Computational and Comparative Musicology -...
Transcript of Computational and Comparative Musicology -...
CH
AP
TE
R 6
Com
putational and Com
parative Musicology
Nicholas C
ook
Introd
uction
The m
iddle of the twentieth century saw
a strong reaction against the comparative
methods that played so large a part in the disciplines of the hum
anities and socialsciences in the first half of the century, and m
usicology was no exception. T
he term“com
parative musicology” w
as supplanted by “ethnomusicology,” reflecting a new
belief that cultural practices could only be understood in relation to the particularsocieties that gave rise to them
: it was sim
ply misleading to com
pare practices acrossdifferent societies, the ethnom
usicologists believed, and so the comparative m
usi-cologist w
as replaced by specialists in particular musical cultures. A
similar reaction
took place in theory and analysis: earlier style-analytical approaches (largely mod-
eled on turn-of-the-century art history) gave way to a new
emphasis on the particu-
lar structural patterns of individual musical w
orks. Perversely, this meant that the
possibility of computational approaches to the study of m
usic arose just as the ideaof com
paring large bodies of musical data—
the kind of work to w
hich computers
are ideally suited—
became intellectually unfashionable. A
s a result, computational
methods have up to now
played a more or less m
arginal role in the development of
the discipline.In this chapter I suggest that recent developm
ents in computational m
usicol-ogy present a significant opportunity for disciplinary renew
al: in the terms intro-
duced in chapter 1, there is potential for musicology to be pursued as a m
ore data-rich discipline than has generally been the case up to now
, and this in turn entailsa re-evaluation of the com
parative method. C
entral to any computational approach,
however, are the m
eans by which data are represented for analysis, and so I begin
with som
e examples of “objective” data representations before introducing the issue
of comparison. (T
he examples I discuss are graphic, but the sam
e points could havebeen m
ade in terms of num
erical representations.) This is follow
ed by an extendedcase study of an im
portant current software package for m
usicological research, theH
umdrum
Toolkit, and the chapter concludes with a brief consideration of the
prospects for computational m
ethods in musicology.
103short
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
03
Figure 6.1. Graphic analyses of Ligeti, Lux aeterna, section 2 (from Clendenning 1995: 948–49).
105
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
05
Issues of R
epresen
tation
A picture, as everyone know
s, is worth a thousand w
ords. Each of the graphs in Fig-
ure 6.1 (from C
lendenning 1995: 248–
249) represents a different aspect of the same
section from Ligeti’s Lux aeterna: (a) charts pitch against tim
e, giving an imm
ediateim
pression of the music’s registral profile (b) show
s how m
any different pitches (notpitch classes) are present at any given point, and (c) highlights voice entries, w
itheach separate voice (four each of soprano, alto, tenor, bass) having a separate hori-zontal line w
hich is shaded whenever the voice enters, w
hile (d) shows essentially
the same inform
ation as (c), only in terms of the total num
ber of entries at any givenpoint. T
hese graphs are “objective” in the sense that, once you have agreed how a
graph is to be laid out, everyone should end up with the sam
e result; the information
is all there in the score, so that the analysis becomes, in effect, a m
atter of reformat-
ting. But the decisions about how to lay them
out are not necessarily self-evident.For exam
ple, in (a) to (c) the minim
um values on the tim
e axis are quavers, whereas
in (d) they are whole bars; other values w
ould have been possible, and the decisionto use these particular values represents a judgm
ent by the analyst that they will be
the most inform
ative in this context. There is also a degree of analytical judgm
ent inthe separate identification of the canons in (a) and (b).
That said, how
ever, it is obvious that there is scope for automation in graphs
like this—and not just in draw
ing them up (the graphs in Figure 6.1 w
ere gener-ated by a com
puter drawing package) but in the processing of the data. T
he graphsin Figure 6.2 (from
Brinkman and M
esiti 1991: 5, 6, 13, 17) were autom
atically gen-erated from
a machine representation of the score. 1
Again, each graph represents a
different way of extracting and view
ing the information in a score, this tim
e bars 1to 26 of the first m
ovement from
Bartók’s String Quartet no. 4: (a) corresponds to
Figure 6.1(a)(b)show
s each instrumental part separately w
ith the vertical lines pro-viding an im
pression of the linear movem
ent from one pitch to another, and
(c)charts the first appearance of each pitch, w
hile (d) represents the dynamic level
of each instrument (based on w
hether it is playing and on notated dynamic value)
and of the whole (in effect a sum
mation of the dynam
ic graphs of each instrument).
But what are w
e to make of such representations? W
hat do they enable us to see thatw
e can’t hear, and in any case, what is the point of seeing
music at all? It is a com
-m
onplace to describe Ligeti’s texture-oriented works as “visual,” and Brinkm
an andM
esiti (who in their article offer a case study of W
ebern’s Symphonie
Op. 21) em
-phasize the spatial sym
metry of W
ebern’s serial structures, comm
enting that “thegraphs allow
us to view all parts together in their spatial environm
ent. [This] is es-
pecially helpful since the pitch symm
etry that is essential to the musical structure of
this work is som
ewhat obscured by the notated score” (Brinkm
an and Mesiti 1991:
21). The question then is the extent to w
hich this problem of obscuring applies to
other repertories, in other words how
useful this kind of visualization is for music
in general.Yet this question is in som
e ways m
isguided: there is no such thing as a good orbad representation of m
usic per se, there are good and bad—
or at least better andw
orse—representations for particular purposes. To say that visual representations are
more appropriate for tw
entieth-century than for other repertories, or for answering
EM
PI
RI
CA
L M
US
IC
OL
OG
Y106
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
06
certain kinds of questions about the music than others, is sim
ply to define their use-fulness, not to criticize them
. Nevertheless Brinkm
an and Mesiti produced som
e in-teresting results for m
usic that one wouldn’t norm
ally think of as particularly “visual,”including som
e strikingly different representations of music by Bach and M
ozart asw
ell as by Bartók, Schoenberg, Wolpe, and Berio. A
t the very least, such comparisons
serve to underline the variety of textures found in different composers or styles—
and texture is at the same tim
e one of the most im
portant aspects of music in term
sof how
we experience it, and one of the hardest to say anything useful about. It is not
hard to imagine how
Brinkman and M
esiti’s software could be used w
ithin a varietyof pedagogical contexts, enabling students to literally and instantaneously “see” dif-ferent pieces of m
usic in a variety of different ways—
and the ability to switch easily
and quickly from one representation of m
usic to another is a crucial element of prac-
tical musicianship. T
hat this has not happened, and that the potential of the approachhas not been fully realized, reflects the cost of translating academ
ic research projectsinto softw
are products for the real world, and illustrates an all-too-fam
iliar pattern inm
usicological software: a sustained burst of initial enthusiasm
is followed by running
out of money, resulting in softw
are that is sometim
es less than fully functional, oftenless than fully docum
ented, rarely properly supported, and usually soon obsolete.But m
y principal answer to the question “w
hat are we to m
ake of such repre-sentations?” is a different one. Brinkm
an and Mesiti (1991: 7) em
phasize the partic-ular usefulness of their graphs “as a device for com
paring different pieces”; Matt
Hughes, w
hose quantitative analysis of Schubert’s Op. 94 no. 1 w
as discussed inchapter 1, sim
ilarly claimed of his m
ethod that it was “a tool for organizing data so
that one may m
ore discernibly view tendencies and interassociations,” adding that
it was “best utilized w
hen viewing groups of com
positions or sections of a composi-
tion rather than a single work” (H
ughes 1977: 145, 150). And there is a general
point to be made here: the m
ore objective an analytical approach is, the more its m
u-sicological value is likely to be realized through m
aking comparisons betw
een dif-ferent pieces—
and sometim
es large numbers of them
. For example, w
hen you carryout a Schenkerian analysis, you are in a sense m
aking a comparison betw
een thepiece in question and the norm
s of comm
on-practice voice leading, as systematized
in the Schenkerian background and middleground. But the value of the analysis
consists primarily in the lengthy process of m
aking it, deciding which notes go w
ithw
hich, which are m
ore important than others, and so forth; the process is lengthy
because it involves a vast number of interpretive judgm
ents, requiring you to weigh
up different factors in relation to one another. At the end of it, you have a know
ledgeof the m
usic—you m
ight call it an intimacy
—that you did not have at the outset,
and there is a sense in which the final graph is significant m
ainly as a record of thislearning process. W
ith any kind of computational approach, by contrast, all of this
happens automatically, and in som
e cases almost instantaneously; the only output is
the graphic or numerical representation of the m
usic that results. And such repre-
sentations rarely tell you anything very useful by themselves: they m
ay well be vul-
nerable to the cheap but telling jibe that they either show you w
hat you hear any-how
(in which case they are redundant), or else they don’t (in w
hich case there areirrelevant). It is w
hen you compare representations of different pieces that useful in-
formation m
ay emerge.
