Proceedings of the International Computer Music Conference (1997)

Classification in Music: A Computational Model for

Paradigmatic Analysis
Christina Anagnostopoulou (1)
Gert Westermann (2)
(1) Faculty of Music and Dept of Artificial Intelligence, University of Edinburgh
chrisa@music.ed.ac.uk
(2) Centre for Cognitive Science, University of Edinburgh
gert@cogsci.ed.ac.uk

Abstract
We present a computational model for the paradigmatic analysis of musical pieces, which is the
classification of musical segments into similarity-based categories. The model requires the analyst
to make explicit choices for the characteristics by which the musical segments are described. The
classification of the segments into categories is determined by these characteristics and is performed
by a self-organizing neural network algorithm. In this way, traditional problems associated with
paradigmatic analysis, namely lack of consistency and objectivity, can be avoided. Moreover, the
model extends the analytical technique by providing different levels of classification, prototypes for
each class, and by showing relations between classes.

1 Introduction within a musical piece are made explicit. Further

analysis is thus facilitatedthis can be distributional,
The paradigmatic analysis (henceforth PA) of mu- comparative or stylistic.
sical pieces has long been criticized for its reliance
on intuition and the resulting inconsistencies [1]. In The main original goal of PA was to give a for-
this paper, we describe a formal model for this task mal, objective account of the material used in a piece,
as a way to address such criticisms. In the next sec- not taking into account the composers intentions
tion, traditional PA is described and its shortcomings or the listeners perceptions. In practice, however,
are discussed. The formal model is then presented, the assignment of the musical segments to different
and its functioning is demonstrated by analyzing De- classes usually relies on intuition: According to Nat-
bussys Syrinx and comparing the results with J.J. tiez, People decide to associate several units in a sin-
Nattiez (the leading figure in PA) second analysis of gle paradigm because of semantic or psychological cri-
this piece [4]. We conclude with a discussion of the teria that they do not express consciously. (quoted
model and suggestions for further work. in [1], p. 180). This lack of explicit criteria underly-
ing the classification will naturally lead to inconsis-
tencies in the analysis, and this has in fact been the
2 Paradigmatic Analysis main criticism of PA. Further criticisms address its
limited character: there is only one level of classifi-
PA consists in the segmentation of a piece of mu- cation, when subcategories could easily be identified
sic and the classification of these segments into cat- and could prove to be useful. Relationships between
egories according to their similarity. The first oc- different classes are not considered, although some
currence of a segment in each class is called the classes will be more similar than others. Moreover, as
paradigm, and subsequent segments are compared to in most other analytical techniques, the segmentation
these paradigms to determine their class membership. of a piece has been criticized as being usually infor-
The paradigms therefore play the role of class proto- mal, which is a problem for the subsequent classifica-
types. tion. Finally, the paradigms against which other seg-
The motivation for this kind of analysis is that ments in a category are compared are merely a first
repetition, variation, transformation, and contrast occurrence and not necessarily prototypical of their
 Segmentation
Featurevector Formal
Representation Paradigmatic
Classification OK
of Segments Analysis
Algorithm
Representation
of Musical Features

