Pattern Recognition Letters 26 (2005) 299–309
www.elsevier.com/locate/patrec
On musical stylometry—a pattern recognition approach
Eric Backer
a
a,*
, Peter van Kranenburg
b
Faculty of Electrical Engineering, Mathematics and Computer Science, Delft University of Technology, Delft, Netherlands
b
Faculty of Arts, Department of Musicology, University of Utrecht, Utrecht, The Netherlands
Received 21 July 2004; received in revised form 12 October 2004
Abstract
In this short communication we describe some experiments in which methods of statistical pattern recognition are
applied for musical style recognition and disputed musical authorship attribution.
Values of a set of 20 features (also called ‘‘style markers’’) are measured in the scores of a set of compositions, mainly
describing the different sonorities in the compositions. For a first study over 300 different compositions of Bach, Handel, Telemann, Mozart and Haydn were used and from this data set it was shown that even with a few features, the
styles of the various composers could be separated with leave-one-out-error rates varying from 4% to 9% with the
exception of the confusion between Mozart and Haydn which yielded a leave-one-out-error rate of 24%. A second
experiment included 30 fugues from J.S. Bach, W.F. Bach and J.L. Krebs, all of different style and character. With this
data set of compositions of undisputed authorship, the F minor fugue for organ, BWV 534 (of which BachÕs authorship
is disputed) then was confronted. It could be concluded that there is experimental evidence that J.L. Krebs should be
considered in all probability as the composer of the fugue in question.
2004 Elsevier B.V. All rights reserved.
Keywords: Musical style recognition; Authorship attribution; Style markers; Machine learning
In memoriam
It was within the development of the international conferences on pattern recognition, a field
of continuing growth in the early seventies, and
*
Corresponding author.
E-mail addresses: e.backer@ewi.tudelft.nl (E. Backer),
p.vankranenburg@lodebar.nl (P. van Kranenburg).
the establishment of the International Association
for Pattern Recognition (IAPR), starting from the
first ICPR held in Washington, DC in 1973, that, I
first met Azriel and after that, almost yearly in
Board meetings and alike, aiming at serving the
pattern recognition community in the context of
a strong international association, and world wide
organization of the series of biannual conferences.
He was strongly driven and motivated to strengthen
the organization and the impact of the IAPR.
0167-8655/$ - see front matter 2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.patrec.2004.10.016
300
E. Backer, P. van Kranenburg / Pattern Recognition Letters 26 (2005) 299–309
Azriel was most supportive in the process of
founding Pattern Recognition Letters (PRL), in
October 1982, and, in his capacity as president of
IAPR, he established the fact that IAPR became
the official sponsor of the journal PRL. My colleague in founding and managing the journal was
Edzard Gelsema who regretfully passed away
much too early on March 2, 2000. I had the privilege to serve as co-chairman, again together with
Edzard Gelsema, in 1992, in organizing the 11th
IAPR International Conference on Pattern Recognition, at The Hague, The Netherlands. Prof.
RosenfeldÕs compliments on the scientific contents,
outspoken while being there, were meant in the
context of his scientific ideas on image modeling
and picture processing which were the major
subjects at that time. We were proud of his
judgments.
With this short contribution we want to honor a
unique personality as Azriel Rosenfeld was and his
life-long dedication to the pattern recognition
community.
(Eric Backer)
1. Introduction
In the past decades, the ever-increasing power
of computers made it possible to execute pattern
recognition algorithms on a large scale. Those
algorithms can also be of great value in authorship
attribution, resulting in a research area called nontraditional authorship attribution (Love, 2002;
Mason, 1985). This kind of research, tries to quantize the representation of the style of a certain
author (text) or composer (music). Studies of this
kind are called stylometric studies. It is not obvious what exactly has to be quantized but something in the structure of text or musical
composition should bear the ‘‘fingerprint’’ of its
maker. Many so-called style markers are developed in order to classify text or composition to
certain styles and to discriminate between alternatives of authors and composers.
