MARSYAS-0.2: A Case Study in Implementing Music Information Retrieval Systems - Intelligent Music Information Systems: Tools and Methodologies - page 36

Figure 1. Audio feature extraction, segmentation and classification

in 1 minute. Whether programming for a week is
worth the trouble is a tough question. Even though
it is impossible to provide a definite answer, we
discuss how our experiences with MARSYAS
can be informative. If the task at hand can take
hours, as MIR algorithms frequently do, runtime
performance becomes a critical issue that can
affect research. For example, if an experiment
that used to take 5 hours takes 10 minutes, it
can be executed multiple times with different
parameters to find the best choice. Programming
is a very time consuming task so choosing to
build your own tools works best if you have a
large enough timeframe to do it. For example a
PhD student has a better chance at completing a
significant software framework than a Masters
student pressed for time. A supportive advisor is
also important and the best way to achieve this is
to provide earlier proof that the time you spend
developing software pays off in terms of research
results. Finally, an important consideration is that
the development of the software framework itself
especially in emerging applications such as MIR
is research in itself. MARSYAS has been used as

a test bed for many ideas in Software Engineer-
ing and Computer Science that arise based on
the particular characteristics and constraints of
audio processing.
MARSYAS can be obtained from http://
marsyas.sourceforge.net . It is written in portable
C++ (as much as possible) and it compiles in Linux,
OS X, Cygwin and Windows Visual Studio 2003
and 2005. Subversion is used for version control
and the latest unstable source code can be obtained
from the Web page.

Requirements

The canonical application of MARSYAS is audio
feature extraction (Tzanetakis & Cook, 2002)
which forms the basis of many MIR algorithms
such as classification, segmentation and similarity
retrieval. Figure 1 shows a schematic diagram of
audio feature extraction and how it can be used
for segmentation and classification. The audio
signal is broken into small slices and by per-
forming some form of frequency analysis such
as the Discrete Fourier Transform followed by