the T hidden state is dominated by ^1, ^3, and ^5,
“cadence”), the two relationships do not overlap in the pop- while the P hidden state uses primarily ^ 1, ^
2, ^4, and ^6. At first
music corpus: the unidirectionality of the S1–T “cadence” is glance this is not a surprising finding; however, recall that our
not the same as the S–T bidirectional tonic/not-tonic models are strictly context-based and thus entirely blind to
dichotomy. scale-degree content. In fact, a closer look at the distributions
Finally, each model contains less-probable pathways that, seems to suggest that certain scale-degree combinations seem
rather than constituting their own systems, expand the pri- to connote certain functions. In the Kostka-Payne corpus, for
mary tonic–antitonic or tonic–pretonic pair. These pathways instance, ^4 prominently occurs in two functions; however,
comprise the prolongational networks that either act as precur- when it occurs alongside ^ 6 it would likely appear in the P
sors or successors to the model’s most frequent pathways. For function, while if it appeared with ^ 7, that sonority would likely
instance, adding S1 between S and T in the McGill- be classified in the D function.51 Despite this limitation, our
Billboard model prolongs the tonic–antitonic dichotomy simi- models reveal a deep connection in our corpora between a
lar to the way that the P and P– prolong the T–D/ chord’s scale-degree content and its contextual tendencies.
T–transitions of the Kostka-Payne model. On a larger scale, Based on these results, the basic principle of harmonic func-
we could even imagine the “weak” inner pathways of the thir- tion that identifies function with scale-degree content seems to
teen-state Bach model prolonging the cadential “strong” pro- hold in most cases.
gressions of the outer pathways that provide the model’s
antitonic/pretonic function.
50 The process did not compare identical chords to one another. Our test is
designed to ask the question, “How much overlap exists between noniden-
COMMON TONES, FREQUENCY, AND WHAT IT MEANS TO BE A
tical chords within a function?”, removing any effect that a single chord’s
FUNCTION
overwhelming frequency within a particular function might have. We
have represented this property in terms of standard deviations (a z-score)
Example 27 shows several pieces of information concerning in order to use the same scale between corpora.
the scale-degree content of each corpus’s hidden states. The 51 This topic is taken up in more depth in Quinn (2017).
334 MUSIC THEORY SPECTRUM 40 (2018)

Downloaded from [Link] by SMT Member Access user on 12 August 2019

example 27. Common tones in each functional model: (a) The Kostka-Payne Corpus; (b) The McGill Billboard Corpus; (c) Bach Chorale
Corpus, 3 states; (d) Bach Chorale Corpus, thirteen states

But content and context do not always coexist peacefully. model. These cases are, however, in the minority: in fact, con-
The D/T– category in the Kostka-Payne corpus has a lower- text- and content-oriented functional theories seem to overlap
than-average common-tone score: chords that progress to T more than we might at first expect.
can have various scale-degree structures. Nowhere is this more But what of the hierarchies and qualities we usually associ-
evident than the “dominant” IV of the Kostka-Payne corpus: ate with functions? For instance, we usually couple tonic with
our models’ analysis in this case suggests that a completely the quality of resolution and accept it as the deepest structural
contextual definition of function might conflate traditional harmony. Our results support previous research that connects
subdominant and dominant functions. The identical logic these sorts of qualities and hierarchies with the frequency with
applies to the T1 function in the popular music corpus, and which we hear stimuli. Over the past several decades, the work
the sx and R functions in Example 20: in all these cases, the of many researchers in the field of music cognition, most nota-
diversity of chords within these categories creates a disjunction bly that of Krumhansl, Huron, and Aarden,52 has made a
between a content-oriented theory and our context-oriented compelling case that our perceived hierarchy of scale degrees is
 chord context and function 335

Downloaded from [Link] by SMT Member Access user on 12 August 2019

example 27. [Continued]

due to the frequency with which we hear them in music, recommendations concerning the general discourse surround-
especially at the ends of phrases. Their research has shown ing harmonic function.
that the most frequent scale degree, tonic, is imparted with a First, our models and analyses have shown degrees of sub-
feeling of resolution and hierarchical superservience; in con- tlety, stylistic regularities, and syntactic categories that are not
trast, chromatic scale degrees occur very infrequently and feel captured by traditional three-function harmonic theory. This
hierarchically subservient, unsettled and unresolved.53 It fol- can stem from the decoupling of proteinic and antitank roles
lows that the same might be true for harmonic functions. Our (as in the popular music repertoire). It can also result in the di-
model’s tonic categories may be more fundamental or struc- vision of predominant into more than one syntactic category
tural simply because they are more frequent. Antitonic chords (as in the P– chords in the Kostka-Payne corpus), the articula-
may be viewed as hierarchically secondary inasmuch as they tion of unique passing and neighboring functions (as in the
participate in the second-most-frequent function. Similarly, Bach corpus), or even some other contextual regularity not
the pretonic functions are only hierarchically subservient or addressed in the current work.
points of highest tension because they precede the most- Second, we suggest that discussions of harmonic function
frequent function. In sum, relative frequency produces a sense be moored to particular repertoires and be sensitive to differen-
of relative stability. ces between corpora. As theorists and teachers, we should em-
phasize the syntactic norms of a corpus rather than universal
rules. Instead of a tonic–subdominant–dominant paradigm
CULTURAL AND PEDAGOGICAL IMPLICATIONS that reaches from the Baroque to the Beatles, we might em-
phasize the constrained categories of chords that Bach deploys
Importantly, we are advocating for a theoretical framework at phrase endings, versus the recurrent riffs in rock music.
rather than any particular instantiation of that theory. We be- Assuming the three Riemannian functions as harmonic uni-
lieve that contextual modeling can provide insights into the versals is at best an oversimplification and at worst culturally
concept of harmonic function but remain open to revisions, hegemonic. Importing a model associated with the German-
caveats, and addenda to our specific models. Our findings language common practice inspired by Hegelian dualism onto
therefore apply more to how we approach teaching or discus- other culturally specific repertoires problematically asserts the
sing harmonic function rather than to precisely what to teach power of one culture over another. By deriving our models
or discuss. For instance, we readily acknowledge that any from preexisting and wide-ranging corpora, our methods pro-
classroom-oriented theory of harmony should involve bass vide a way to sidestep the difficult implications of using a
lines and melodic progressions, both absent in the current “common practice” as the main driver of theory and as a yard-
models. However, we will close by offering two stick for tonal norms.

