Representation and alignment of sung queries for music information
retrieval.  Norman H. Adams and Gregory H. Wakefield (Dept. of 
Electrical Engineering and Computer Science, Univ. of Michigan, 
1101 Beal Avenue, Ann Arbor, MI 48109-2110)

The pursuit of robust and rapid query-by-humming systems, which
search melodic databases using sung queries, is a common theme in music
information retrieval.  The retrieval aspect of this database problem has
received considerable attention, whereas the front-end processing of sung
queries and the data structure to represent melodies has been based on
musical intuition and historical momentum. The present work explores three
time series representations for sung queries: a sequence of notes, a
`smooth' pitch contour, and a sequence of pitch histograms.  The
performance of the three representations is compared using a collection of
naturally sung queries. It is found that the most robust performance is
achieved by the representation with highest dimension, the smooth pitch
contour, but that this representation presents a formidable computational
burden.  For all three representations, it is necessary to align the query
and target in order to achieve robust performance.  The computational cost
of the alignment is quadratic, hence it is necessary to keep the dimension
small for rapid retrieval.  Accordingly, iterative deepening is employed to
achieve both robust performance and rapid retrieval.  Finally, the
conventional iterative framework is expanded to adapt the alignment
constraints based on previous iterations, further expediting retrieval
without degrading performance.

Using music structure to improve beat tracking. Roger B. Dannenberg
(Computer Science Dept., Carnegie Mellon Univ., Pittsburgh, PA 15213)

Beats are an important feature of most music. Beats are used in music
information retrieval systems for genre classification, similarity
search, and segmentation. However, beats can be difficult to identify,
especially in music audio. Traditional beat trackers attempt to (1)
match predicted beats to observations of likely beats, and (2) maintain
a fairly steady tempo. A third criterion can be added: when repetitions
of musical passages occur, the beats in the first repetition should
align with the beats in all other repetitions. This third criterion
improves beat tracking performance significantly. Repetitions of musical
passages are discovered in audio data by searching for similar sequences
of chroma vectors. Beats are "tracked" by first locating a sequence of
likely beats in the music audio using high frequency energy as an
indicator of beat likelihood. This beat sequence is then extended by
searching forward and backward for more matching beats, allowing slight
variations in tempo, and using a relaxation algorithm to optimize the
proposed beat locations with respect to the three criteria. Other
high-level music features may offer further improvements in beat
identification.

Using instrument recognition for melody extraction from polyphonic 
audio. Jana Eggink (Sony Deutschland GmbH, European Technology Center, 
Hedelfinger Strasse 61, 70327 Stuttgart, Germany) and Guy J. Brown
(Dept. of Computer Science, Univ. of Sheffield, Regent Court, 211 
Portobello St., Sheffield S1 4DP, U. K.)

A system is proposed that identifies the solo instrument in accompanied
sonatas and concertos, and uses this knowledge to extract the melody line
played by this instrument. The approach uses a feature representation based
solely on the spectral peaks belonging to the harmonic series of a
fundamental frequency (F0). Based on an initial approximate F0 estimation,
this representation proved to be sufficient for instrument classification
even in the presence of highly unpredictable background accompaniment. Once
the solo instrument is known, a more accurate estimation of the melody line
is carried out based on so-called melody models, which are trained on
instrument specific training material. In every time-frame multiple F0
candidates are extracted and their likelihood is evaluated according to the
chosen melody model. Additional temporal constraints take the form of
frame-to-frame transition probabilities, and are obtained from the same
training material. The two knowledge sources are combined in a statistical
search for the overall most likely melody line. When evaluated on realistic
recordings of classical sonatas and concertos, the system was able to find
the correct F0 in 72% of frames, an improvement of over 20% compared to a
simple salience based approach.

Music to Knowledge: a visual programming environment. Andreas F. Ehmann
(Dept. of Electrical and Computer Engineering, Univ. of Illinois 
Urbana-Champaign, Urbana, IL 61801) and J. Stephen Downie (Graduate 
School of Library and Information Science, Univ. of Illinois 
Urbana-Champaign, Champaign, IL 61820}

The objective of the International Music Information Retrieval Systems
Evaluation Laboratory (IMIRSEL) project is the creation of a large,
secure corpus of audio and symbolic music data accessible to the music
information retrieval (MIR) community for the testing and evaluation of
various MIR techniques. As part of the IMIRSEL project, a cross-platform
JAVA based visual programming environment called Music to Knowledge
(M2K) is being developed for a variety of music information retrieval
related tasks. The primary objective of M2K is to supply the MIR
community with a toolset that provides the ability to rapidly prototype
algorithms, as well as foster the sharing of techniques within the MIR
community through the use of a standardized set of tools. Due to the
relatively large size of audio data and the computational costs
associated with some digital signal processing and machine learning
techniques, M2K is also designed to support distributed computing across
computing clusters. In addition, facilities to allow the integration of
non-JAVA based (e.g. C/C++, MATLAB, etc.) algorithms and programs are
provided within M2K. [Work supported by the Andrew W. Mellon Foundation
and NSF Grants No. IIS-0340597 and No. IIS-0327371.]

