Automatic Tempo-based Classification of Argentine Tango Music

Vlasia Anagnostopoulou
([email protected])

Report for the Fall 2010 CS290I class taught by Prof. Yuan-Fang Wang

Abstract

Tempo-based classification is a popular technique to

classify music, with applications in media-art projects,
systematic study of music etc. The success of this
technique using dance music genres has typically been
found to be over 70%, yet it was mostly evaluated using
standardized music genres, such as ballroom music.
Argentine tango is a music genre, originating in Buenos
Aires at the end of the 19th century, which has gained a
lot of popularity since its creation and is currently being
danced in social settings all over the world (see side
Photo taken in San Francisco, courtesy of Ref. [10]).
There are three music types which define the Argentine
tango played in social settings; although these types have
different meters from one another, their tempi feature
variations influenced by the trends of the era of
composition and the composersʼ personal style.
Therefore, Argentine tango seemingly makes a good, yet
interesting candidate to explore for tempo-based
classification schemes.

In this report, we construct and evaluate an automatic,

bayes tempo classifier with application to Argentine tango
music, using a library with more than 2,000 pieces. We
find that the accuracy of our classifier stands at 79%,
which is well within the bound set by similar schemes
applied to ballroom music, proving our original point that
tempo is a promising feature in the classification of
Argentine tango. We also come up with interesting
insights about how to improve classification schemes for this particular musical genre.

Introduction

Tempo-based classification is a popular method to classify a collection of music, as can be

shown by the proliferation of music classifying tools [References 5 and 6], which has been
particularly successful in the classification of dancing music. The reason why this method has
been particularly successful with dancing music is because dancing music typically features a
tight tempo distribution, and different dancing genres have distinct tempi, making it a "clear"
choice for the classifier to guess the right genre based on tempo. Yet the accuracy of
approaches using tempo alone cannot typically exceed 80% (see References [1] and [3]). In
order to achieve higher accuracy, typically other musical features (e.g. melody, pitch etc) need
to be used in filters [Reference 4]. At any case with or without complex filters, automatic music
classification is a very useful technique for incorporating into media-art projects and systematic
study of music.

Unfortunately, tempo-based classification has mostly been applied to Ballroom dancing music.
Argentine tango is a dancing musical genre with origins in Buenos Aires at the end of the19th
century, which quickly spread over the world and is currently danced socially in many places all
over the world. Argentine tango is completely distinct compared to its Ballroom peer, both in
terms of its tempo distribution and its feeling. What is more, the music which is played in social
settings consists of three types, the tango, the milonga and the vals. Yet although the meter
corresponding to each type (4/4, 2/4 and 3/4 in respect) is applied uniformly, the tempi tend to
be influenced by trends at the decade of composition and tastes of the composers. Because
Argentine tango has been overlooked by previous works on dancing music, and because of its
wide popularity and interesting characteristics, in this work we focus on the study of this
particular genre.

Our main contribution with this report is the creation and evaluation of an automatic bayes
classifier for Argentine tango music, which we construct using a library with over 2,000 full (not
sample) tango pieces. Besides, we are able to prove our original point that tempo-based
classification is successful with Argentine tango music, based on the fact that our classifier has
an accuracy of 79%, comparing well with similar approaches used with Ballroom music. Yet, our
analysis suggests that more features will need to be used in order to improve further a
classifierʼs accuracy. The rest of this report is structured as follows: in the section “Related
Work”, we present the tools and techniques related to our work, in the section “Methodology” we
describe our experimental methodology, including the details of the bayes filter and our library,
in the section “Results and Discussion” we present our results and some discussion, whereas in
the section “Conclusion” we conclude this report.

Related work

Automatic music classification has been a topic of broad research interest. It can be roughly
characterized based on the level of the features used, i.e. low-level, such as tempo, or high-
level, such as timbre, melody, instrumentation etc. Although low-level features are easy to
extract using a beat-extracting tool, and extremely useful for automatic characterization of
music, the level of accuracy of low-level classifiers typically gets saturated below 80%.
Techniques using more complex rhythmic patters, as an alternative to the strong-beat tempo,
have been used in order to improve a classifiersʼ accuracy. One such technique, for example,
uses periodicity patterns to classify ballroom music, and achieves accuracy of up to ~80% (see
References [2] and [3]). High-level features on the other hand, are harder to extract and
process, yet they can improve the accuracy of the prediction significantly, and especially when
used in combination with low-level features (as in Reference [4]). In our work, since according to
our knowledge this is the first work to explore the genre of traditional Argentine tango, and since
the scope of this report is limited, we only experiment with tempo as a feature.