Interesting work has been done by Dannenberg
and Watson (1997). They used machine learning
tools to recognize the ‘‘mood’’ of music, such as
lyrical, frantic, etc. They showed very low error
rates, however, they do not mention all the features
that were used. Also, the work of Pedro Ponce de
León and José Iñesta is worth mentioning, (Ponce
de León and Iñesta, 2003). They used self-organizing neural maps to classify musical styles. Extracted features included basic melody properties
like number of notes, pitch range, etc.
The main problem of stylometry is the lack of
an underlying theory, (Love, 2002). Many style
markers turn out to be distinctive, but often it is
not clear why. Until the study is done, it is not
known which of the style markers (or which combination) will be the discriminator. As a method
for automatically obtaining style markers would
be very desirable but has not been developed up
to now, we have to generate a large number of
potentially interesting features (style markers)
which it is hoped will be suitable for stylometric
studies. This will be the subject of Section 3.
As it is the aim of this study to contribute to the
problem of a disputed authorship of a specific
composition, a fugue known as BWV 534, two
experiments were defined to show that a pre-defined set of 20 style markers (low-level properties
of counterpoint) could be successful.
Experiment 1. To indicate the difference between the style of J.S. Bach and other composers
like Telemann and Handel, as well as to distinguish between composers, like Haydn and Mozart,
whose styles are very alike.
Experiment 2. To test the hypothesis that the
piece BWV 534 is not composed by J.S. Bach,
and most likely is composed by J.L. Krebs and
most likely is not composed by W.F. Bach (J.S.
BachÕs son).
It should be noted that for more than two decades, there are indeed a number scattered musicological contributions about the disputed
authorship of J.S. Bach with respect to BWV 534
(Humphreys, 1985), though not conclusive. The
conjecture that the piece could have been written
by J.L. Krebs is just one of the outcomes of a more
fundamental study of Peter van Kranenburg in his
thesis (Kranenburg, 2004), about the disputed
authorship of BWV 534. The application of
pattern recognition methods on a large scale is
thereby just an attempt to verify some of the presently formulated hypotheses.
E. Backer, P. van Kranenburg / Pattern Recognition Letters 26 (2005) 299–309
2. Data and data preparation
A large corpus of encoded music is available
from the Center for Computer Assisted Research
in the Humanities at Stanford University
(CCARH). 1 From this collection, a number of
compositions are drawn to construct the dataset
that is used in the present study.
The collection of encoded music at the CCARH
consists almost entirely of music from the eighteenth and early nineteenth centuries. Not all of
this is suited for our purpose. Many movements
from cantatas, oratorios and operas have a basso
continuo, which is not completely written out.
So, some harmonic characteristics cannot be determined. These movements are only used when more
than two other voices are active most of the time.
In order to reduce the variance in the computed
feature values, it is also important not to include
too short compositions. After examining the
behavior of the feature values a minimum of 30
bars is taken. Another issue is the presence of
transposing instruments. Sometimes several parts
had to be transposed. Apart from this, many files
needed some adaptations before CPNView 2 could
parse them. With these limitations in mind, a number of compositions is chosen from the CCARH
library.
For experiment 1, the resulting dataset consists
of the following groups of pieces:
• J.S. Bach: 40 cantata movements;
• J.S. Bach: 33 fugues from ‘‘Das Wohltemperierte Clavier’’;
• J.S. Bach: 11 movements from the ‘‘Kunst der
Fuge’’;
• J.S. Bach: 8 movements from the violin
concertos;
• G.F. Handel: 39 movements from the Concerti
Grossi, op. 6;
• G.F. Handel: 14 movements from trio sonatas,
op. 2 and op. 5;
1
<http://www.ccarh.org>.
CPNView: Donncha Ó Maidı́n (University of Limerick,
Ireland).