Com
putational and Com
parative Musicology
107
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
07
Com
putational and Com
parative Musicology
109
Figu
re 6.2.G
raphic analyses of Bartók, String Quartet N
o. 4, I, bars 1–
26 (fromBrinkm
an and Mesiti 1991: 5, 6, 13, 17).
(d)
The value of objective representations of m
usic, in short, lies principally in thepossibility of com
paring them and so identifying significant features, and of using
computational techniques to carry out such com
parisons speedily and accurately. At
this point, however, the issue of reduction (already broached in chapter 1, this vol-
ume) has to be confronted. T
his can conveniently be explored in relation to theanalysis of folksongs, since a large of proportion of early w
ork in computational
analysis was based on such m
aterial: there were substantial existing collections of
folksongs to work on, and the m
usic was m
onophonic and scalar, hence could becoded very com
pactly—
a vital consideration in the days when com
puter mem
oryw
as a scarce and expensive comm
odity. But of course, if you code folksongs as aseries of pitches, or scale degrees, then you are leaving out all the variables of into-nation, ornam
entation, timbre, and other aspects of singing style, not to m
ention the
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
09
original contexts of use and cultural connotations of the songs. And that obviously
limits w
hat you can conclude from com
paring them.
At this point a concrete exam
ple may be helpful, and Figure 6.3 is taken from
the Essen M
usical Data Package, a series of databases of popular and traditional
vocal repertories together with tools for encoding and analysis. 2
Coordinated until
his death in 1994 by Helm
ut Schaffrath, the Essen databases consisted by the fol-
lowing year of som
e 10,000 songs, mainly E
uropean (and largely Germ
an), but with
some representation of other traditions, particularly C
hinese; they are still beingadded to, though developm
ent has been transferred from E
ssen to Warsaw
(Self-ridge-Field 1995). Figure 6.3 show
s “Deutschland über alles” in the custom
codeused by the package, EsA
C(the E
ssen Associative C
ode), and several of the datafields are self-explanatory: BO
EH
ME
is the name of the database in w
hich this songis located, C
UT
contains the title, RE
G is the region from
which it com
es, and FCT
is the functional category of the song, while K
EY contains not only its key (F) but
also the numbering of the song in the database (B0001), the value of the sm
allestnote value (16, a sixteenth or sem
i-quaver), and the time signature (4
/4). The tune
itself is contained in the field ME
L, using a simple code based on the scale degree
(with # for sharp and b for flat degrees), �
and �to indicate shifts to a higher or
lower register, respectively, and a rhythm
ic notation based on the shortest value(sem
i-quaver), with a dot prolonging the value by 50 percent, a single underline
character (_) by 100 percent, and a double underline character (_ _
) by 200 percent.T
hat should suffice for the tune to be read; the only other information contained in
the ME
L field is //(for the end of the m
elody), the use of spaces to indicate bar lines(the m
elody has been notated across the bar lines), and the splitting up of the mel-
ody into separate lines, each of which represents a phrase (this represents a judg-
ment on the part of the transcriber, and as such is a rather unusual feature of the
EsAC
code).W
hat can be done using this information? T
he distributed version of the Essen
database comes w
ith a simple analytical utility (A
NA
), which takes a com
plete data-
EM
PI
RI
CA
L M
US
IC
OL
OG
Y110
Figu
re 6.3
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
10
base as its input, and generates as its output an annotated version of the file in which
up to 12 different kinds of analytical information are added for each song. T
hese in-clude straightforw
ard statistical information such as the distribution of ascending
and descending intervals (how m
any ascending minor seconds?), the distribution of
durations (how m
any crotchets, quavers, semi-quavers?), and distribution of scale
degrees (how often does the fifth scale degree appear in the low
er register?); thereare also som
e derived measures such as the percentage of ascending steps and leaps.
They also include som
e less obvious information, including the pattern of phrase
repetitions in pitch and duration, a coding of the contour of each phrase, them
elodic spine (based on the notes at metrically stressed beats), and the final notes
of each phrase (this is where the coding of phrases com
es in). 3A
limitation of the
AN
A softw
are is that it provides data for each song separately, whereas the statistical
information w
ould be more useful if it w
ere provided across the database as a whole,
but there is a reason for this: the package was intended for use w
ith a comm
ercialrelational database package, w
hich would facilitate this kind of analysis. N
eedless tosay, the original package is no longer obtainable. 4
But what can be done
using this information? Schaffrath him
self published anum
ber of studies based on his datasets, in which he used this kind of analytical in-
formation to support broad stylistic generalizations, for exam
ple as between differ-
ent traditions: “Germ
an folksongs,” he concluded from one such study (1992: 107),
“generally skip more often up than dow
n; in the Chinese pool the opposite ap-
plies.... Germ
an songs use the interval of a fourth 51 percent more often than C
hi-nese songs do. T
his might be explained by the preference for pentatonic scales.” But
there is also a different way in w
hich such information can be used, w
hich is for pur-poses of searching, identification, and classification. A
database of 10,000 folksongsis not unlike a bibliographic database, and to locate individual item
s or groups of re-lated item
s you search for particular features, combining them
so as to narrow the
search down to a m
anageable number of hits. In som
e contexts all you need to searcha bibliographic database is the author’s nam
e; in other cases you need to combine an
author name, tw
o title keywords, and year to locate the article you need. In the sam
ew
ay, you might try to locate a particular song by searching for its first few
notes orincipit; this is the w
ay that traditional dictionaries of themes operate (for instance,
Barlow and M
orgenstern 1983), and it works w
ell with “art” m
usic, where there are
fixed scores. But it is not such a good way to search for folksongs, w
hich typicallyexist in a large num
ber of variants, often involving variation or the interpolation ofnotes. T
hat is why A
NA
includes a routine to derive the final notes of each phrase: asSchaffrath (1997: 349) explains, “although variants m
ay contain different numbers
of notes or melodic contours, they tend to retain underlying harm
onic structures.”O
r you could combine such a search w
ith a particular melodic spine and
/or patternof phrase repetition. A
nd whereas I have described the use of such features in term
sof sim
ply locating items, the sam
e processes can be used to group together songs orrepertories in w
ays that might reflect, for exam
ple, historical development, linguis-
tic associations, or conditions of use. In such ways the autom
ated analysis of musi-
cal style can become a m
eans of generating or verifying musicological hypotheses.
