TOWARDS STRUCTURAL ALIGNMENT OF FOLK SONGS
J¨ org Garbers and Frans Wiering
Department of Information and Computing Sciences
We describe an alignment-based similarity framework
for folk song variation research. The framework makes
use of phrase and meter information encoded in Hum-
drum scores. Local similarity measures are used to
compute match scores, which are combined with gap
scores to form increasingly larger alignments and
higher-level similarity values. We discuss the effects
of some similarity measures on the alignment of four
groups of melodies that are variants of each other.
In the process of oral transmission folk songs are re-
shaped in many different variants. Given a collection
of tunes, recorded in a particular region or environ-
ment, folk song researchers try to reconstruct the ge-
netic relation between folk songs. For this they study
historical and musical relations of tunes to other tunes
and to already established folk song prototypes.
It has often been claimed that their work could ben-
efit from support by music information retrieval (MIR)
tice however it turns out that existing systems do not
work well enough out of the box . Therefore the re-
search context must be analyzed and existing methods
must be adapted and non-trivially combined to deliver
1.1 Similarity and alignment
Similarity and alignment can be considered two sides
of the same coin. In order to produce an automatic
alignment we need a measure for the relatedness of
musical units. Conversely, in order to compute the (lo-
cal) similarity between two melodies we must know
which parts of the melody should be compared.
similarity measures. In a previous paper we derived
generalized queries from a group of folk song vari-
ants . For a given group of musically related query
melodies aligned by the user, we were able to retrieve
bers for this group.
Making a manual alignment is time-consuming and
involves edit decisions, e.g. ‘shall one insert a rest in
one melody or delete a note in the other?’. When look-
ing for good additional group members in a database,
one should allow both options. However, keeping track
of all options quickly becomes impracticable. In this
paper we therefore look into automatic alignment of
corresponding score positions and ways of controlling
the alignment with basic similarity measures.
1.2 Overview and related work
In this paper we first discuss why automatic detection
of (genetically related) folk song variants is very de-
manding and is a major research topic in its own. Next,
to support research into similarity measures based on
musically meaningful transformations, we develop a
framework that helps to model the influence of local
similarity measures on variation detection and align-
ment. Starting from the information encoded in our
folk song collection, we motivate the use of available
structural and metrical information within alignment
directed similarity measures. Finally we compare au-
tomatically derived alignments with alignments anno-
tated by an expert.
Generally, we follow a similar approach to Mon-
geau and Sankoff’s , who tackled selected transfor-
mational aspects in a generic way. They set up a frame-
work to handle pitch contour and rhythm in relation to
an alignment-based dissimilarity (or quality) measure.
They based their framework assumptions on musical
be assigned to the second occurrence. This ambiguity
accounts for many “failures” in the test logs.
Group G4 (Heer Halewijn) was chosen because of
its complexity. Only when looking at the annotation
for a while the chosen alignment becomes understand-
able. It ismainlybased ontonal functionof pitchesand
contains many inner gaps. For pairs of phrases, other
alignments are plausible as well, but in the multiple
alignment several small hints together make the given
annotation convincing. Therefore there are only few
correct automatic alignments. Interestingly, however,
the algorithm manages to align a subgroup (72256,
74003, and 74216) without failure.
ilarity framework for folk song melodies represented
as scores. We have done initial tests that show both the
usefulness and limitations of our segmentation, align-
ment and evaluation approach. We see two continu-
ations. First, we should use the framework to study
similarity seeds that take the observed stability of be-
ginnings and endings into account (see section 2.2).
Second, the alignment framework needs to be de-
veloped further into several directions. 1) We did not
so far pay any attention to the relationship between the
statistical properties of the distance measure, its nor-
malization and the value of the gap penalties.
should support the modeling of states and non-linear
gap costs. 3) Multiple alignment strategies should be
incorporated in order to relate more than two melodies.
The need for this became apparent in the last alignment
group. Multiple alignments are particularly needed for
group queries . Therefore we will not only evaluate
the quality of the alignments but also the performance
of melody retrieval using these alignments.
Acknowledgments. This work was supported by the
Netherlands Organization for Scientific Research within the
WITCHCRAFT project NWO 640-003-501, which is part
of the CATCH-program.
 Log files for this paper. http://pierement.
Signalverarbeitung. Modified Mongeau-Sankoff
 J. Garbers, P. van Kranenburg, A. Volk, F. Wier-
ing, L. Grijp, and R. C. Veltkamp. Using pitch sta-
bility among a group of aligned query melodies
to retrieve unidentified variant melodies. In Si-
mon Dixon, David Bainbridge, and Rainer Typke,
editors, Proceedings of the eighth International
Conference on Music Information Retrieval, pages
451–456. Austrian Computer Society, 2007.
 J. Garbers, A. Volk, P. van Kranenburg, F. Wier-
ing, L. Grijp, and R. C. Veltkamp. On pitch and
chord stability in folk song variation retrieval. In
Proceedings of the first International Conference
of the Society for Mathematics and Computation in
Music, 2007. (http://www.mcm2007.info/
 P. van Kranenburg, J. Garbers, A. Volk, F. Wiering,
L.P. Grijp, and R. C. Veltkamp. Towards integra-
tion of MIR and folk song research. In ISMIR 2007
proceedings, pages 505–508, 2007.
 M. Mongeau and D. Sankoff. Comparison of mu-
sical sequences. In Computers and the Humanities,
volume 24, Number 3. Springer Netherlands, 1990.
 C. Sapp. Humdrum extras (source code). http:
 Rainer Typke. Music retrieval based on melodic
similarity. PhD thesis, Utrecht University, 2007.
 A. Volk, P. van Kranenburg, J. Garbers, F. Wiering,
R. C. Veltkamp, and L.P. Grijp. A manual annota-
tion method for melodic similarity and the study of
melody feature sets. In ISMIR 2008 proceedings.
 Walter Wiora. Systematik der musikalischen Er-
scheinungen des Umsingens. In Jahrbuch f¨ ur
Volksliedforschung 7, pages 128–195. Deutsches