2
• G.Ph. Telemann: 30 movements
‘‘Fortsetzung des Harmonischen
dienstes’’;
• G.Ph. Telemann: 24 movements
‘‘Musique de table’’;
• F.J. Haydn: 54 movements from
quartets;
• W.A. Mozart: 53 movements from
quartets.
301
from the
Gottestfrom the
the string
the string
Of the three baroque composers works in
different genres are added. Orchestral works as
well as compositions for small instrumentation,
and, in the case of J.S. Bach, works for keyboard.
Of Mozart and Haydn only string quartets were
added.
As mentioned above, the main point of interest
is the difference between the style of J.S. Bach and
the other composers. But it is also interesting to try
to distinguish between composers whose style is
very much the same. Especially the set with Haydn
and Mozart will be challenging, since only compositions of the same genre are included.
For experiment 2, we have been collecting relevant material for comparison of each of the
three—in this study considered—candidates of
authorship of BWV 534.
• J.S. Bach: 11 fugues (different keys, different
time signatures and different date of origin; it
is assumed that all pieces have been composed
by J.S. Bach);
• J.L. Krebs (pupil of J.S. Bach): 8 fugues (as
above; all composed by J.L. Krebs);
• W.F. Bach (J.S. BachÕs son): 5 fugues (as above;
all composed by W.F. Bach).
In order to escape from the curse of dimensionality (and thus aiming at producing a sufficient
amount of data), and at the same time making
use of the length of a composition, we explore
(overlapping) windowing over the entire composition as shown in Fig. 1.
Clearly, we are facing a trade-off between the
number of fragments (as high as possible) and
the variance of the feature values (as small as possible) computed on the basis of the number of bars
in the fragment. From Fig. 2, we observe that a
302
E. Backer, P. van Kranenburg / Pattern Recognition Letters 26 (2005) 299–309
Fig. 1. Windowing over an entire composition.
Also, as the data points generated from one
composition are ordered in time, a composition
is represented in the feature space as a path.
We denote a data set resulting from windowing
as 30_10 data set if we use 30 bars as the window size and 10 bars as the offset. Likewise a
30_01 data set means 30 bars as the window size
and only 1 bar offset (this is the largest overlap
possible).
3. Features (style markers)
Fig. 2. Mean variance of feature values as a function of the
number of bars in a window (fragments of the composition).
choice of 30 bars per fragment seems to be adequate to generate reliable feature values (in terms
of variance). As a consequence, data points obtained from overlapping fragments will be close
to each other in the feature space. Decomposing
a composition (windowing) results in a number
of related data points, enabling us to represent a
composition as a cloud of data points on the
basis of which global densities can be estimated.
We note that the amount of overlap is a free
parameter, which can be used in the analysis at later stage.
For each composition in the dataset, the values
of 20 features are computed (Kranenburg and
Backer, 2004). Most of these features are low-level
properties of counterpoint. When composing polyphonic music, the composer must control the distances between the voices. The way he is doing
this can be expected to be consistent for compositions in different genres and of different dates.
Apart from the distances between the voices, some
other features are computed which can be expected
to be discriminative. Higher-level features (e.g. the
key, modulations, the development of a theme, the
use of certain motifs, etc.) are expected to be less
suitable for our purpose, since they reflect the
characteristics of the individual compositions.
The following features are computed. Some of
them come with an explanation, although in gen-
E. Backer, P. van Kranenburg / Pattern Recognition Letters 26 (2005) 299–309
eral, from a musicological point of view, much is
just speculation.
3.1. StabTimeslice
The ‘‘stability’’ of the length of the successive
time slices. With a time slice the time interval between two changes in the music is meant. This is
shown in Fig. 3. The stability is computed by
dividing the standard deviation of the lengths of
the time slices by the mean length of the time
slices. This normalization is necessary to compare
pieces with different time signatures. So, when having a low value, the music is more like a steady
stream, while a larger value indicates more diversity in rhythm.