Or at least that is the aspiration. In practice, EsA
Cand A
NA
suffer from obvious
limitations, ranging from
the pragmatic (the output of A
NA
is only marginally m
ore
Com
putational and Com
parative Musicology
111
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
11
user-friendly than machine code) to those of principle: EsA
Ccan cope w
ith only anarrow
range of music (m
onophonic, restricted to three octaves, and so on), and isw
ell adapted for only a narrow range of applications. A
NA
extracts a limited num
berof features, each of w
hich represents a radical reduction of the music and is intended
for a particular purpose (such as statistical generalization or pattern matching);
moreover, the usefulness of these reductions depends on a num
ber of quite specificassum
ptions, such as that the final-note patterns of phrases vary less than incipits.A
ll this makes sense in term
s of what the E
ssen package is designed for; it is, to adoptoutdated softw
are jargon, a “turnkey” solution to folksong analysis, or at least to acertain sort of folksong analysis. A
nd it would be possible to im
agine trying to de-velop the code, and the analytical softw
are, in such a way as to be m
ore flexible—to cope w
ith music of any desired degree of com
plexity and variety, and to extractfrom
it every feature that could be of any possible interest under any possible cir-cum
stances. That w
ould be like the approach of business software developers in the
1980s, who aim
ed at prepackaged, turnkey solutions in which every eventuality had
been foreseen and the appropriate feature bolted on somew
here, if only you couldfind it.
In musicology, as perhaps also in the office, this represents a basic m
isappre-hension
—in term
s not only of the unlimited variety of m
usical materials, but also
of the variety of purposes for which people w
ill want to represent them
; as Huron
(1992: 11) puts it, “The types of goals to w
hich music representations m
ay serve arelegion and unpredictable.” C
omplexity is not the solution. T
here are, after all, much
more com
plicated codes than EsAC
which fulfill certain purposes extrem
ely well
while being very poorly adapted for others. M
IDI code has transform
ed entire sec-tors of the m
usic industry, but because it tries to understand all music on the m
odelof keyboard m
usic its handling of microtonal inflection is inelegant, to say the least;
it also doesn’t distinguish enharmonic equivalents, and has to be drastically recon-
figured if useful analytical information is to be extracted from
it—even som
ethingas sim
ple as locating a C m
ajor triad. 5A
gain, DA
RM
S6
is an extremely pow
erful codedesigned for purposes of printing and publication (it is the basis of all A
-R E
ditionscores), and accordingly it understands m
usic as a kind of graphics; it doesn’t thinkof an e
1but of a notehead on the bottom
line of a stave that has an F clef on it. As a
result, it too needs drastic reconfiguration if analytically useful information is to be
extracted from it (Brinkm
an 1986). And though the problem
s of multiple codes are
alleviated by the existence of interchange utilities (e.g., to turn DA
RM
S into MID
I),there are lim
itations to what is possible: since M
IDI does not distinguish betw
een C �and D �, the inform
ation is simply not there to translate into D
AR
MS, w
hich doesm
ake the distinction.A
s with business softw
are, the solution to this problem of m
ultiple requirements
is not the creation of ever more com
plex and unwieldy integrated solutions, but a
modular approach involving an unlim
ited number of individual softw
are tools de-signed to serve individual purposes; that w
ay, when you w
ant to do something new
,you sim
ply design a new tool to do it. M
odular approaches like this, however, need
some kind of fram
ework w
ithin which to w
ork—
in essence, a set of rules for therepresentation of data in m
achine-readable files, and for the transfer of information
between different softw
are tools. Hum
drum is an exam
ple of just such a framew
ork.
EM
PI
RI
CA
L M
US
IC
OL
OG
Y112
EQ
MSP 111
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
12
A C
ase Study: T
he H
um
drum
Toolkit
Developed by D
avid Huron in the early 1990s, H
umdrum
7consists of tw
o distinctelem
ents: in the first place a syntax—
the set of rules I referred to—
and in the sec-ond, a “toolkit” consisting at present of som
ething over 70 separate software routines
for manipulating the data in different w
ays for different purposes. It runs in a UN
IXenvironm
ent, 8and its m
odular approach is much m
ore readily understood by thosew
ho have experience of UN
IX than those w
ho do not. Like Hum
drum, U
NIX
can bethought of as a set of rules plus a set of tools: individual U
NIX
comm
ands (which
have names like grep, sort, and cat) do sim
ple things like searching files for a partic-ular string and extracting, deleting, or altering it, or rearranging data w
ithin or be-tw
een files. The pow
er and flexibility of the system com
es not from these individual
tools but from the possibility of “pipelining” them
so that the output of one tool be-com
es the input of the next, forming chains of com
mands of unlim
ited length. The
Hum
drum toolkit is sim
ply a set of additional UN
IX tools designed specifically for
manipulating m
usic-related data. It is this open and modular philosophy, w
hich al-w
ays allows for an additional tool to be added w
hen required, that explains why
—as H
uron (1997: 375) puts it—“H
umdrum
is especially helpful in music research
environments w
here new and unforeseen goals arise.” U
nfortunately, as we shall see,
it also means that the user has to have a detailed understanding of the w
ay the dataare encoded and the w
ays in which the tools m
anipulate it. Hum
drum, in short, has
a steep learning curve.A
ll this may sound very abstract, so Figure 6.4 show
s the first two phrases of
“Deutschland über alles” in kern
code, the normal H
umdrum
format for represent-
ing notated music (only tw
o phrases because when printed out—
though not interm
s of internal storage—kern
is much less com
pact than EsAC
). 9T
his is a directtranslation into kern
of Figure 6.3, so the same inform
ation may be found in the data
fields preceded by !!! (the exclamation m
arks identify them as com
ments); only the
information concerning Schaffrath, the copyright statem
ent, and the !!!AR
L line(w
hich gives latitudinal and longitudinal coordinates) are new. A
s for the other lines,**kern m
eans that what follow
s is in kern(the double asterisk m
eans that this labeldefines the data that follow
s), and the lines preceded by a single asterisk define theinstrum
ent category and instrument (in each case “vox” or voice), m
eter (4/4), key
signature (B �, the “flat” being designated by “�”—
the sign for sharp is “�”), and
key (F). 10T
he tune itself begins on the following line, preceded by a curly bracket
(this indicates the phrases, while the lines beginning w
ith “�” are bar lines). “4.f”
means that the first note is the F above m
iddle C (if it w
ere an octave lower it w
ouldbe F, if an octave higher ff, if an octave higher than thatfff, and so on); the “4” m
eansthat it is a quarter note (crotchet), and the “ .