52 Krumhansl (1988, 1990); Huron (2006); and Aarden (2003).

53 See Huron (2006) for a complete list of scale-degree qualities and the sta-
tistical rationale for each.
335a MUSIC THEORY SPECTRUM 40 (2018)

APPENDIX A Example A2 shows a Hidden Markov Model. Added to

the Markov model are several observations—x, y, and z—and
In this appendix, we present the article’s formal models and
A, B, and C are now the hidden states. The transitions follow
algorithms in more details, including Hidden Markov Models,
the same logic as before: the probability of the hidden-state
expectation-maximization, and silhouette width. Mathematical
sequence <A, B, A> would still be 8.1%. However, the
formalizations of these processes can be found in Jurafsky
observations add another layer. As in this article’s functional
and Martin (2008).
analyses, such a two-layered model is used to make judg-
Example A1 shows a simple Markov model that uses three
ments about something you can’t currently observe using
states: A, B, and C. Each letter progresses to the next in the
something you can. For instance, economists might use ob-

Downloaded from [Link] by SMT Member Access user on 12 August 2019

alphabet with a 90% probability, and backward only 10% of
servable weather temperatures to predict demand for air con-
the time. We can use the model to assess the probability of
ditioners; a spell-checking model might observe a typed
some sequence: the progression <A, B, C> would be 72.3%
misspelling and then replace it with the intended autocor-
probable, (.9 x .9 x .9), whereas <A, B, A> would be 8.1%
rected word. In these models, the hidden state depends on
likely (.9 x .9 x .1). Markov models are useful at assessing
the lexical probabilities connecting the hidden state and the
the probability of series of simple observations. If we, say,
observation (i.e., how likely is a hot day to inspire air condi-
knew how likely individual chords were to progress to one
tioner sales, or is it more probable that I meant “soup” or
another, we could assign a probability to sequences of chords,
“soap” when I typed “can of sop”?). In HMMs, the hidden
and therefore know which sequences were more probable
states also depend on the preceding hidden states, or the
than others given the model.
transition probabilities. For instance, today’s air conditioner
sales are influenced by yesterday’s, and the best autocorrected
word is dependent on the previous words. In Example A2, if
we observed the string <x, y, x>, there are six options for
hidden-state sequences: <A, B, C>, <A, B, A>, <B, A,
B> <C, B, C>, <C, A, C>, and <C, A, B>. If we assess
the probability of each, given the model’s lexical and transi-
tion probabilities, we would realize that <C, A, B> provides
the highest combined probabilities. It is the only option that
traverses only 90% pathways. The maximum probability of
the sequence <x, y, x> given the model would be the prod-
uct of transition probabilities of the sequence <C, A, B>
(or, .9 x .9 x .9) and the probabilities of x, y, and x being
produced by those respective hidden states (again, .9 x .9 x
example a1. A Markov Model

example a2. A Hidden Markov Model

 chord context and function 335b

Downloaded from [Link] by SMT Member Access user on 12 August 2019

example a3. A Hypothetical Improvement to Example A2

.9). The reader is encouraged to experiment with the model, string. In order to claim that a model is generalizable and
proposing observations and adducing the hidden states. not over fitted to the peculiarities of a particular observation
HMMs can also derive their parameters from a body of string, the training and test corpus should not be the same
data. A model can compile lexical and transition probabilities data set. For the current article, we use the Viterbi decoding
by tallying how frequently observations accompany hidden algorithm for this task (again, formalizations appear in
states and how often the hidden states follow one another. Jurafsky and Martin 2008).
Given a body of data connecting air conditioner sales with As an example, consider the “good” two-state HMM and
temperature or typos with intended spellings, a programmer toy corpus in Example A4. The example includes (a) the lex-
could model those connections using an HMM. ical probability table, (b) the transition probability matrix,
However, in the current article, we begin only with a series and (c) the probabilities assigned by model to the observation
of observations—chords—and attempt to find a model to as- string. Here, the model includes a category A that has a
sess the unobserved hidden states—functions—not even 100% probability of manifesting I chords, a 50% probability
knowing how many hidden states underlie the observations. to transition to itself, and a 50% probability of transitioning
To do this, we use an iterative process of trial and error to to state B. B, on the other hand, shows V half the time and
determine which models best explain the observation series, V7 the other half of the time; B transitions to A 100% of the
called the training corpus. One could imagine, for instance, time. The probability for the sequence would be the product
that if we tried to explain an observation sequence that only of each lexical and each transition probability. (Note that V
ever moved from x to y to z back to x, the model in and V7 are grouped into the same hidden state not because
Example A2 would perform poorly. The (admittedly trivial) of their common tones, but because they occur in the same
model of Example A3 would return higher probabilities, contexts.) Compare this model to Example A5, and the
namely always 100%. This process of tweaking the transition “poor” two-state HMM. The toy solution now groups V7
and lexical probabilities within a model to increase the fit be- with I in state A, and both states move between one another
tween the model and the observations is called expectation with equal probability. The transition and lexical probabilities
maximization. In this paper, we use the Baum-Welch expec- of observed and hidden states given the model are lower
tation maximization algorithm, which cycles through possible (worse) than those in the Good two-state solution. Because
lexical and transition settings, trying to improve the probabil- the good solution has a higher probability than the poor solu-
ity of the analysis given the model for each iteration. The al- tion, an expectation maximization process would discard the
gorithm returns a final model either after a certain number of latter in favor of the former. In our modeling, we will prefer
cycles, or after it starts returning diminishing improvements solutions like Example A4 over Example A5 for this reason.
(this being a parameter that can be set by the analyst). While the comparison between the Good and Poor two-
After training, the HMM can then be used to analyze a state models illustrates how the algorithm chooses between
test corpus. During the testing phase, an HMM assigns the different possible solutions, the process also must choose be-
most probable Hidden States to the test corpus’ observation tween different numbers of hidden states. That is, how would
335c MUSIC THEORY SPECTRUM 40 (2018)