3.2. DissPart
The fraction of the score that consists of dissonant sonorities. Consonants are: perfect primes,
minor and major thirds, perfect fourths and fifths
and minor and major sixths. But a fourth is considered dissonant if it is between the lowest voice
and one of the upper voices. All other intervals
are considered dissonant. The total duration of
dissonant sonorities is divided by the total duration of the composition.
303
regardless of inversion or pitch. A unique number
represents each sonority. For each sonority the
total duration of all occurrences is computed.
Then the probabilities of occurrence are estimated
using these weighted frequencies. With this probabilities the entropy is computed according to:
N
X
pi logðpi Þ
i¼1
where N is the total number of sonorities and pi the
probability of occurrence of sonority i.
3.5. HarmonyEntropy
Mason also defines the concept ‘‘Harmony’’. It
is much like sonority, but now difference is made
in pitch. So e.g. a F-major triad and a G-major
triad are the same sonority but different harmonies. Again the inversion is not taken into account.
The value of this feature is computed the same way
as the Sonority Entropy.
3.6. PitchEntropy
A list of occurrences of all pitches is made.
Again the occurrences are weighted by the durations. Of the resulting list, the entropy is
computed.
3.3. BeginBarDiss
3.7. VoiceDensity
The fraction of bars that begins with a dissonant sonority.
3.4. SonorityEntropy
For this feature, the concept ‘‘sonority’’ is used
according to the definition of Mason (1985) In this
definition sonority is a certain type of chord. So
e.g. all the major triads are the same sonority,
In a polyphonic composition not all voices are
active during the whole composition. The voice
density indicates the average number of active
voices. This is normalized with the total number
of voices. For this feature only bars that are
strictly polyphonic are taken into account i.e. bars
in which no voice has more than one note and in
which more than one voice is active.
3.8. PartSeconds, PartThirds, PartFourths,
PartAugFourths, PartDimFifths, PartSixths,
PartSevenths, PartOctave
Fig. 3. Boundaries of the time slice.
When combining the different voices of a polyphonic composition, the composer has to obey certain constraints. In many of these constraints the
304
E. Backer, P. van Kranenburg / Pattern Recognition Letters 26 (2005) 299–309
vertical distances between the voices are important. This set of features measures the amount of
a number of intervals between the different voicepairs. Systematically all voice-pairs are examined.
The total duration of all occurrences of each specific interval is computed and at the end divided
by the total duration of all intervals in all voicepairs. The intervals are taken modulo one octave.
So e.g. a tenth is a third. When the same pitch occurs in more than one voice, it is taken into account once.
3.9. ParThirds, parFourths, parSixths
It can happen that in a voice pair two intervals
of the same size succeed each other. This is called a
parallel. For these three features the amount of
parallel thirds, fourths and sixths is computed in
the same way as the previous group of features.
The total duration of all intervals involved in these
parallels is added and divided by the total duration
of all intervals in all voice pairs.
3.10. StepSuspension
When a dissonant is sounding between two
voices, it often is suspended into a consonant by
lowering the lower voice one step. This feature
indicates how many dissonances are suspended
this way. It is computed in the same way as the
previous features.In the remaining these features
are referred to by their index numbers. These can
be found in Table 1.
Table 1
The feature set (style markers)
Index
Feature
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
StabTimeslice
DissPart
BeginBarDiss
SonorityEntropy
HarmonyEntropy
PichEntropy
VoiceDensity
PartSeconds
PartThirds
PartFourhs
PartAugFourths
PartDimFifths
PartFifths
PartSixths
PartSevenths
PartOctaves
ParThirds
ParFourths
ParSixths
StepSuspension
Likewise, we have class arrangements like {J.S.
Bach}{W.F.
Bach}{J.L.
Krebs},
{J.S.
Bach}{W.F. Bach}, {J.S. Bach}{J.L. Krebs} and
{W.F. Bach}{J.L. Krebs} for experiment 2.