Distributed digital music archives and libraries. Ichiro Fujinaga
(Faculty of Music, McGill Univ., Montreal, Canada)

The main goal of this research program is to develop and evaluate
practices, frameworks, and tools for the design and construction of
worldwide distributed digital music archives and libraries. Over the
last few millennia, humans have amassed an enormous amount of musical
information that is scattered around the world. It is becoming
abundantly clear that the optimal path for acquisition is to distribute
the task of digitizing the wealth of historical and cultural heritage
material that exists in analog formats, which may include books and
manuscripts related to music, music scores, photographs, videos, audio
tapes, and phonograph records. In order to achieve this goal, libraries,
museums, and archives throughout the world, large or small, need
well-researched policies, proper guidance, and efficient tools to
digitize their collections and to make them available economically. The
research conducted within the program addresses unique and imminent
challenges posed by the digitization and dissemination of music media.
There are four major research projects in progress: development and
evaluation of digitization methods for preservation of analog
recordings; optical music recognition using microfilms; design of
workflow management system with automatic metadata extraction; and
formulation of interlibrary communication strategies.

Speech-recognition interfaces for music information retrieval.
Masataka Goto (National Institute of Advanced Industrial Science and 
Technology (AIST), 1-1-1 Umezono, Tsukuba, Ibaraki 305-8568, JAPAN, 
m.goto@aist.go.jp)

This paper describes two hands-free music information retrieval (MIR)
systems that enable a user to retrieve and play back a musical piece by
saying its title or the artist's name. Although various interfaces for
MIR have been proposed, speech-recognition interfaces suitable for
retrieving musical pieces have not been studied. Our MIR-based jukebox
systems employ two different speech-recognition interfaces for MIR,
speech completion and speech spotter, which exploit intentionally
controlled nonverbal speech information in original ways. The first is a
music retrieval system with the speech-completion interface that is
suitable for music stores and car-driving situations. When a user only
remembers part of the name of a musical piece or an artist and utters
only a remembered fragment, the system helps the user recall and enter
the name by completing the fragment. The second is a background-music
playback system with the speech-spotter interface that can enrich
human-human conversation. When a user is talking to another person, the
system allows the user to enter voice commands for music playback
control by spotting a special voice-command utterance in face-to-face or
telephone conversations. Experimental results from use of these systems
have demonstrated the effectiveness of the speech-completion and
speech-spotter interfaces. (Video clips:
http://staff.aist.go.jp/m.goto/MIR/speech_if.html)

MPEG-7 - standardized tools for music information retrieval. 
Juergen Herre (Audio & Multimedia Departments, Fraunhofer Institute for 
Integrated Circiuts (IIS), Am Wolfsmantel 33, 91058 Erlangen, Germany)

Today, many applications in Music Information Retrieval (MIR) employ
audio features which have been tailored individually by the algorithm
developers. For a broader use also in commercial applications, MIR
technology can benefit significantly from a "common languag" in audio
signal description that can be used to annotate any type of multimedia
assets in order to facilitate search & retrieval according to a wide
range of conceivable criteria in an interoperable way. The audio part of
the ISO/MPEG-7 "Multimedia Content Description Interface" provides
such a common signal description language by defining a rather
comprehensive set of standardized features (called "Low Level
Descriptors", LLDs), application-centric subsets, and a unified way of
exchanging this data based on XML. The talk provides an overview of the
MPEG-7 Audio tool chest, including existing and forthcoming extensions.
While the idea is clearly to create a universal platform for any
conceivable MIR task, some of the initially conceived applications of
MPEG-7 Audio are illustrated.

Automatic detection of the dominant melody in acoustic musical signals.
Anssi P. Klapuri (Inst. of Signal Processing, Tampere Univ. of 
Technology, Korkeakoulunkatu 1, 33720 Tampere, Finland}

An auditory-model based method is described for estimating the
fundamental frequency contour of the dominant melody in complex music
signals. The core method consists of a conventional cochlear model
followed by a novel periodicity analysis mechanism within the subbands.
As the output, the method computes the salience (i.e., strength) of
different fundamental frequency candidates in successive time frames.
The maximum value of this vector in each frame can be used to indicate
the dominant fundamental frequency directly. In addition, however, it
was noted that the first-order time differential of the salience vector
leads to an efficient use of temporal features which improve the
performance in the presence of a large number of concurrent sounds.
These temporal features include particularly the common amplitude or
frequency modulation of the partials of the sound that is used to
communicate the melody. A noise-suppression mechanism is described which
improves the robustness of estimation in the presence of drums and
percussive instruments. In evaluations, a database of complex music
signals was used where the melody was manually annotated. Use of the
method for music information retrieval and music summarization is
discussed.

Using string alignment in a query-by-humming system for real world 
applications. Christian Sailer (Fraunhofer IDMT, Langewiesener Str. 
22, 98693 Ilmenau, Germany)

Though Query by Humming (i.e. retrieving music or information about
music by singing a characteristic melody) has been a popular research
topic during the last decade, few approaches have reached a level of
usefulness beyond mere scientific interest. One of the main problems is
the inherent contradiction between error tolerance and dicriminative
power in conventional melody matching algorithms that rely on a melody
contour approach to handle intonation or transcription errors. Adopting
the string matching / alignment techniques from bioinformatics to melody
sequences allows to directly assess the similarity between two melodies.
This method takes an MPEG-7 compliant melody sequence (i.e. a list of
note invervals and length ratios) as query and evaluates the steps
necessary to transform it into the reference sequence. By introducing a
musically founded cost-of-replace function and an adequate post
processing, this method yields a measure for melodic similarity. Thus,
it is possible to construct a query by humming system that can properly
discriminate between thousands of melodies and still be sufficiently
error tolerant to be used by untrained singers. The robustness has been
verified in extensive tests and real world applications.