”, unsurprisingly, that it is dotted. Itshould now
be possible to read the melody quite straightforw
ardly. (The apparently
simplistic reciprocal notation for durations is unexpectedly flexible, coping w
ithm
ost things short of Ferneyhough; triplet quavers, for instance, work out as 6, quin-
tuplet semi-quavers as 20). T
hat explains everything in Figure 6.4, except the “ * ”,w
hich marks the end of the record. It should also explain everything in Figure 6.5,
which show
s the same phrases (w
ith the comm
ents omitted), as set in G
major by
Haydn in the second m
ovement of his String Q
uartet Op. 76 no. 3: instead of the
Com
putational and Com
parative Musicology
113
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
13
single column (or spine) of Figure 6.4, there are now
four, one for each line of them
usic. It is helpful to think of this as being laid out like staff notation, only rotatedby 90 degrees.
Kern
is not as easy to read as EsAC
(which, apparently, w
as designed to be easyto sight read). A
nd it would seem
to be better adapted for linear textures than, saykeyboard m
usic—though in fact there is no lim
itation in this respect, partly becauseyou can put m
ore than one note on a single line within any one spine, and also be-
cause you can collapse spines into one another or create new ones at any point. It is
also more flexible than it looks. In the first place, you don’t have to specify every-
thing in Figures 6.4 or 6.5; there is no need to encode phrases, or bar lines, or du-rations, or pitches (in fact the only
mandatory lines are “**kern” and “*
�”). In the
EM
PI
RI
CA
L M
US
IC
OL
OG
Y114
Figu
re 6.4
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
14
second place, there is a large number of further codes and interpretations built into
kern, allowing you for instance to include sym
bols for articulation or ornamentation,
clefs, stem direction, or the num
ber of staff lines, if any (and whether they should
be colored or dotted). But more than that, **kern spines can be m
ixed with others,
including, for example, such predefined representations as **text (for lyrics) or
**IPA (lyrics, but now
notated using the International Phonetic Alphabet), or new
representations defined by the user: you might, for exam
ple, create spines labeled**bow
ing, to show how
a given pattern is (or might be) bow
ed, or **timing, to en-
code the rhythmic nuances in a particular perform
ance, or **heartbeat, to record lis-teners’ physiological responses to hearing the m
usic. Then again, you m
ight not use**kern at all, but an alternative representation for pitch; predefined representationsinclude **pitch (using the International Standards O
rganization format), **deg
(scale degrees, as in EsAC
), **solfg (French fixed-do solfège), **freq (frequency
Com
putational and Com
parative Musicology
115
Figu
re 6.5
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
15
expressed in cycles per second), or the self-explanatory **midi; any of these repre-
sentations can be automatically translated to any other, provided of course that the
relevant information is there to be translated (recall the problem
with M
IDI and en-
harmonic equivalents). T
here is also a particularly well-developed **fret represen-
tation, which turns pitch notation into tablature notation based on any specified
number of strings and tuning, or of course you m
ight always invent a new
repre-sentation for som
e particular repertory or purpose.A
nd what does all this enable you to do? For a start, you can replicate the func-
tions of AN
A: to extract the m
elodic spine of Figure 6.4, for instance, you locate thedow
nbeat of each bar (search for lines beginning “�” and then skip to the next line),
skip over the duration figures, and extract the pitch symbol(s) to a new
file. Or you
could just as easily search an entire “BOE
HM
E” database for m
elodic motions from
^4 to ^5 and establish how frequently they occur by com
parison with the reverse
motion (you w
ould translate the database into **deg records in order to do this, andsearch for adjacent lines containing 4 and 5, skipping over irrelevant sym
bols likedurations, and finally counting the num
ber found). Or you could search all of Bach’s
chorales (the Riem
enschneider collection is available in kerncode) in order to es-
tablish how often the m
elodic leading-notes are approached from beneath, as against
from above: to do this, you extract the spines containing the m
elodies, turn theminto **deg records, and count the occurrences of “^7” as against “v7.” 11 (T
his par-ticular task is easier than it sounds, because **deg autom
atically codes the directionin w
hich a given scale degree was approached, though not the size of the interval in-
volved). Then again, shifting to a different level of com
plexity, you could analyze therelationship betw
een a particular succession of vowels in song texts and the m
elodicor rhythm
ic contexts within w
hich they occur. You could establish how far particu-
lar harmonic form
ations are correlated with specific m
etrical locations (and analyzethe results by com
poser, period, or geographical origin). Or you could classify dif-
ferent melodic, harm
onic, or structural contexts in Chopin’s com
plete mazurkas,
and use this as the basis for analyzing timing inform
ation extracted from a large
number of recordings.But the best w
ay to see what H
umdrum
is capable of is to examine published
studies that have used it, of which the m
ajority (though by no means all) are by
Huron and his co-w
orkers; I shall briefly describe five. An article on the m
elodic archin folksongs (H
uron 1996) is based on the Essen collection and provides a direct
comparison w
ith Schaffrath’s own w
ork. Its purpose is to test systematically the fre-
quent, but inadequately supported, claim that “arch” contours are prevalent through-
out Western folksongs, and H
uron points out the dangers of “unintentional bias”(1996: 3) w
hen such generalizations are supported by a small num
ber of possiblyhandpicked exam
ples: they can be firmly established only through the analysis of
data sets large enough to allow the draw
ing of statistically significant conclusions.H
e puts forward tw
o basic ways in w
hich folksongs might be said to exhibit arch-
shaped contours—w
ithin each phrase, and overall—and system
atically analyzesthe E
ssen collection for each. For the first test, individual phrases throughout thecollection w
ere sorted according to the number of notes in them
, and the averagecontour for each phrase-length com
puted; for the second, the average pitch of eachphrase of each song w
as computed, and the resultant contour classified. T
he overall
EM
PI
RI
CA
L M
US
IC
OL
OG
Y116
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
16
conclusion was that there is a strong tendency tow
ard arch-shaped contours on bothm
easures (though a third approach, based on the nesting of arches within one an-
other, was not supported).
A project like this m
ight be described as regulative: it takes an existing musico-
logical claim and tests it rigorously against a large body of relevant data, adopting cri-
teria of significance and certainty that may be unfam
iliar in musicology but are nor-
mal in data-rich disciplines. Tw
o further studies illustrate the range of musicological
claims that can be tested in this m
anner. A study carried out in collaboration w
ithPaul von H
ippel (Hippel and H
uron 2000) focused on the “gap-fill” model of m
elody,according to w
hich listeners expect melodic leaps to be follow
ed by a change of di-rection; this is an im
portant element of N
armour’s “im
plication-realization” theory,w
hich seeks to explain how listeners experience m
usic on the basis of whether or not
implications set up by the m
usic are realized in what follow
s. What is at issue is not
whether or not m
elodic leaps are usually followed by a change of direction, w
hich isundoubtedly the case; it is w
hether or not this happens (as Krum
hansl has concludedon the basis of experim
ental studies) 12because of listener expectations. T
he alterna-tive is that it is a trivial consequence of the lim
ited registral range of vocal music, a
possibility which von H
ippel and Huron (2000: 63) explain by com
parison with
a simplified m
elody that is confined to a range of three adjacent pitches—for exam
ple, A, B, and C
. In such a melody, the only skip available is betw
eenA
and C. U
pward, this skip m
ust land at the top of the range; downw
ard, itm
ust land at the bottom. A
fter a skip, therefore, there is no way for a m
elodyto continue m
oving in the same direction. O
n the contrary, two of the three
available pitches can be reached only by a reversal. Although m
ost melodies
have a wider range, they are subject to the sam
e basic argument.