Fig. 4 shows the distributions of some of the
‘‘best’’ features to discriminate between the classes
10
5
0.5
1
30
0
0
6
ParThirds
All experiments are carried out with the Matlab-toolbox PRTools. 3
For both experiments, we perform feature selection using the Floating Forward Selection algorithm, proposed by Pudil et al. (1994). The
FFS algorithm is applied to all possible class
arrangements like {Bach}{all other composers},
{Bach}{Telemann}{Handel} etc. for experiment 1.
3
<http://www.ph.tn.tudelft.nl/~bob/PRTOOLS.html>.
20
4
10
2
0
0
2
DissPart
5
0
0
4. Analysis
10
StabTimeslice
0.1
0.2
0.3
0.4
0
4
0.2
0.6
0.8
PitchEntropy
4.5
30
HarmonyEntropy
0.4
5
5.5
StepSuspension
20
1
10
0
4
5
6
7
0
0
0.1
0.2
0.3
Fig. 4. Some of the ‘‘best’’ features for class arrangement
{Bach}{notBach}; Bach = solid, not-Bach = dashed.
E. Backer, P. van Kranenburg / Pattern Recognition Letters 26 (2005) 299–309
I (DissPart > 0.358065)
0.25
ParThirds
0.2
0.15
0.1
0.05
0.2
0.4
0.6
0.8
1
StabTimeslice
Fig. 5. Scatter plot with features that characterize BachÕs style
(+). Compositions with DissPart 6 0.358 are not shown.
{Bach} and {not-Bach} and Fig. 5 a simple decision boundary obtained from these three features.
For comparison of the two extremes in class
arrangements {Bach}{Telemann, Handel, Haydn,
Mozart} (a two-class problem; a) and
{Bach}{Telemann}{Handel}{Haydn}{Mozart}
(as a five-class problem; b), we observe the following classification statistics when using a k-nearest
neighbor classifier (see Table 2).
The five-class problem is obscured by the presence of compositions of Haydn and Mozart. The
arrangement: {Haydn}{Mozart} yields a leaveone-out error (l-o-o-error) of 24.30%. Here we observe a significant limitation of the used music
library (CCARH) as Haydn and Mozart were only
represented by a collection of string quartets,
obscuring the recognition results for the five-class
arrangement. All other arrangements not including discrimination between Haydn and Mozart
yield errors between 5% and 9%. A full account
of the above results is given in (Kranenburg and
Backer, 2004).
From these results we conclude that BachÕs style
can be isolated from the style of other composers
with such a performance that it might be regarded
as a valuable addition to the traditional methods
of musical style classification. It offers a quantitative evaluation of the styles rather than the traditional qualitative descriptions. It is important not
305
to see this as a replacement, but as an addition.
Combining results from different viewpoints, will
give more robust knowledge. The results of the
above studies are a promise for the future, in
which we can expect further increase in the computational power as well as further increase in the
understanding and application of pattern recognition techniques.
This also means that this kind of research can
be helpful in authorship disputes. This is the origin
of experiment 2. 4
Some of the features (from Section 2) are displayed in Fig. 6 for the class arrangement {J.S.
Bach}{J.L. Krebs}{W.F. Bach}. The densities
are estimated using a 30_01 data set (maximum
overlap). None of the features are perfect discriminants, however a combination of six features used
for training of a quadratic Bayesian classifier with
10-fold-cross-validation yields a (still optimistic)
error of 1.2% for the discrimination of {J.S.
Bach}{J.L. Krebs} arrangement, 1.6% for the discrimination of {J.S. Bach}{W.F. Bach} arrangement, and 1.6% for the discrimination of {W.F.
Bach}{J.L. Krebs} arrangement.
We are using the Fisher Linear Discriminant
transformation over the entire feature space to
visually interpret the ‘‘best’’ two-dimensional scatter plot (discriminants 1 and 2). Fig. 7 shows the
resulting scatter plot of the transformed data set
with classes {J.S. Bach}, {J.L. Krebs} and {W.F.