Reevaluate
Figure 1: A formal model for Paradigmatic Analysis

category (see section 4 for an example from Nattiez In summary, a formal model of paradigmatic
analysis), and potential inconsistencies may arise by analysis serves as a tool for the analyst, forcing her to
comparing each segment to a paradigm which is not make her choices of representation explicit and pro-
prototypical. viding a well-defined algorithm for the clustering of
segments, without restricting the freedom to choose
the classification criteria.
3 A Formal Model
3.1 Segmentation
Classifying objects is a fundamental task which has
been studied in depth in other disciplines, such as It is generally accepted that there is no single cor-
formal learning theory, computer science, and psy- rect way of segmenting a piece of music. Segmen-
chology. However, the existing classification theories tation is a problematic issue for any kind of musi-
and techniques from those disciplines have not gen- cal analysis, and therefore ideally a system should
erally influenced music analysis (but see [3]). accept any analysts choice on segmentation. The
In a formal model, the algorithm by which the modular character of the present system allows this
segments of the piece are classified has to be explicitly approach. In that way, different segmentations can be
defined. Also, the feature representations on which comparedit is obvious that the most sensible ones
the classification is based have to be made explicit, will result in the most sensible classifications.
so they can serve as input to the classification algo- In our example we used J.J. Nattiez segmenta-
rithm. A formal system further leads to the modu- tion from his second PA of Debussys Syrinx for solo
larization of PA into different subtasks, each of which flute [4].
can be solved independently. Figure 1 shows such a
formal model. In the first step, the musical piece is 3.2 Representation of Musical Fea-
segmented, and the analyst has to decide on the way
tures
in which musical features are to be represented. In
the second step, each segment is expressed as a list of Each segment is described in the formal model as
features (a feature vector ), which is then used as in- a list of features. The term feature is used here
put to a classification algorithm. The output of this not only in the traditional sense, that is with bi-
module constitutes a PA. If the analyst is not satis- nary values (yes/no) , but also more generally, to
fied with the clustering obtained, he will re-evaluate include multi-valued attributes, and in fact any hier-
the feature representations. For example, if two seg- archic relation in a semantic network. An example
ments which the analyst considers to be different are of a feature with a binary value would be the exis-
grouped together by the model, he will introduce a tence of a grace-note in a segment. An example of
feature that distinguishes these segments from one a multi-valued feature would be instrument register:
another. Based on the resulting new representation it could be the first octave of the flute, the second
of the segments, a different classification will occur. or the third, or any combination of these. Examples
This process is repeated until a satisfactory classifi- of hierarchic relations are shown below for melodic
cation is obtained. This final classification will be shape and rhythmic movement.
based on explicit segments and features and will be It is obvious that the results of the PA will de-
free from inconsistencies. pend crucially on the feature selection. The analyst
 Segm. Music Notation Feat. vector Class

2
 6
1, 6 1 0 0 0 0 I

2
 2
Figure 2: Construction of a GNG network. Small cir-
4 0 1 0 0 0 I or II cles represent input data, and large circles connected

2
 6
with edges are the units of the network.