If this second possibility is the case, then gap-fill patterns will be no m
ore comm
onin real folksongs than in random
ly generated melodies w
ith otherwise com
parablecharacteristics—
and, to cut a long story short, this is exactly what von H
ippel andH
uron found. If their argument (w
hich I have grossly simplified) is accepted, then
it provides a striking instance of how studies based on large data sets can raise ques-
tions about conclusions derived even from experim
ents as meticulously controlled
as Krum
hansl’s. 13
A third study (H
uron 2001) makes a sim
ilar kind of point in relation to “art”m
usic. It reassesses Allen Forte’s (1983) set-theoretical analysis of the first m
ove-m
ent of Brahms’s String Q
uartet Op. 51, no. 1
—an application to high R
omantic
music of an approach originally developed for tw
entieth-century music, bringing
with it an appearance of objectivity and rigor in stark contrast to conventional m
o-tivic descriptions of such m
usic. (In other words, Forte’s article is a late expression
of the postwar culture of objectivity described in chapter 1, this volum
e.) Whereas
a “motive” is sim
ply a characteristic melodic figure that recurs prom
inently through-out a piece, Forte’s “alpha” ic (interval class) pattern can be defined m
uch more pre-
cisely: it is a particular intervallic configuration that may occur in any of the stan-
dard transforms of set theory (prim
e, inversion, retrograde, retrograde inversion), inany rhythm
ic configuration, and at any metrical point. So H
uron proceeds to searchsystem
atically for Forte’s “alpha” pattern, not only in the first movem
ent of Op. 51,
Com
putational and Com
parative Musicology
117
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
17
No.1 but also in the first m
ovements of Brahm
s’s other quartets—a com
parablerepertory in w
hich the “alpha” pattern, if it is really characteristic of Op. 51 no. 1,
should be significantly less comm
on. First, he demonstrates that, as defined by
Forte, the “alpha” pattern is comm
on in Op. 51 no. 1, but no m
ore so than in theother quartets. H
owever, if you restrict the search to instances of the pattern that
begin on metrically strong beats and com
e at the beginning of slurs, then it is more
prevalent in Op. 51 no. 1 than in the others. T
his is particularly the case if you lookfor the pattern only in its prim
e form, and even m
ore so if you consider only thosecases w
here the pattern coincides with a long-short-long rhythm
ic pattern. So theprevalence of Forte’s “alpha” pattern is confirm
ed—
but only under precisely thoseconditions that are captured by the conventional idea of a m
otive! The conclusion
must be that the appearance of rigor in Forte’s analysis as against traditional analyti-
cal descriptions is just that, an appearance.Im
portant and indeed salutary as this regulative function may be, it w
ould bew
rong to give the impression that com
putational analysis is good only for testing thevalidity of existing theoretical or analytical claim
s. It can also be used to undertakew
ork that could hardly be achieved in any other way. A
spectacular example is the
correlation of musical features w
ith geography, a project once again based on theE
ssen database (Aarden and H
uron 2001). Recall that the E
ssen records include in-form
ation concerning the geographical origins of the songs contained in them—
sometim
es down to the level of an individual tow
n or village, and that the kernver-
sion of “Deutschland über alles” added longitudinal and latitudinal coordinates in
the !!!AR
L field (a new field defined specifically for A
arden and Huron’s project).
Standard Hum
drum com
mands can be used to extract records from
the databaseand output their coordinates; these are then used as input to a m
apping program. By
way of exam
ple, Figure 6.6 shows the distribution of m
ajor- and minor-m
ode songsw
ithin the Essen database; these m
aps were generated using the G
EO
-Music site
(http://ww
w.m
usic-cog.ohio-state.edu/cgi-bin
/Mapping/m
ap.pl), established as partof the project. It is hard to know
what conclusions m
ight be drawn from
this com-
parison (other than that the major m
ode is considerably more w
idespread than them
inor); the southward skew
ing of the major-m
ode data as against the minor sim
plyreflects the three occurrences in Italy
—hardly a basis for robust generalization
—w
hile the concentration in Germ
any and central Europe evident in both cases re-
flects the bias of the Essen database. A
fully fledged musicological research project
would no doubt involve dealing w
ith more features at a greater level of detail. N
ever-theless this exam
ple does indicate the potential for computational m
ethods to drawtogether quite diverse kinds of inform
ation; it also illustrates the value of graphicpresentation of the com
plex data generated by work in com
parative musicology.
A final study (H
uron and Berec, forthcoming) is perhaps even m
ore suggestivein term
s of possible musicological applications. Like H
ofstetter’s attempt (m
entionedin chapter 1) to turn a claim
about the “spirit of nationalism” into an em
piricallytestable proposition, H
uron’s and Berec’s starting point is a woolly, com
mon-language
concept: idiomaticism
. How
might you set about defining w
hat is idiomatic on, say,
the trumpet in such a w
ay that a computer could m
ake evaluations of just how idio-
matic a particular piece of m
usic is? In the first place, Huron and Berec say, idiom
at-icism
is not the same as difficulty: a piece of trum
pet music can be difficult but idio-
EM
PI
RI
CA
L M
US
IC
OL
OG
Y118
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
18
matic, or easy and unidiom
atic. It is rather a matter of achieving the desired m
usicaleffect in the easiest possible m
anner. For example, a piece w
ill be idiomatic if it is sig-
nificantly easier to play as written than w
hen transposed up or down by say a sem
i-tone; this is “transposition idiom
aticism.” So H
uron and Berec evaluated the diffi-culty of a num
ber of pieces when played in all transpositions from
an octave belowthe original to an octave above, and in several cases the original proved to be signif-icantly easier to play than m
ost or all of the transposed versions. (They also ran a
parallel test based on how m
uch more difficult the m
usic was w
hen played at a dif-ferent tem
po from the notated one, w
ith similar results.) But there w
as also a furtherfinding: the pieces had been selected so as to include exam
ples by both trumpet vir-
tuosi and nontrumpeters, and the conclusion in respect of both transposition and
tempo idiom
aticism w
as the same—
that expert trumpeters com
pose more idiom
at-ically than com
posers who cannot play the instrum
ent.T
hat conclusion might be considered predictable, not to say obvious (just as in
the case of Hofstetter’s discovery that there are differences betw
een national styles).A
nd it would have been of lim
ited interest if Huron and Berec had sim
ply askedtrum
peters to play the pieces at different transpositions and at different tempos, and
to say how difficult they w
ere, an approach that would not have required H
umdrum
,of course. But instead
—and this is the real point of their study
—they attem
ptedsom
ething much m
ore challenging: what they refer to as “the design and im
ple-m
entation of a computer m
odel of a trumpet perform
er.” Their m
odel takes a kernfile as its input, and evaluates the difficulty of perform
ance along a variety of differ-ent dim
ensions: in terms of the valve transitions betw
een successive notes; in terms
of register, particularly in the case of sustained notes; in terms of dynam
ics; and interm
s of tonguing at different speeds. (The inform
ation on which they based the
model included perform
er assessments in the case of valve transitions, tests of per-
formance in the case of tonguing, and physiological data concerning lung capacity
and diaphragm support in the case of sustained notes.) T
he model w
as tested usinga series of trum
pet studies set for different grades by the Royal C
onservatory ofM
usic of Toronto; there was a generally fairly high correlation betw
een its estimates
of their difficulty and the Royal C
onservatory’s. But of course the main test w
as theevaluation of idiom
aticism, and the fact that the m
odel come up w
ith what every-
body knows confirm
s the extent to which it actually w
orks. That is the real conclu-
sion to be drawn from
the study.D
o musicologists really need
a computer m
odel of a trumpet perform
er? One
answer to this is that the approach could be readily generalized to other instrum
ents:to the recorder (w
here finger transitions would be of particular im
portance, owing
to the complexity of cross-fingerings), to the piano (the m
odel might propose the
best fingering, which could be tested against those specified by editors or used by
performers), or to the guitar (in effect a m
odeling of the dance of the fingers on thefretboard). But the m
ore substantial answer lies behind this. O
ne of the chronicproblem
s with analyzing m
usic is the tendency toward abstraction: the m
ore so-phisticated our analytical m
odels, the more divorced the object of analysis seem
s tobecom
e from the experience of m
usic, and especially from the sense of physical en-
gagement in w
hich much if not all m
usic has its source. It is conventional to deplorethis, but to do nothing about it. In this context a com
puter model of the electric gui-
EM
PI
RI
CA
L M
US
IC
OL
OG
Y120
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
20
tarist’s left hand could become, perhaps paradoxically, a m
eans of thinking the bodyback into m
usical analysis: you could, to give just one example, chart the interaction
of the dancing fingers with the ear—
of the imperatives of the fretboard w
ith thoseof “purely m
usical” implication and realization
—in rock guitar im
provisation. Inthis w
ay computational m
usicology holds out the promise not only of providing
more robust answ
ers to questions that have already been asked, but also of making
it possible to ask new questions.