a) The HMM’s Lexical Probabilities:

State A B

Observed I 100% 0%
Chord

Downloaded from [Link] by SMT Member Access user on 12 August 2019

V7 0% 50%

V 0% 50%

b) Transition Probabilities (shown as the probability of row moving to column)

A B

A 50% 50%

B 0% 100%

c) The HMM’s analysis of a toy corpus

Observations: I V7 I I V I I I V7 I I V I I I V7 I
Good 2-State: A B AABAAABAABAAABA
Lexical Probs: 1 .5 1 1 .5 1 1 1 .5 1 1 .5 1 1 1 .5 1 = .0625
Transition Probs: .5 1 .5 .5 1 .5 .5 .5 1 .5 .5 1 .5 .5 .5 1 = .00049

example a4. Good two-state HMM analyzing Toy Corpus no. 1

we know that two states optimally underpin the observation To quantify this, we use a measurement called silhouette
string, as opposed to three, four, or more? width, a calculation used in cluster analyses. The silhouette
Consider the two hypothetical Good four-state analyses in width of point i can be found first by computing a(i), the av-
Example A6. Both distinguish between I, V7, and V as dif- erage distance between i and each other point in its cluster.
ferent functional entities (states A, B, and C), and both iden- The a(i) therefore tells us how well i is matched with other
tify the fourth state, D, as one which outputs I chords that points in its own cluster. We then compute b(i), the average
occur in particular relationships to states B and C. In the first distance between i and the points in its closest neighboring
model, D occurs before B or C, and in the second it occurs cluster. The b(i) therefore tells us how distinct i is from its
after. If the lexical and transition probabilities were constant closest cluster. The silhouette width then subtracts a(i) from
between both models and the latter only differed from the b(i) and divides by the larger of the two. Assuming the point
former in whether D proceeds to or from B and C, the prob- is situated in the best cluster, the number will always be be-
ability of the sequence would be identical between both mod- tween zero and one (if the point is in the wrong cluster the
els. It would not be clear which model represented the better number will be below zero). The closer the value is to one,
option for an expectation maximization procedure. Because the better the fit to its cluster. For an overall measure of an
there is no obvious solution, the algorithm would sometimes entire clustering solution, the average silhouette width simply
produce the first solution, and sometimes the second. This averages of the silhouettes of all the individual points.
lack of consistency suggests that four is not an ideal number The current article uses each model’s analyses of the hidden
of states to produce this observation sequence. In our model- states of the test corpus to create our cluster analyses. For
ing, we will prefer solutions like Example A4 over Example each chord in the test corpus, every HMM with a particular
A6 for this reason. cardinality of hidden states assesses the chord’s state, and
 chord context and function 335d

a) The HMM’s Lexical Probabilities:

State A B

Observed I 80% 20%

Chord

Downloaded from [Link] by SMT Member Access user on 12 August 2019

V7 20% 0%

V 0% 80%

b) Transition Probabilities (shown as the probability of row moving to column)

A B

A 50% 50%

B 50% 50%

c) The HMM’s analysis of a toy corpus

Observations: I V7 I I V I I I V7 I I V I I I V7 I
Poor 2-State: A A ABBAAB B ABB AA B A A
Lexical Probs: .8 .2 .8 .8 .2 .8 .8 .2 .8 .8 .2 .8 .8 .8 .2 .2 .8 = .0000055
Transition Probs: .5 .5 .5 .5 .5 .5 .5 .5 .5 .5 .5 .5 .5 .5 .5 .5 = .0000153

example a5. Poor two-state HMM analyzing Toy Corpus no. 1

Observations: I V7 I I V I I I V7 I I V I I I V7 I
Good 4-State #1: A C ADBAADC ADBA AD C A
Good 4-State #2: A C AABDAAC DABD AA C D

example a6. Toy Corpus no. 1 analyzed by two Good four-State HMMs

that sequence of assessments is used to produce a dissimilar- particular iii chord in the test corpus might have a vector
ity matrix, which is then in turn used for a k-means cluster more similar to tonic, while another might be more similar
analyses where k equals the number of hidden states used in to the dominant. Each observation’s vector—each chord’s
the HMM. This process allows us to measure consistency in string of hidden states—within the test corpus is placed in a
a situation where “the first hidden state” might mean some- matrix, where the dissimilarity between each observation is
thing different for each training/test session. For instance, if calculated. The dissimilarity between [ADBCA] and
we were modeling five runs of the HMM process using four [BACBB] would be five, since there is no overlap. The dis-
hidden states, a possible vector might be [ADBCA], suggest- similarity matrix now captures units of difference between
ing that the first HMM assigned the slice to state A, the sec- how the models analyze every observation in the test cor-
ond to state D, and so on. If this were a I chord, one might pus.54 The cluster analysis is then performed using this ma-
imagine that most other I chords would receive the same or trix, dividing the observations into the number of hidden
similar analysis vector. Again, note that the state assignments states within the constituent HMMs. The result is an
are arbitrary: “A” might be “tonic” in one HMM, while “D”
could be tonic in another. A “dominant” slice would then 54 The current work squares the dissimilarities, simply to make the differen-
have a completely different vector, say, [BACBB]. A ces more pronounced, which in turn makes the data easier to work with.
335e MUSIC THEORY SPECTRUM 40 (2018)