Various tools have been created to assist with the extraction of both low and high level features
(as in References [5] and [6]). Most tools offer a graphical interface to communicate the
dynamic variation of the extracted features. Because we focus on tempo-based classification,
we choose a tool which has been widely used for projects requiring tempo extraction, called
Beatroot (see Reference [5]). Besides a graphical interface, Beatroot also provides with a batch
processing mode, which is useful for the massive tempo extraction over our large library, which
is required in our experiments.

Methodology

Our methodology consists of the following steps:

1) Database preparation:

First of all, all pieces in our library are labeled according to their genre. Had this not been the
case, we will have had to apply the labels manually. Second, we transform the format of the
pieces in the library (mp3) into a format which is readable by Beatroot, e.g. wav. For this
operation, we use a format converting tool which is part of iTunes.

2) Tempo extraction using Beatroot:

In the batch processing mode, Beatroot spits the interarrival beats of all strong beats that it can
detect into a text file. We run all pieces in our database through Beatroot to get a text file for
each song. Then, for each piece, we calculate its average beat using the text file.

3) Creation of the tempo distribution for each genre:

We calculate the tempo distributions for each of the tango, milonga and vals types, by
calculating both the average tempi and the variations. The results of this step are summarized in
Table 1:

Table 1: Tempo Distributions for the Three Tango Types.

Tango Music Type Average Tempo [BPM] Variation
[BPM^2]

Tango Detected: 125/ Actual: 63 108

Milonga 129 1016

Vals 142 1192

Note that tango has a detected average tempo of 125 BPM (Beats Per Minute). Yet, its actual
average tempo is half the detected one, i.e. 63 BPM. This is because Beatroot detects all strong
beats in a meter, and because in the case of tango, each meter has two strong beats, one on
the first beat of the meter and one on the third one. Another interesting observation which we
can make from this table is that tango has a very tight tempo distribution, i.e. the variation is
only 108 BPM^2, while both Milonga and Vals have a 10x wider distribution, their variations
being 1016 and 1192 BPM^2 in respect.

4) Construction of the naive bayes classifer:

The posterior probabilities according to the bayes formula for the three tango music types can
be expressed as:
- P(tango | tempo) = P(tango) * P(tempo | tango) / Evidence
- P(milonga | tempo) = P(milonga) * P(tempo | milonga) / Evidence
- P(vals | tempo) = P(vals) * P(tempo | vals) / Evidence

where:

- P(tango), P(milonga) and P(vals) are the prior probabilities, based on the number of pieces
belonging to a type over the total number of pieces in the library

and where:

- P(tempo | tango), P(tempo | milonga) and P(tempo | vals) are the likelihood probabilities, which
are calculated using the tempo distributions

and where:

- Evidence = P(tango)*P(tempo | tango) + P(milonga)*P(tempo | milonga) +

! ! ! ! ! ! ! ! P(vals)*P(tempo | vals)

5) Evaluate accuracy of classifier

In order to evaluate the accuracy of the classifier, we run the whole library through the classifier,
and evaluate the classifierʼs prediction. The number of successes constitutes the accuracy of
our classifier.

Results and Discussion

The details of our evaluation are summarized in Table 2 below:

Table 2: Classifier Accuracy per Type.