Con
clusion
: Brave N
ew W
orld?
Promises, prom
ises... and it m
ust be admitted that com
putational software has a
hardly better record of delivering on its promises than dot com
companies. Let’s stick
with H
umdrum
as the most likely current candidate for a pow
erful, general-purposecom
putational aid for musicology, and explore som
e of the practicalities of doingm
usicology with it.
Q. Is it easy to get hold of?
A. Yes, at the tim
e of writing you can dow
nload it for free, or you can buy itfor the cost of the m
aterials from the C
enter for Com
puter Assisted
Research in the H
umanities at Stanford. 14
Q. W
ill it actually run on my com
puter? Don’t I have to be using a U
NIX
machine?
A.You can run it on a PC
or a Mac, but you w
ill need to install a UN
IXtoolkit. You can probably get that for free, too. 15
Q.Butw
illI be able to find the music I w
ant to work on in kern
code?A
. That depends w
hat you want to w
ork on. The E
ssen package is available in kern, and so is som
e of the Densm
ore collection of Native A
merican
music, so if you are interested in folk m
usic you can start work right
away. A
s for Western “art” m
usic, there is quite a lot out there: not onlyBach’s chorales but also m
ost of his cantatas, as well as the W
ell-Tempered
Clavier; substantial am
ounts of Corelli, V
ivaldi, Handel, and Telem
ann;cham
ber and orchestral music by H
aydn and Mozart; and m
ost of theBeethoven sym
phonies. 16T
he hope, of course, is that as more m
usicolo-gists use H
umdrum
, so more m
usic will be encoded, encouraging m
orem
usicologists to use Hum
drum—
and in this way you get a virtuous
circle. (The sam
e applies to the development of the tools them
selves.)Q
.And if I can’t find w
hat I want?
A. If it exists in som
e other code you may be able to translate it into kern.
Translation routines exist for the MuseD
atacode used by the C
enter forC
omputer A
ssisted Research in the H
umanities (C
CA
RH
), the Plaine andE
asie Code used for R
ISM (R
épertoire international des sources musicales),
and MU
STR
AN
as well as the code used by Leland Sm
ith’s music notation
program SC
OR
E, and the “E
nigma” code used by Finale. N
ot all of thesetranslators are readily available or unproblem
atic, but if you still can’tfind w
hat you want you can of course encode your ow
n. Rather than
doing it by typing in the code, you can use an encoding routine that
Com
putational and Com
parative Musicology
121
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
21
enables you to enter the music on a M
IDI keyboard
—rather like step-
time entry in a sequencer package. T
here are also utilities for playing kern
code (via MID
I), for displaying music notation on screen, or for out-
putting it as a PostScript file (though not all of these are available if youare running H
umdrum
on a PC or a M
ac). You can turn kernfiles into
Finaleones for printing, too.
Unfortunately, none of this addresses the real difficulties of integrating H
um-
drum into everyday m
usicological life. The fact is that not everybody is happy w
itha U
NIX
comm
and-line environment, or finds it easy to rem
ember the different com
-m
ands with their m
ultiple (and often complex) options; people w
ho use UN
IX (or
Hum
drum) on a daily basis can operate it m
uch more quickly than a graphic envi-
ronment (like W
indows, w
ith its drop-down m
enus and mouse control), but if
Hum
drum is to becom
e part of everyday musicological life then w
hat matters is its
accessibility to the occasionaluser. Michael Taylor (1996) has developed a graphical
user interface (GU
I) for Hum
drum, though it is not publicly distributed at the tim
eof w
riting, 17and this m
akes life considerably easier for the occasional user: there aredrop-dow
n menus listing the various H
umdrum
comm
ands and dialog boxes where
you set the associated options, and you can even select musical elem
ents using ordi-nary language (“any bar line,” “at least one flat,” or “C
major chord,” for instance). 18
For nonspecialist users such an interface—if generally available and supported
—w
ould surely represent a substantial step in the right direction: it not only looks fa-m
iliar but also constantly reminds you of the com
mands and options that are avail-
able, without significantly com
promising H
umdrum
’s functionality and flexibility.But such an approach can only go so far. T
his is because, to use Hum
drum at
all, you need quite a detailed knowledge of its representations (kern
and the rest); agreat deal of H
umdrum
usage has to do with things like extracting the spine you are
interested in, merging spines, or stripping out inform
ation that is not wanted for any
particular operation—
and even ordinary-language definitions cannot protect youfrom
the need to understand what is going on in the code. For this reason A
ndreasK
ornstädt (1996) advocates a different approach: instead of “a Hum
drum‘com
mand
center’ with lots of buttons and gadgets w
ith names like yank
and MID
I|smf
,” ashe puts it (1996: 119), he has developed an analytical environm
ent into which dif-
ferent GU
I modules can be slotted for different applications. E
ach module can be
thought of as a “browser” that let users view
just those musical elem
ents that theyare interested in, and no others; for instance, K
ornstädt has created such a systemfor leitm
otivic analysis, in which users can define leitm
otifs and locate their occur-rence in a score on the basis of on-screen notation and intuitive com
mands. T
his,again, is not publicly available, but another of K
ornstädt’s customized applications
of Hum
drum is: an on-line them
atic dictionary, called Them
efinder. 19You type in
some pitches from
the tune you want to find (or alternatively you can type in the
scale degrees, or the intervals between the notes, or even just its contour), and the
software finds all the m
atches in its database and displays them on screen in m
usi-cal notation. T
here are no Hum
drum com
mands; you don’t even need to know
it isH
umdrum
that is doing the work
—and w
hat makes this possible is that T
heme-
finder is designed to do one thing and one thing only. It is hard at the present time
EM
PI
RI
CA
L M
US
IC
OL
OG
Y122
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
22
to tell how far K
ornstädt’s approach—
the development of specialized system
s fordifferent analytical purposes—
will prove to be com
patible with the flexibility and
unpredictability of real-world m
usicological research, or whether it represents a par-
ticularly sophisticated way of falling into the sam
e traps as 1980s office software.