Downloaded from [Link] by SMT Member Access user on 12 August 2019

example a7. Silhouette widths of each cluster in the Kostka-Payne analysis

example a8. Silhouette widths of each cluster in the McGill-Billboard analysis

example a9. Silhouette widths of each cluster in the Bach-chorale analysis

aggregated picture of which chords tend to be analyzed by diagram and report are summations of the agreement be-
which hidden states. tween the 300 individual models, and—since the clusters
Example A7 shows the silhouette widths that result from capture the emergent properties of the constituent HMMs—
the article’s Kostka-Payne HMMs. The widths capture the we create our final models by treating the clusters like com-
amount of agreement—the tightness of the clustering—of posite functions. Since our clustering involves actual chords
the 300 models produced for each number of hidden states. within our test set, we treat each cluster of chords as exem-
Generally, the widths need to be read as relative to those plifying its own function. The transitions between the clus-
around them since fewer clusters will tend to produce higher ters’ constituent chords constitute the lexical probabilities,
overall values. Therefore, we look for peaks within the con- and the chords associated with each cluster become the emis-
tour: a peak will mean that the models improved their consis- sion probabilities.
tency compared to the predicted decline when adding Examples A8 and A9 show the silhouette widths of the five
clusters. (NB: there is no standard way of quantifying these different test/training pairs for the popular music and Bach cho-
peaks.) In Example A7, the widths peak at four clusters/ rale studies. If peaks are reproduced over several trials, we con-
states, indicating this provides the most consistent group of sider the clustering robust. Note that Example A8 has consistent
models, and therefore the preferred number of states for this peaks at k ¼ 8, while A9 has a consistent peak at k ¼ 3, and a
corpus is four. Throughout the article, the models we more subtle peak around k ¼ 13 or 14.
 chord context and function 335f

APPENDIX B

PROBABILITY TABLES

D/T– T P- P End

D/T– 0.4% 14.2% 0.4% 0.6% 0.1%

Downloaded from [Link] by SMT Member Access user on 12 August 2019

T 15.0% 0.0% 4.3% 5.4% 4.6%

P– 0.0% 0.4% 2.5% 15.5% 0.4%

P 26.7% 0.3% 8.4% 0.3% 0.3%

example b1. Transitions between hidden states of Kostka-Payne four-function model. Table sums to 100%, background colors are graded
to reflect increasing probabilities.

Function

chord root D T P– P

1 0.00% 91.84% 0.00% 0.00%

^
#1 0.48% 0.00% 10.71% 0.00%

^
2 1.90% 0.00% 4.76% 65.00%

^
b3 0.00% 0.00% 0.00% 1.67%
#2

3^ 0.00% 2.04% 16.67% 0.00%

^4 9.52% 0.00% 9.52% 17.50%

^
#4 1.43% 0.00% 1.19% 6.67%

^
5 70.00% 0.00% 0.00% 0.00%

^
b6 1.43% 1.53% 3.57% 3.33%
#5
^
6 1.43% 1.53% 51.19% 0.00%

^
7 12.38% 0.51% 1.19% 0.00%

Other 1.43% 2.55% 1.19% 5.83%

example b2. Roots in Kostka-Payne associated with each hidden state (lexical probabilities). Columns sum to 100%.
335g MUSIC THEORY SPECTRUM 40 (2018)

P Q R T S U X W End

P 0.5% 2.9% 0.1% 0.0% 0.0% 0.1% 0.0% 0.0% 0.3%

Q 2.7% 0.9% 0.1% 0.1% 0.2% 0.0% 0.1% 0.0% 0.5%

Downloaded from [Link] by SMT Member Access user on 12 August 2019

R 0.0% 0.0% 0.0% 10.3% 2.8% 0.1% 0.0% 0.1% 0.6%

T 0.1% 0.1% 3.5% 0.3% 10.3% 10.0% 0.0% 0.0% 1.6%

S 0.1% 0.1% 9.6% 12.0% 0.1% 0.5% 0.0% 0.0% 1.1%

U 0.1% 0.1% 0.0% 1.8% 9.1% 0.4% 0.0% 0.0% 0.7%

X 0.0% 0.0% 0.0% 0.1% 0.0% 0.0% 0.6% 5.9% 0.6%

W 0.0% 0.1% 0.0% 0.1% 0.1% 0.0% 6.0% 1.6% 0.5%

example b3. Transitions between hidden states of McGill-Billboard eight-function model. Table sums to 100%, background colors are
graded to reflect increasing probabilities.
 chord context and function 335h