2, 7 1 0 0 0 0 I classification algorithm. Table 1 shows several seg-

9 ments out of the 79 from Syrinx with part of their
2    
2
24, 28 0 1 0 0 0

3
2 2 2 2 2
52 0 1 0 0 0
Table 1: Classification examples for several segments
from Syrinx.
will choose the features according to the desired out-
come: for example, he might choose to focus on a
rhythmic analysis, or compare several pieces of music
according to a set of common features. Since similar-
ity in music could be argued to be context-dependent
(context being the piece or pieces under analysis),
features can be low-level, piece-specific (e.g., the use
of a specific interval), as well as very general musical
properties (like upward melodic motion).
For our experiments we chose a combination of
general and of piece-specific features, describing
melodic shape (movingup, down, or different
combinations or stationary).
rhythmic movement (continuouswhich can
be quaver, semiquaver or demisemiquaver
movement or interruptedby a dotted
rhythm, syncopation, long note or a pause).
interval patterns (with instances being low-level
successions of intervals) and
instrument register (in order to describe trans-
position).
These features proved to be sufficient for the final
classification.
The features describing a segment were concate-
nated to form a feature vector, in our case with 40
binary values, which was then used as input to the
feature vector representations.
II
3.3 The Classification Algorithm
The classification algorithm which we chose to em-
II
ploy for our experiments was Growing Neural Gas
[2]. This is an unsupervised neural network algo-
rithm that grows units while it learns. Each unit
corresponds to the prototype of one cluster. An in-
put signal, i.e., a feature vector representing a musical
segment through 40 binary features, can be viewed as
having a position in the 40-dimensional input space,
and the units of the network are positioned in the
same space. When an input signal is presented, the
unit of the network which is closest to it (measured by
Euclidean distance) is moved towards this signal by
a fraction of the distance to this signal, together with
its topological neighbours. The distance between the
signal and the winning unit is added to a local er-
ror variable of this unit. The winning unit and the
second closest unit are then connected by an edge,
or the age of the edge is reset to zero, if it already
exists. The edges reflect neighbourhood relations be-
tween the network units. At each step, all edges in
the network are aged, and edges which have reached
a pre-defined maximum age are deleted. This process
ensures a continuous updating of the neighbourhood
relations.
A new unit is inserted into the network at regular
intervals, between the unit with the highest accumu-
lated error and its neighbour with the highest error.
The built-up structure of the network reflects the dis-
tribution and density of the input signals: The units
move towards the input signals, and a high density
of inputs in an area will lead to more units being
allocated in this area. Figure 2 shows the develop-
ment of a network in a two-dimensional input space
with four distinct clusters. The network starts with
two units and can therefore distinguish only between
the two main clusters. In effect, the network answers
the question: If there were two clusters, what would
their prototypes be? After a certain amount of epochs
(presentation of the input signals), a new unit is in- 5 Conclusions
serted and the units move to the positions indicated
in the second picture. When the fourth unit is in- We have demonstrated a formal model of paradig-
serted, the units distribute over the four clusters. matic analysis. The model requires the analyst to
In principal, insertion of units proceeds forever. make the categorization criteria explicit without re-
The GNG algorithm thus lets the analyst define the stricting his particular choices. The classification pro-
level of grainedness of her analysis and does not im- duced by the model is then entirely based on these
pose a priori constraints on the number of clusters. choices and depending on the obtained results, they
Each unit forms a prototype of a cluster, expressed can be revised by the analyst. The model yields dif-
in the probability distribution of the feature val- ferent levels of classification, prototypes for each class
ues of their cluster members. These prototypes are and relations of classes of the same level.
frequency-dependent, since the position of each unit The model can be extended in various ways.
is updated with each presentation of an input sig- Firstly, it is possible that the analyst wishes to
nal. Neighbourhood relations between clusters are attribute different importance to different features.
expressed in the connections between the network Such a weighting of features should be incorporated
units. into an extended model. Secondly, while the present
version relies on a given segmentation of a piece, in
principle a revision of this initial segmentation could
be incorporated into the classification process, com-
4 Experimental Results bining the stages of PA in a single unified model. This
The GNG algorithm was trained on the musical fea- will be our main direction of future research.
ture vectors for 2000 iterations, inserting a new unit
every 100 iterations (2 minutes CPU time on a Sun 6 Acknowledgements
Ultra workstation). Thus, the final classification con-
sisted of 20 categories, and by comparing this to pre- We are grateful to Peter Nelson for comments on a
vious stages, the hierarchy of clusters could be ob- draft of this paper. The second author was funded
served. by the ESRC (award no. R00429624342) and by
We ran various experiments with different input the Gottlieb Daimler-und Karl Benz-Stiftung (grant
representations. With our final representation (which no. 02.95.29).
is mentioned above) we obtained an intuitively sen-
sible analysis which was surprisingly close to Nat-
tiez second paradigmatic analysis. Due to the lack References
of space it is impossible to give the various results
obtained in full length. In table 1, segments 1, 2, 6 [1] N. Cook. A Guide to Musical Analysis. Oxford
and 7 belong to the same class; segments 24, 28 and University Press, 1987.
52 belong to another. Segment 4 is a problematic [2] Bernd Fritzke. A growing neural gas algorithm
segment in that it can be classified to either of these learns topologies. In G. Tesauro, D. S. Touretzky,
two categories, according to different input represen- and T. K. Lean, editors, Advances in Neural In-
tations. It would be classified with the first category formation Processing Systems 7, pages 625632,
when rhythmic features are taken into account, and Cambridge, MA, 1995. MIT-Press.
with the second if melodic shape is emphasized. In
our experiments, after 2000 iterations, this segment [3] R.O. Gjerdingen. Using connectionist models to
formed a class by itself, but was linked by edges to explore complex musical patterns. In Music and
both class I and class II. Connectionism, pages 138149. MIT Press, 1991.
Nattiez classified segment 4 according to melodic
shape and not to rhythm. The problem with this is [4] J.J. Nattiez. Fondements dune Semiologie de la
that since segment 4 is the first occurrence for such Musique. Union Generale dEditions, 1975.
a melodic shape, it becomes a paradigm. However,
further segments are more and more varied, and the
result is that segment 4 is in no way prototypical of
the whole category. In our analysis, this problem was
avoided because prototypes represented the weighted
average of all class members and not the first occur-
rence of a class member.