There is also the question of how
the substantial costs of developing and updating acom
prehensive system of this kind are to be found.
Maybe, in the com
ing years, the combination of user-friendly interfaces for
Hum
drum and a developing body of m
usicological work using it w
ill provide thenecessary m
omentum
for it to become m
ore generally used, so that in time it w
illbecom
e just one of those skills that musicologists have to acquire (like using nota-
tion software, for exam
ple). Or m
aybe that is just not going to happen; in that case,com
putational software like H
umdrum
might still becom
e part of the wider m
usi-cological scene, but as a specialist professional service, rather than a skill you acquirefor yourself (rather like note processing w
as a decade or so ago). In other words, if
you had a project in which com
putational approaches could play an important
part—in generating or testing hypotheses on the basis of a large body of data, for in-
stance—then you w
ould call in the services of a specialist computational m
usicolo-gist, w
ho would arrange for any necessary encoding of data, carry out the analysis,
and provide advice on the interpretation of the results. What the ordinary m
usicolo-gist w
ould need to know, then, w
ould be not how to extract or collapse H
umdrum
spines, but how to form
ulate a research question in such a way that advantage can
be taken of computational m
ethods to answer it m
ore securely than is possible usingthe m
ethods traditional in data-poor fields. Whether anyone could m
ake a decentliving as a specialist com
putational musicologist is another m
atter.U
sing computers for m
usicology is like using computers for anything else: you
need to have reasonable expectations about what they can do. H
owever open and
flexible its design, any musicological softw
are is a tool, and like any tool it is goodfor som
e things and not for others. The articles by H
uron and his coworkers that I
have described in this chapter might be characterized as finding m
usicological usesto w
hich computational tools can be put; in essence, they illustrate the softw
are’s po-tential, and they do so in a m
anner that owes m
ore to social-scientific discourse thanto traditional hum
anities writing, w
ith their explicit hypotheses, control groups,quantification of significance, and form
ulation in what H
uron (1999b) has termed
the “boilerplate language” of scientific inquiry. Perhaps that is only to underline thedistinction betw
een “cognitive” or “systematic” m
usicology, as such work is som
etimes
called, and the close attention to context and what m
ight be termed epistem
ologicalpluralism
of the traditional discipline. But what I am
suggesting is that musicology in
the broadest sense can take advantage of computational m
ethods and transform itself
into a data-rich discipline, without giving up on its hum
anist values. Understood that
way, scientific discourse is unlikely to offer a general m
odel for the musicology of the
future: it is up to musicologists to develop such a m
odel. And that does not m
ean try-ing to find m
usicological uses to which com
putational tools can be put, but the re-verse: discovering the tools that are m
ost appropriate to a particular musicological
purpose, and taking full advantage of them. A
t present we hardly have such a thing as
a model of how
sophisticated computational approaches m
ight be integrated within
a sustained, musicologically driven project. Yet the tools are ready and w
aiting. 20
Com
putational and Com
parative Musicology
123
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
23
Notes
1.T
he routines used to create these graphs, which took advantage of the N
eXT
com-
puter’s Postscript display, have not to date been disseminated, but the principles
underlying them w
ere described in chapter 20 of Brinkman 1990.
2.H
elmut Schaffrath, Essen M
usical Data Package
[Menlo Park, C
alif.: Center for C
om-
puter Assisted R
esearch in the Hum
anities (CC
AR
H), 1995] [data and analytical
software on four floppy discs for M
S-DO
S]); 6,225 folksongs are included in thisrelease. Both data and softw
are are now available at w
ww
.esac-data.org. (All U
RLs
were correct at press tim
e, but Web resources change rapidly. Search engines m
ayprovide a better m
eans of access to the resources listed in this chapter.3.
A full listing of the analytical output for “D
eutschland über alles,” reformatted and
annotated for clarity, may be found in Selfridge-Field 1995; see also Schaffrath
1992, 1997.4.
Since EsAC
consists entirely of ASC
II characters, it is possible to import the records
into modern databases. H
owever researchers now
generally prefer to access the kernversion of the databases and process them
using Hum
drum (see note 7 below
).5.
Hence the need for translation program
s such as POC
O (see chapter 8, this volum
e).6.
Digital A
lternate Representation of M
usical Scores; for details see Selfridge-Field1997, chapters 11
–15.
7.D
avid Huron, T
he Hum
drum Toolkit: Softw
are for Music R
esearchers[Stanford, C
alif.:C
enter for Com
puter Assisted R
esearch in the Hum
anities, 1993 (three floppy discsand 16-page installation guide)]. A
wide range of inform
ation may be accessed
from the H
umdrum
Toolkit home page (w
ww
.humdrum
.net) and from the m
ainH
umdrum
site (ww
w.m
usic-cog.ohio-state.edu/H
umdrum
or http://dactyl.som.ohio-state.edu
/Hum
drum).
8.U
NIX
is an operating system, like M
S-DO
S or Window
s, but includes a large num-
ber of general-purpose tools and thus fulfils many of the functions of a program
ming
environment as w
ell. Developed in the late 1960s, it rem
ains the environment of
choice for many professional softw
are developers and researchers.9.
For a summ
ary of kernsee H
uron 1997; for detailed descriptions see Huron 1995,
1999a.10.
Lines beginning with asterisks are know
n as “interpretation records,” those with tw
oasterisks being inclusive
interpretations (something is either in kern or it isn’t), and
those with one asterisk being tandem
interpretations (of which you can have any
number).
11.In a review
of the Hum
drum Toolkit, Jonathan W
ild (1996) includes what is essen-
tially a tutorial on this precise task; assuming that the chorales have been concate-
nated into one file, the entire analysis consists of seven comm
ands pipelined toanother—
in other words, it can be executed w
ith a single comm
and line. Furtheranalytical applications of H
umdrum
are discussed in Huron 2002, w
hich appearedafter this chapter had gone into production.
12.See chapter 1, this volum
e.13.
A follow
-up study (Hippel 2000) reanalyzed the experim
ental data on the basis ofw
hich Burton Rosner and Leonard M
eyer had claimed that gap-fill patterns reflect
listener expectations; von Hippel concludes that that the apparent effects of gap-fill
expectations were an artifact of experim
ental design (specifically, effects of training).O
f course listeners may have such expectations, but von H
ippel’s and Huron’s point
is that these are the result of melodic patterns and not the other w
ay around.
EM
PI
RI
CA
L M
US
IC
OL
OG
Y124
short
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
24
14.See the FA
Q page at the H
umdrum
Toolkit home page (http://dactyl.som
.ohio-state.edu/H
umdrum
/index.html). A
dditional tools developed by Craig Sapp
can be downloaded from
http://ww
w.ccarh.org
/software/hum
drum/m
useinfo.15.
There is a further alternative: a num
ber of Hum
drum com
mands can be run online
at http://musedata.stanford.edu
/software/hum
drum/online; you select the com
-m
and you want, supply the input data (by pasting it into a w
indow, uploading a file,
or supplying its UR
L), set the options, and receive the output as plain text or inH
TM
L. Twenty of the 70 or m
ore Hum
drum com
mands are currently available in
this way, w
ith an emphasis on data translation and conversion.