Functions

chords P Q R T S U X W

I 26.8% 0.0% 0.0% 83.6% 0.0% 0.0% 8.2% 2.0%

Downloaded from [Link] by SMT Member Access user on 12 August 2019

IV 0.0% 2.3% 0.0% 0.0% 85.0% 0.0% 20.2% 0.0%

V 0.0% 10.5% 80.9% 0.0% 0.0% 30.2% 0.3% 0.2%

bVII 2.1% 0.9% 0.7% 3.2% 0.0% 22.0% 0.0% 23.3%

vi 2.1% 0.9% 0.0% 4.1% 0.4% 16.5% 0.0% 0.0%

V sus 4 0.0% 0.0% 9.1% 0.6% 0.4% 6.8% 0.0% 0.0%

i 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 27.3% 0.0%

bVI 0.0% 0.0% 0.0% 0.6% 0.5% 0.0% 8.0% 10.6%

II 0.5% 0.0% 0.0% 0.3% 4.8% 2.2% 0.0% 0.0%

ii 1.1% 0.0% 0.0% 1.1% 2.8% 2.4% 1.4% 0.0%

iv 0.0% 1.4% 1.9% 0.2% 0.9% 0.0% 0.0% 8.9%

V7 0.0% 11.4% 3.7% 0.0% 0.7% 0.3% 0.3% 0.0%

iii 0.0% 0.5% 0.1% 2.0% 0.0% 4.4% 0.0% 0.0%

i7 0.0% 2.7% 0.0% 0.0% 0.0% 1.2% 0.0% 9.9%

example b4. Chords in Kostka-Payne associated with each hidden state (lexical probabilities). Columns sum to 100%. Only chords that
occur thirty times or more within the corpus are shown (>.05% of the corpus’s chords).
335i MUSIC THEORY SPECTRUM 40 (2018)

I no 3rd 0.0% 0.0% 0.0% 0.4% 0.0% 0.0% 11.1% 0.0%

iii 7 1.1% 1.4% 0.0% 0.8% 0.0% 4.6% 0.0% 0.0%

vi 7 3.2% 10.0% 0.0% 0.7% 0.3% 0.2% 0.3% 0.0%

Downloaded from [Link] by SMT Member Access user on 12 August 2019

v 0.0% 2.7% 0.1% 0.0% 0.0% 0.7% 3.4% 4.5%

ii7 15.8% 0.5% 0.1% 0.0% 0.6% 0.0% 0.0% 0.2%

III 0.0% 0.0% 0.0% 0.5% 0.0% 1.0% 0.9% 5.7%

IV9 0.0% 0.9% 0.0% 0.0% 3.2% 0.0% 0.0% 0.0%

v7 3.2% 3.2% 0.4% 0.0% 0.0% 1.7% 0.3% 2.2%

I9 0.0% 15.1% 0.0% 0.0% 0.1% 0.0% 0.0% 0.0%

^ ^ ^
{ 1, 2, b7} 0.0% 0.0% 0.0% 1.2% 0.0% 0.2% 0.0% 4.0%

example b4. [Continued]

D P T End

D 5.61% 0.91% 25.10% 0.49%

P 17.07% 22.61% 0.49% 1.27%

T 8.25% 16.73% 0.17% 1.30%

example b5. Transitions between hidden states of Bach Chorale three-function model. Table sums to 100%, background colors are graded
to reflect increasing probabilities.
 chord context and function 335j

^ ^^
Functions { 1 , 4, 5} 1.2% 1.7% 0.0%

chords D P T II7 0.0% 2.2% 0.0%

I 0.0% 0.0% 73.1% IV9 0.0% 2.2% 0.0%

V 30.9% 5.7% 0.0% V/vi 0.5% 1.3% 0.6%

Downloaded from [Link] by SMT Member Access user on 12 August 2019

IV 3.0% 15.8% 0.0% { 1^, ^2, 4^} 1.1% 0.9% 0.0%

V7 21.1% 0.0% 0.0% vii/vi 0.0% 1.7% 0.0%

vi 2.0% 8.3% 5.4% V7 no 5th 2.1% 0.0% 0.0%

ii 0.5% 9.3% 0.8% V7/IV 0.0% 1.5% 0.0%

ii7 1.6% 6.1% 0.0% V7/vi 0.0% 1.2% 0.4%

^ ^ ^
iii 1.9% 3.0% 4.0% { 1, 5, 6} 0.0% 0.5% 1.5%

vii 7.6% 0.4% 0.0% V add 4 0.7% 0.8% 0.0%

IV7 0.0% 4.8% 0.0% { ^1 , 2^, 6}

^
0.0% 1.3% 0.0%

^ ^ ^
{ 1, 2 , ^5}
^ ^
1.2% 3.5% 0.0% { 2, 4, 5} 1.6% 0.0% 0.0%

^ ^ ^ ^
vi 7 0.0% 4.1% 0.2% { 1, 4, 6, 7} 1.4% 0.2% 0.0%

vii 7 5.0% 0.2% 0.0% V/iii 0.5% 0.8% 0.0%

I7 0.2% 3.3% 0.1%

example b6. Chords in Bach Chorales associated with the three hidden states (lexical probabilities). Columns sum to 100%. Only chords
that occur nineteen times or more within the corpus are shown (>.05% of the corpus’s chords).
335k MUSIC THEORY SPECTRUM 40 (2018)

T+ R Px p d T Tx P D sx Rx D+ Dx End

T+ 0.2% 0.5% 0.3% 3.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.3% 0.0% 0.0% 0.3% 0.0%