Bach} with in overlay—as an example—BWV
535, a fugue of J.S. Bach of which authorship is
certainly not-disputed.
In order to interpret the features used in decision making of the different class arrangements,
we generate the corresponding decision trees
(C4.5).
We observe:
1. For the class arrangement {J.S. Bach}{J.L.
Krebs} the decision tree (Fig. 8a) uses four features StabTimeslice, PartSeconds, PartThirds,
and PartFourths; only one fragment of the
30_10 data set is misclassified.
4
For all details about data: http://www.musical-stylerecognition.net.
306
E. Backer, P. van Kranenburg / Pattern Recognition Letters 26 (2005) 299–309
Table 2
Minimal errors with nearest neighbor classsifiers
Class arrangement
k
Feature subset
l-o-o-error (%)
Two-class problem (a)
Five-class problem (b)
5
11
1, 17, 2
2, 13, 8, 9, 1, 17, 5, 10, 14, 19, 11, 6, 7, 20
6.62
26.47
100
70
60
80
50
60
40
30
40
20
20
10
0
–0.05
0
0.05
0.1
0.15
0
–0.05
0
PartSeconds
0.05
0.1
0.15
0.2
PartSevenths
25
20
15
10
5
0
–0.2
0
0.2
0.4
0.6
0.8
StabTimeslice
Fig. 6. Some densities of features for the class arrangement {J.S. Bach}{J.L. Krebs}{W.F. Bach}; J.S. Bach = solid, J.L.
Krebs = stripes, and W.F. Bach = dashed.
2. For the class arrangement {J.S. Bach}{W.F.
Bach} the decision tree (Fig. 8b) uses four features: ParthSevenths, StabTimeslice, PartOctaves, and StepSuspension; only one fragment
is misclassified.
3. For the class arrangement {J.S. Bach}{W.F.
Bach} the decision tree uses two features:
BeginBarDiss and ParThirds; no errors occur.
We are now ready to classify the disputed f fugue for organ, BWV 534.
1. If we assume that J.S. Bach and W.F. Bach were
the only candidates, we observe the following.
The quadratic Bayesian classifier, trained with
all features from the transformed Fisher space
assigns all fragments of BWV 534 to {J.S.
Bach}.
With the six best features selected by FFS, 10
(out of 11) fragments are assigned to {J.S.
Bach} for 30_10 data.
The decision tree from Fig. 8b also assigns all
fragments of BWV 534 to {J.S. Bach}. There-
E. Backer, P. van Kranenburg / Pattern Recognition Letters 26 (2005) 299–309
bwv535
4
Discriminant 2
2
0
−2
−4
−6
−8
−6
−4
−2
0
2
4
6
Discriminant 1
Fig. 7. Scatter plot of the fugue data set in the Fishertransformed feature space; J.S. Bach (+), J.L. Krebs (*) and
W.F. Bach ( ); BWV 535—as an example—in overlay (bold
stars).
307
fore, it is save to conclude from these observations that the hypothesis of W.F. Bach being
considered as the composer is false.
2. If we assume that J.S. Bach and J.L. Krebs were
the only candidates, we observe the following.
The quadratic Bayesian classifier, trained with
all features of the transformed Fisher space assigns all fragments of BWV 534 to {J.L. Krebs}.
Also, with the best six features selected by FFS,
all fragments are assigned to {J.L. Krebs} for
30_10 data.
The decision tree from Fig. 8a assigns five fragments to {J.S. Bach} and six fragments to {J.L.
Krebs}. Therefore, it is still save to conclude
that the style of BWV 534 resembles the style
of {J.L. Krebs} more than the style of {J.S.
Bach} and that the hypothesis of J.L.
Krebs being considered as the composer, is true.
3. If we compare the styles of J.L. Krebs and W.F.
Bach, we observe the following.