16.T
here are two principal sources for such m
aterial (from both of w
hich they may be
obtained free): the KernScores site at http://kern.hum
drum.net, and C
CA
RH
’sM
useData site (http://w
ww
.musedata.org).
17.“H
umdrum
Toolkit GU
I” currently exists only in a 16-bit version, running underW
indows 3.1, but further developm
ent (and eventual distribution) is planned by theSonic A
rts Centre at Q
ueen’s University, Belfast.
18.O
nly a few of these “regular expressions,” as they are know
n in UN
IX jargon, are
built into the interface, but you can add your own, along w
ith new interpretations;
this is where you w
ould define **timing or **heartbeat.
19.A
ccessible at http://ww
w.them
efinder.org; for a brief description see Kornstädt
1998.20.
My grateful thanks to D
avid Huron, w
ho has influenced this chapter in many w
ays,not least through m
aking it possible for me to w
ork with him
in 2000 as a visitingscholar at the C
ognitive Musicology Laboratory, O
hio State University, C
olumbus.
Referen
ces
Aarden, B., and H
uron, D. (2001). “M
apping European folksong: G
eographical localiza-tion of m
usical features,” in Walter B. H
ewlett and E
leanor Selfridge Field (eds.), T
he Virnal Score: R
epresentation, Retrieval, R
estoration(C
omputing in M
usicology 12).C
ambridge, M
ass.: MIT
Press, 169–
183.Barlow
, H., and M
orgenstern, S. (1983). A D
ictionary of Musical T
hemes, rev. ed. London:
Faber.Brinkm
an, A. R
. (1986). “Representing m
usical scores for computer analysis.” Journal of
Music T
heory30: 225
–275.
Brinkman, A
. R. (1990). Pascal Program
ming for M
usic Research. C
hicago: Chicago U
niver-sity Press.
Brinkman, A
. R., and M
esiti, M. R
. (1991). “Graphic m
odeling of musical structure.”
Com
puters in Music R
esearch3: 1
–42.
Clendenning, J. P. (1995). “Structural factors in the m
icrocanonic compositions of G
yörgyLigeti,” in E
. W. M
arvin and R. H
ermann (eds.), C
oncert Music, R
ock, and Jazz since1945: Essays and A
nalytical Studies. Rochester, N
.Y.: University of R
ochester Press,229
–256.
Forte, A. (1983). “M
otivic design and structural level in the First Movem
ent of Brahms’s
String Quartet in C
Minor.” M
usical Quarterly
69: 471–
502.H
ippel, P. von (2000). “Questioning a m
elodic archetype: Do listeners use gap
–fill to
classify melodies?” M
usic Perception18: 139
–153.
Hippel, P. von, and H
uron, D. (2000). “W
hy do skips precede melodic reversals? T
heeffect of tessitura on m
elodic structure.” Music Perception
18: 59–
85.
Com
putational and Com
parative Musicology
125
short
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
25
Hughes, M
. (1977). “[Schubert, Op. 94 N
o. 1:] a quantitative approach,” in M. Yeston
(ed.), Readings in Schenker A
nalysis and Other A
pproaches. New
Haven: Yale U
niversityPress, 144
–164.
Huron, D
. (1992). “Design principles in com
puter-based music representation,” in
A.M
arsden and A. Pople (eds.), C
omputer R
epresentations and Models in M
usic.London: A
cademic Press, 5
–39.
Huron, D
. (1995). The H
umdrum
Toolkit: Reference M
anual. Menlo Park, C
alif.: Center for
Com
puter Assisted R
esearch in the Hum
anities [accessible online at http://dactyl.som
.ohio-state.edu/H
umdrum
/comm
ands.toc.html].
Huron, D
. (1996). “The m
elodic arch in Western folksongs.” C
omputing in M
usicology10:
3–
23.H
uron, D. (1997). “H
umdrum
and Kern: Selective feature encoding,” in E
. Selfridge-Field(ed.), Beyond M
IDI: T
he Handbook of M
usical Codes. C
ambridge, M
ass.: MIT
Press,375
–401.
Huron, D
. (1999a), Music R
esearch Using H
umdrum
: A U
ser’s Guide. Stanford, C
alif.:C
enter for Com
puter Assisted R
esearch in the Hum
anities [accessible online athttp://dactyl.som
.ohio-state.edu/H
umdrum
/guide.toc.html].
Huron, D
. (1999b). “Methodology: O
n Finding Field-Appropriate M
ethodologies at theIntersection of the H
umanities and the Social Sciences.” 1999 E
rnest Bloch lectures,U
niversity of California at Berkeley, #3 [accessible online at http://dactyl.som
.ohio-state.edu/M
usic220/Bloch.lectures/3.M
ethodology.html].
Huron, D
. (2001). “What is a m
usical feature? Forte’s analysis of Brahms’s O
pus 51,N
o.1, Revisited.” M
usic Theory O
nline7.4.
Huron, D
. (2002). “Music inform
ation processing using the Hum
drum Toolkit: C
oncepts,exam
ples, and lessons.” Com
puter Music Journal 26
/2: 11–
26.H
uron, D., and J. Berec (forthcom
ing). “Characterizing idiom
atic organization in music:
Atheory and case study.” Journal of N
ew M
usic Research.
Kornstädt, A
. (1996). “SCO
RE
-to-Hum
drum: A
graphical environment for m
usicologicalanalysis.” C
omputing in M
usicology10: 105
–122.
Kornstädt, A
. (1998). “Them
efinder: A w
eb-based melodic search tool.” in W
alter B.H
ewlett and E
leanor Selfridge-Field (eds.), Melodic Sim
ilarity: Concepts, Procedures,
and Applications (C
omputing in M
usicology, II). Cam
bridge, Mass.: M
IT Press,
231–
236.Schaffrath, H
. (1992). “The retrieval of m
onophonic melodies and their variants: C
on-cepts and strategies for com
puter-aided analysis,” in A. M
arsden and A. Pople (eds.),
Com
puter Representations and M
odels in Music. London: A
cademic Press, 95
–109.
Schaffrath, H. (1997). “T
he Essen Associative C
ode,” in Eleanor Selfridge-Field (ed.),
Beyond MID
I: The H
andbook of Musical C
odes. Cam
bridge, Mass.: M
IT Press,
343–
361.Selfridge-Field, E
. (1995). Essen Musical D
ata Package(C
CA
RH
Technical Report N
o. 1),M
enlo Park, Calif.: C
enter for Com
puter Assisted R
esearch in the Hum
anities [dis-tributed w
ith the Essen Musical D
ata Package].Selfridge-Field, E. (1997). Beyond
MID
I: The H
andbook of Musical C
odes. Cam
bridge, Mass.: M
IT Press.
Taylor, W. M
. (1966). “Hum
drumG
raphical User Interface.” M
.A. dissertation., M
usicTechnology program
, Queen’s U
niversity of Belfast.W
ild, J. (1996). “A review
of the Hum
drum Toolkit: U
NIX
tools for music research,
created by David H
uron.” Music T
heory Online
2.7.
EM
PI
RI
CA
L M
US
IC
OL
OG
Y126
06
clarke
.10
3_
12
6 3
/14
/04
8:2
8 P
M P
ag
e 1
26