Information Technology Reference
In-Depth Information
vance the public resource-computing paradigm:
to encourage the creation of many projects, and to
encourage a large fraction of the world's computer
owners to participate. BOINC has been used in
a number of projects, ranging from searching
for extraterrestrial intelligence to searching for
gravitational waves. The DART project described
here builds on BOINC by extending it by two
ways: firstly, it provides a decentralized overlay
for accessing information across the system; and
secondly, it reverses the role of the participants. In
BOINC the participants provide CPU cycles for
the analysis of external data (downloaded from a
central point), whereas, in DART, the participants
provide MIR metadata information to the network
by analysing their own data that is, audio that is
stored on their hard drives.
The analysis component of MIR requires
extensive computational resources. Distributed
environments and P2P networks are already be-
ing used for this purpose (Tzanetakis, Gao, &
Steenkiste, 2005). The idea of using MIR over
P2P was proposed in Wang, Li, and Shi (2002),
however this system suffered from problems with
scalability. More recently, the JXTA programming
framework was used by Baumann (2003) to aid in
the content-based retrieval over a P2P network.
The proposed DART system differs from the
distributed MIR system proposed in Tzanetakis,
Gao, and Steenkiste (2005) however, in that only
metadata is returned to the main Triana server
for analysis, as opposed to actual audio data
files, and has a different overall goal; DART is
not intended to act as a file sharing system, but
instead a distributed P2P MIR system with the
main application scenario focussing on the rec-
ommendation of music based on the audio files
the users hard drive.
Pandora (http://www.pandora.com/) is a novel
music recommendation system from the makers
of the Music Genome Project (http://www.pan-
dora.com/mgp.shtml). Pandora allows users to
enter the names of artists or songs they like, and
Pandora will consult return a play list of artists
and songs that the user may like. Again, DART
hopes to build on these concepts by using the user's
actual audio files to base the recommendations
on, and analysing a much broader range of music,
potentially across millions of users.
In summary, DART's P2P architecture aims to
build upon all of these developments to provide an
advanced, fully scalable platform for developing,
testing and deploying new search and analysis
algorithms on an Internet scale. Furthermore, as
explained later in the chapter, the DART system
can be adapted to fulfil a variety of applications
other than music recommendation by modify-
ing the Triana workflow that is distributed to the
worker nodes. The Triana framework and Triana
workflows are discussed in the next section.
related mIr studies
Audio analysis algorithms and frameworks for
Music Information Retrieval (MIR) are expanding
rapidly, providing new ways to garnish informa-
tion from audio sources well beyond what can
be ascertained from ID3 tags. Modest successes
have been made in audio-based musical genre
classification audio-analysis algorithms such as
musical genre classification (Tzanetakis & Cook
2002; Aucouturier & Pachet 2003), beat detection
and analysis (Foote & Cooper 2002), similarity
retrieval (Aucouturier & Pachet 2002; Logan &
Salomon 2001; Yang, 2002), and audio finger-
printing (Haitsma & Kalker, 2002). This work
uses Short-Time Fourier Transforms to track the
means and variances of the Spectral Centroid,
standard deviations of the spectrum around its
centroid, spectral envelopes, and signal power to
represent sound textures, beat and pitch content
(Tummarello, Morbidoni, Puliti, & Piazza, 2005).
These values are then transformed into attribute-
value pairs for pattern matching and semantic
retrieval. There is still much to be done in this
field. Refinements to existing strategies, as well
as new strategies are still needed.
Search WWH ::
Custom Search