The quadratic Bayesian classifier, trained with
all features from the transformed Fisher space
Fig. 8. Decision tree: (a) J.S. Bach versus J.L. Krebs (30_10 data set); (b) J.S. Bach versus W.F. Bach (30_10 data set); (c) J.L. Kerbs
versus W.F. Bach (30_10 data set).
308
E. Backer, P. van Kranenburg / Pattern Recognition Letters 26 (2005) 299–309
bwv534
Discriminant 2
4
2
0
−2
−4
−6
−8
−6
−4
−2
0
Discriminant 1
2
4
6
Fig. 9. Scatter plot of the disputed fugue BWV 534 (bold stars)
in the Fisher-transformed feature space; J.S. Bach (+), J.L.
Krebs (*) and W.F. Bach ( ).
assigns all fragments of BWV 534 to {J.L.
Krebs}.
With the best six features selected by FFS, all
fragments are assigned to {J.L. Krebs} for
30_10 data.
The decision tree from Fig. 8c also assigns all
fragments of BMV 534 to {J.L. Krebs}. Therefore, it is save to conclude that if the choice had
to be made between J.L. Krebs and W.F. Bach,
there is no doubt in considering J.L. Krebs as
the composer.
20 low-level properties of counterpoint to be measured in the represented score of a composition. It
was concluded that it is very possible to isolate the
style of J.S. Bach from other composers like Telemann, Handel, Haydn or Mozart. Given the positive outcome, it has been a challenge to enter the
field of non-traditional author attribution.
So, we conducted a second experiment to investigate how well disputed musical authorship of a
given composition could be solved if a limited
number of alternatives are given a priori. In our
case, the fugue BWV 534 which has been attributed to J.S. Bach but of which real authorship
has been disputed on musicological grounds. His
son, W.F. Bach and his pupil, J.L. Krebs have
been put forward as serious candidates for true
authorship.
From the experiment, it safely could be concluded that there is experimental evidence that
J.L. Krebs has to be considered in all probability
as the real composer.
Acknowledgement
The continuous support of Dr. Albert Clement,
musicologist and Bach-expert of the University of
Utrecht, The Netherlands, is very much
appreciated.
References
In Fig. 9, the fragments of the disputed fugue
BWV 534 are displayed in overlay with the twodimensional Fisher transformed feature space, nicely indicating how well the fragments fit into the
available data of J.L. Krebs.
5. Conclusions
In this short communication, we have presented
an attempt to apply pattern recognition techniques
in the area of musical style characterization and
disputed musical authorship.
First, we conducted an experiment to investigate how well the style of different composers
could be identified. For that purpose, we designed
Dannenberg, T., Watson, D., 1997. A machine learning
approach to musical style recognition. In: Proceedings of
the 1997 International Computer Music Conference, pp.
344–347.
Humphreys, D., 1985. Did J.S. Bach Compose the F minor
Prelude and Fugue BWV 534? In: Williams, P. (Ed.), Bach,
Handel, Scarlatti—Tercentenary Essays. Cambridge, pp.
173–184.
Kranenburg, P. van, Backer, E., 2004. Musical style recognition, a quantitative approach. In: Proceedings of the
Conference on Interdisciplinary Musicology, Graz, Austria.
Kranenburg, P. van, 2004. De fuga in f (BWV 534) van Johann
Sebastian Bach? Thesis. University of Utrecht. Available
from: <www.Musical-Style-Recognition.net> (in Dutch).
Love, H., 2002. Attributing Authorship: an Introduction.
Cambridge.
Mason, R.M., 1985. Modern Methods of Musical Analysis
using Computers. Peterborough.
E. Backer, P. van Kranenburg / Pattern Recognition Letters 26 (2005) 299–309
Ponce de León, P.J., Iñesta, J.M., 2003. Feature-driven
recognition of musical styles. Lecture Notes in Computer
Science 2652, 773–781.
309
Pudil, P., Novovicova, J., Kittler, J., 1994. Feature Search
Methods in Feature Selection. Pattern Recognition Letters
15, 1119–1125.