Tango Music N of samples Likelihood Classifier
Type Probabilities Success

Tango 1,797 0.76 93.6%

Milonga 255 0.11 7.5%

Vals 306 0.13 49.8%

Total: 2,358 Weighted Average:

79%
Simple Average:
50%
Our results indicate that tango is the most successfully predicted type with 93.6% accuracy,
followed by vals with 49.8% accuracy, and milonga with 7.5% accuracy. Interestingly, the
accuracy of our simulator diverged significantly across the three types. There are several
reasons how to explain this. First of all, tango has a much higher probability of likelihood, and a
much tighter tempo distribution compared to the other two types. These facts obviously gave a
lot of confidence to the classifier regarding tango. Milonga on the other hand, has a very low
likelihood probability and a 10x greater variance compared to tango. These facts alone reduce
the confidence of the classifier regarding the detection of milonga. To make matters worse, the
actual milonga tempo is very close to the detected tango tempo (129 BPM vs. 125 BPM), which
combined with the high confidence of the classifier in regards to tango, lead to the very high
missprediction rate for milonga. Lastly, the classifierʼs accuracy for vals is 49.8%. This is
because vals has quite a distinct tempo compared to milonga and tango, therefore it is possible
for the classifier to make more confident predictions regarding vals compared to milonga, yet
the accuracy is only modest because of its low likelihood probability and fairly large variance.

A straightforward solution in order to improve the accuracy of the simulator is to adjust Beatroot
such that it detects the beat corresponding to the tempo, instead of any strong beat within the
meter. One way to do so is by adding hints from the melody, for example a melodic phrase
typically starts at the beginning of the meter and finishes at its end, and so the classifier could
mark these beginning and end points in order to try to figure out the tempo. Using such hints,
the prediction accuracy of all types would improve.

Another idea which can be used in order to improve the classifierʼs accuracy is to include
features related to the era of composition. Most Argentine tango pieces were composed in
between the late 1910ʼs and 1960ʼs, where the pieces composed at earlier decades, and
specifically the pieces of milonga and vals types, have a slower tempo compared to the ones
composed in the later decades. We could therefore create various filters based on the tempo
distributions per decade.

Yet Argentine tango music is mostly influenced perhaps by specific composersʼ tastes. Although
most composers who created what we perceive as classical Argentine tango music followed
specific rules, such as the same meter, mostly similar tempo and similar dramatic, each of them
incorporated their personal influences into their work. For example, Biagiʼs compositions are
influenced by his affinity to jazz, which resulted in the tempo not always falling exactly on the
beat (Reference [7]). DʼArienzo compositions, on the other hand, fall precisely on the beat (he
was called the “King of Beat”), and his compositions are complexly enorchestrated (Reference
[8]). For a finer grain classification therefore, we could use musical features specific to
composers.

Conclusion

In this report, we proved our initial intuition that tempo-based classification is a promising
technique for the classification of the unexplored yet very popular Argentine tango music. Yet,
because the composition of this music evolved significantly over the decades and across
various composers, we observed that there is generally a lot more room to improve the
accuracy of classification.
References

[1]: Fabien Goyon and Simon Dixon. Dance Music Classification: A tempo-based approach. In
Proceedings of the 5th International Conference on Music Information Retrieval (ISMIR'04),
2004.
[2]: Elaine Chew, Anja Volk, and Chia-Ying Lee. Dance Music Classification Using Inner Metric
Analysis. Book chapter in The Next Wave in Computing, Optimization and Decision
Technologies, Volume 29, VI, pages 355 - 370, 2005. Springer.
[3]: Simon Dixon, Fabien Gouyon, and Gerhard Widmer. Towards Characterization of Music via
Rhythmic Patterns. In Proceedings of the 5th International Conference on Music Information
Retrieval (ISMIR'04), pages 509 - 516, 2004.
[4]: Core McKay, and I. Fujinaga. Automatic Music Classification and the Importance of
instrument identification. In Proceedings of the Conference on Interdisciplinary Musicology
(CIM05), 2005.
[5]: Beatroot webpage: http://www.eecs.qmul.ac.uk/~simond/beatroot/
[6]: Marsyas webpage: http://marsyas.info/community/ideas
[7]: Reference for Biagi: http://www.todotango.com/english/creadores/rbiagi.asp
[8]: Reference for DʼArienzo: http://www.todotango.com/english/creadores/jdarienzo.asp
[9]: Argentine tango reference: http://en.wikipedia.org/wiki/Argentine_tango
[10]: Photo courtesy of Evelyn Lu. Website: http://luluaperture.smugmug.com/