R 0.1% 0.0% 0.9% 0.4% 0.1% 0.6% 0.0% 0.2% 0.5% 0.2% 1.1% 0.0% 0.2% 0.1%

Downloaded from [Link] by SMT Member Access user on 12 August 2019

Px 0.1% 0.1% 0.2% 1.3% 0.6% 0.0% 0.0% 0.6% 0.6% 0.5% 0.0% 0.3% 0.1% 0.1%

p 0.0% 0.0% 2.6% 0.0% 3.6% 1.2% 0.0% 0.5% 1.0% 0.1% 0.0% 1.0% 0.1% 0.4%

d 0.0% 0.2% 0.1% 0.1% 0.0% 3.7% 0.1% 0.2% 0.0% 0.2% 0.1% 0.1% 0.0% 0.0%

T 3.4% 1.0% 0.1% 3.9% 0.1% 0.0% 3.2% 4.6% 3.2% 0.6% 0.3% 0.5% 0.1% 1.1%

Tx 0.1% 0.0% 0.0% 0.0% 0.0% 4.0% 0.2% 0.2% 0.1% 0.0% 0.0% 0.1% 0.0% 0.0%

P 0.0% 0.0% 0.0% 0.0% 0.1% 0.2% 0.0% 1.3% 5.5% 0.0% 0.1% 0.4% 0.1% 0.1%

D 0.3% 0.4% 0.0% 0.5% 0.0% 5.3% 0.2% 0.0% 0.1% 0.6% 0.2% 4.5% 1.8% 0.4%

sx 0.0% 0.3% 0.2% 0.4% 0.2% 0.2% 0.0% 0.2% 0.4% 2.9% 0.3% 0.3% 0.1% 0.4%

Rx 0.0% 1.6% 0.0% 0.5% 0.0% 0.1% 0.0% 0.0% 0.1% 0.1% 0.9% 0.0% 0.1% 0.3%

D+ 0.0% 0.1% 0.0% 0.0% 0.0% 6.3% 0.9% 0.0% 0.1% 0.1% 0.1% 0.0% 0.0% 0.0%

Dx 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 2.5% 0.0% 0.2% 0.3% 1.1% 0.1%

example b7. Transitions between hidden states of Bach Chorale thirteen-function model. Table sums to 100%, background colors are
graded to reflect increasing probabilities.
 chord context and function 335l

functions

chords T+ R Px p d T Tx P D sx Rx D+ Dx

I 0.0% 0.0% 0.0% 0.0% 0.0% 87.2% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.6%

V 1.0% 0.0% 0.0% 0.0% 0.0% 0.0% 2.5% 0.0% 84.8% 0.0% 0.0% 0.0% 0.0%

Downloaded from [Link] by SMT Member Access user on 12 August 2019

IV 0.0% 0.0% 0.0% 72.1% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%

V7 0.0% 0.0% 0.0% 0.0% 1.0% 0.1% 0.0% 0.6% 0.2% 0.0% 0.0% 86.0% 0.0%

vi 0.5% 97.3% 4.3% 1.2% 0.0% 2.8% 2.5% 0.0% 0.0% 0.0% 0.0% 0.0% 1.2%

ii 0.0% 0.0% 5.9% 12.7% 1.0% 0.0% 0.0% 0.3% 0.0% 44.0% 0.0% 0.0% 0.0%

ii7 0.0% 0.0% 0.0% 1.6% 7.9% 0.0% 0.0% 29.3% 0.0% 3.3% 0.0% 0.0% 0.0%

iii 7.2% 2.2% 4.8% 0.0% 0.5% 4.9% 3.0% 0.6% 5.0% 0.4% 4.5% 0.0% 1.2%

vii 3.6% 0.0% 0.0% 0.0% 39.9% 0.0% 6.0% 0.0% 0.2% 0.0% 0.0% 1.9% 0.0%

IV7 0.0% 0.0% 23.7% 2.6% 0.0% 0.0% 0.0% 8.1% 0.0% 0.0% 0.0% 0.0% 0.0%

^ ^ ^
{ 1 , 2 , 5} 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 4.0% 17.1% 1.5% 0.0% 0.0% 0.0% 2.3%

vi 7 0.0% 0.0% 27.4% 3.3% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 4.1%

vii 7 0.0% 0.0% 0.0% 0.0% 25.6% 0.0% 0.0% 0.0% 0.7% 0.0% 7.1% 0.0% 0.6%

I7 28.4% 0.0% 0.0% 0.0% 0.0% 0.0% 1.5% 0.0% 0.0% 0.0% 0.0% 0.0% 0.6%

^ ^ ^^
{1, 2, 4, 5} 0.0% 0.0% 0.5% 0.0% 0.0% 0.0% 3.0% 15.0% 0.0% 0.0% 0.0% 0.3% 0.0%

example b8. Chords in Bach Chorales associated with the thirteen hidden states (lexical probabilities). Columns sum to 100%. Only
chords that occur nineteen times or more within the corpus are shown (>.05% of the corpus’s chords).
335m MUSIC THEORY SPECTRUM 40 (2018)

I9 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 23.5% 0.0% 1.5% 0.0% 0.0% 0.0% 0.0%

I add 4 2.6% 0.0% 0.0% 0.0% 0.0% 0.0% 20.0% 0.0% 1.9% 0.0% 0.0% 0.0% 0.0%

iii 7 19.1% 0.0% 0.5% 0.0% 0.0% 1.3% 1.5% 0.0% 0.0% 0.0% 0.0% 0.0% 0.6%

^ ^ ^
{1, 4, 5} 0.0% 0.0% 0.0% 0.0% 2.5% 0.0% 5.5% 7.8% 0.0% 0.0% 0.0% 1.3% 0.0%

Downloaded from [Link] by SMT Member Access user on 12 August 2019

II7 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 3.7% 0.0% 0.0% 0.0% 0.0% 15.2%

IV9 6.2% 0.0% 9.1% 0.9% 0.5% 0.0% 0.0% 0.9% 0.0% 0.0% 0.0% 0.0% 0.0%

V/vi 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 22.7% 0.0% 0.0%

^ ^ ^
{ 1, 2, 4} 0.0% 0.0% 0.0% 2.8% 0.0% 0.0% 6.0% 2.2% 0.0% 0.0% 0.0% 0.0% 0.0%

vii/vi 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 1.6% 0.0% 0.0% 0.0% 0.0% 14.0%

V7 no 5th 0.0% 0.0% 0.0% 0.0% 2.5% 0.0% 3.0% 0.0% 0.2% 0.0% 0.0% 4.8% 0.0%

V7/IV 13.4% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%

V7/vi 2.1% 0.0% 3.8% 0.2% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 7.0%

^ ^ ^
{ 1, 5, 6} 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 15.6% 0.0% 0.0%

V add 4 5.7% 0.0% 0.0% 0.0% 0.0% 0.0% 2.5% 0.0% 0.0% 0.0% 0.0% 0.0% 3.5%

^ ^ ^
{ 1 , 2 , 6} 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 6.2% 0.0% 0.0% 0.6% 0.0% 0.6%

^ ^^
{ 2 , 4, 5} 0.0% 0.0% 0.5% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 11.7%

^ ^^ ^
{1, 4, 6, 7} 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 4.5% 0.3% 0.0% 0.8% 0.0% 2.9% 0.0%

V/iii
0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 8.7% 0.0% 0.0% 0.0%

^ ^ ^
{2, 5, 6}
1.0% 0.0% 0.0% 0.0% 9.4% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%

v
1.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 1.5% 3.7% 0.0% 0.0% 0.0%

example b8. [Continued]

 chord context and function 335n

V7/ii 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.4% 0.0% 0.0% 10.5%

II 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 7.9% 0.0% 0.0% 0.0%

^ ^ ^ ^ 0.0% 0.0% 2.2% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 5.4% 0.0% 0.0% 0.0%
{2, 3, 4, 6}

vii/ii 0.0% 0.0% 0.0% 0.0% 0.0% 0.6% 0.0% 0.0% 0.0% 4.1% 0.0% 0.0% 0.0%

Downloaded from [Link] by SMT Member Access user on 12 August 2019

example b8. [Continued]

works cited Everett, Walter. 2004. “Making Sense of Rock’s Tonal

Systems.” Music Theory Online 10.4.
Aarden, Bret J. 2003. “Dynamic Melodic Expectancy.” Ph.D. F!etis, François-Joseph. 1844. Trait!e complet de la th!eorie et de
diss., Ohio State University. la pratique de l’harmonie. Paris: Schlesinger.
Agmon, Eytan. 1995. “Functional Harmony Revisited: A Forte, Allen. 1962. Tonal Harmony in Concept and Practice.
Prototype-Theoretic Approach.” Music Theory Spectrum 17 New Haven: Yale University Press.
(2): 196–214. Harrison, Daniel. 1994. Harmonic Function in Chromatic
———. 2013. The Languages of Western Tonality. Heidelberg: Music: A Renewed Dualist Theory and an Account of Its
Springer Verlag. Precedents. Chicago: University of Chicago Press.
Bello, Juan P., and Jeremy Pickens. 2005. “A Robust Mid- Hauptmann, Moritz. 1853 [1888]. Die Natur der Harmonik
level Representation for Harmonic Content in Music und der Metrik. Leipzig, Breitkopf und H€artel. Trans. W.
Signals.” In Proceedings of the International Society for Music Heathcote as The Nature of Harmony and Metre. London:
Information Retrieval. 304–11. London. S. Sonnenschein.
Biamonte, Nicole. 2010. “Triadic Modal and Pentatonic Patterns Huron, David. 2006. Sweet Anticipation: Music and the
in Rock Music.” Music Theory Spectrum 32 (2): 95–110. Psychology of Expectation. Cambridge, MA: MIT Press.
Burgoyne, J. Ashley. 2012. “Stochastic Processes and Jacoby, Nori, Naftali Tishby, and Dmitri Tymoczko. 2015.
Database-Driven Musicology.” Ph.D. diss., McGill “An Information Theoretic Approach to Chord
University. Categorization and Functional Harmony.” Journal of New
Burgoyne, J. Ashley, Jonathan, Wild, and Ichiro, Fujinaga. Music Research 44 (3): 219–44.
2013. “Compositional Data Analysis of Harmonic Jurafsky, Daniel, and James H. Martin. 2008. Speech and
Structures in Popular Music.” In Proceedings of the 4th Language Processing: An Introduction to Natural Language
International Conference on Mathematics and Computation in Processing, Computational Linguistics, and Speech
Music. 52–63. Heidelberg: Springer. Recognition, 2nd edition. Upper Saddle River, NJ: Prentice
Creel, Sarah C., Elissa L. Newport, and Richard N. Aslin. Hall.
2004. “Distant Melodies: Statistical Learning of Klumpenhouwer, Henry. 2007. “Dualist Tonal Space and
Nonadjacent Dependencies in Tone Sequences.” Journal of Transformation in Nineteenth-Century Musical Thought.”
Experimental Psychology: Learning, Memory, and Cognition In The Cambridge History of Western Music Theory, 5th edi-
30: 1119–30. tion. Ed. Thomas Christensen. 456–76. Cambridge:
Cuthbert, Michael, and Christopher Ariza. 2011. “Music21: Cambridge University Press.
A Toolkit for Computer–Aided Musicology and Symbolic Kopp, David. 2002. Chromatic Transformations in
Music Data.” In Proceedings of the International Symposium Nineteenth-Century Music. Cambridge: Cambridge
on Music Information Retrieval. 637–42. University Press.
Dahlhaus, Carl. 1968/1990. Studies on the Origin of Harmonic Kostka, Stefan, and Dorothy Payne. 2012. Tonal Harmony
Tonality. Trans. Robert O. Gjerdingen. Princeton: with an Introduction to Twentieth-Century Music, 4th edi-
Princeton University Press. tion. New York: McGraw-Hill.
deClercq, Trevor, and David Temperley. 2011. “A Corpus Krumhansl, Carol L. 1988. “Tonal and Harmonic
Analysis of Rock Harmony.” Popular Music 30 (1): 47–70. Hierarchies.” In Harmony and Tonality. Ed. J.
Doll, Christopher. 2007. “Listening to Rock Harmony.” Sundberg. Stockholm: Royal Swedish Academy of
Ph.D. diss., Columbia University. Music.
335o MUSIC THEORY SPECTRUM 40 (2018)

———. 1990. The Cognitive Foundations of Musical Pitch. Riemann, Hugo. 1893. Vereinfachte Harmonielehre, oder die
Oxford: Oxford University Press. Lehre von den tonalen Funktionen der Akkorde. London:
Lerdahl, Fred. 2001. Tonal Pitch Space. Oxford: Oxford Augener.
University Press. ———. 1900–1901. Pr€aludien und Studien III. Leipzig:
Lester, Joel. 1992. Compositional Theory in the Eighteenth Hermann Seemann.
Century. Cambridge, MA: Harvard University Press. Riepel, Joseph. 1755. Grundregeln zur Tonordnung insgemein:
Loui, Psyche, David Wessel, and Carla L. Hudson Kam. Abermal Durchgehends mit musicalischen Exempeln abgefaßt
2010. “Humans Rapidly Learn Grammatical Structure in a und Gespr€ach-weise vorgetragen. Frankfurt: Krippner.
New Musical Scale.” Music Perception 27: 377–88. Saffran, Jenny R., Elizabeth K. Johnson, Richard N. Aslin,

Downloaded from [Link] by SMT Member Access user on 12 August 2019

Loui, Psyche. 2012. “Learning and Liking of Melody and and Elissa L. Newport. 1999. “Statistical Learning of Tone
Harmony: Further Studies in Artificial Grammar Sequences by Human Infants and Adults.” Cognition 70:
Learning.” Topics in Cognitive Science 4: 1–14. 27–52.
Mavromatis, Panayotis. 2005. “A Hidden Markov Model of Saffran, Jenny R., Karelyn Reeck, Aimee Niebuhr, and Diana
Melody Production in Greek Church Chant.” Computing Wilson. 2005. “Changing the Tune: The Structure of the
in Musicology 14: 93–112. Input Affects Infants’ Use of Absolute and Relative Pitch.”
———. 2009. “HMM Analysis of Musical Structure: Developmental Science 8 (1): 1–7.
Identification of Hidden Variables through Topology- Smith, Charles. 1981. “Prolongations and Progressions as
Sensitive Model Selection.” In Communications in Computer Musical Syntax.” In Music Theory: Special Topics. Ed.
and Information Science: Mathematics and Computation in Richmond Browne. 139–74. New York: Academic Press.
Music. Eds. Ellen Chew, Adrian Childs, and Ching-Hua Stein, Deborah. 1983. “The Expansion of the Subdominant in
Chuan. 205–217. Heidelberg: Springer Verlag. the Late Nineteenth Century.” Journal of Music Theory 27
———. 2012. “Exploring the Rhythm of the Palestrina Style: (2): 153–80.
A Case Study in Probabilistic Grammar Induction.” Journal Stephenson, Kenneth. 2002. What to Listen for in Rock: A
of Music Theory 56: 169–223. Stylistic Analysis. New Haven, CT: Yale University Press.
McHose, Allen Irvine. 1947. The Contrapuntal Harmonic Swinden, Kevin J. 2005. “When Functions Collide: Aspects of
Technique of the 18th Century. New York: F. S. Crofts Plural Function in Chromatic Music.” Music Theory
and Co. Spectrum 27 (2): 249–82.
Momigny, J!erôme-Joseph. 1806. Cours complet d’harmonie et de Temperley, David. 2009. “A Statistical Analysis of Tonal
composition d’après une th!eorie neuve et g!en!erale de la musique. Harmony.” Accessed September 2, 2018. [Link]
Paris: Bailleul. [Link]/kp-stats/.
Moore, Allan.1992. “Patterns of Harmony.” Popular Music 11 Tymoczko, Dmitri. 2003. “Function Theories: A Statistical
(1): 73–106. Approach.” Musurgia 10 (3–4): 35–64.
Nobile, Drew. 2014. “A Structural Approach to the White, Christopher W. 2013. “An Alphabet-Reduction
Analysis of Rock Music.” Ph.D. diss., CUNY Graduate Algorithm for Chordal n-grams.” In Proceedings of the 4th
Center. International Conference on Mathematics and Computation in
Quinn, Ian. 2010. “What’s ‘Key for Key’: A Theoretically Music. Eds. Jason Yust, Jonathan Wild, and John Ashley
Naive Key–Finding Model for Bach Chorales.” Zeitschrift Burgoyne. 201–12. Heidelberg: Springer.
der Gesellschaft für Musiktheorie7(ii): 151–63. White, Christopher W., and Ian Quinn. 2016. “The Yale-
———. 2017. “Tonal Harmony.” In The Oxford Handbook of Classical Archives Corpus.” Empirical Musicology Review
Concepts in Music Theory. Ed. Alexander Rehding and 11 (1): 50–58.
Steven Rings. New York: Oxford University Press. Wittig, Susan. 1977. “The Computer and the Concept of
Quinn, Ian, and Panayotis Mavromatis. 2011. “Voice Leading Text.” Computers and the Humanities. 11 (4): 211–15.
and Harmonic Function in Two Chorale Corpora.” In
Mathematics and Computation in Music. Ed. Carlos Agon.
230–40. Heidelberg: Springer.
Rameau, Jean-Philippe. 1737. G!en!eration harmonique, ou
Trait!e de musique th!eorique et pratique. Paris: Jean-Baptiste-
Christophe Ballard.
Raphael, Christopher, and Joshua Stoddard. 2004. “Functional
Analysis Using Probabilistic Models.” Computer Music
Journal 28 (3): 45–52.
Music Theory Spectrum, Vol. 40, Issue 2, pp. 314–37, ISSN 0195-6167,
Rehding, Alexander. 2011. “Tonality between Rule and electronic ISSN 1533-8339. V C The Author(s) 2018. Published by Oxford
Repertory; Or, Riemann’s Functions—Beethoven’s University Press on behalf of The Society for Music Theory. All rights
Function.” Music Theory Spectrum 33 (2): 109–23. reserved. For permissions, please e-mail: [Link]@[Link].
DOI: 10.1093/mts/mty021