Identifying Saxophonists from Their Playing Styles - Intelligent Music Information Systems: Tools and Methodologies - page 104

audio recordings. The identification of performers
by using the expressive content in their perfor-
mances raises particularly interesting questions
but has nevertheless received relatively little
attention in the past. Given the capabilities of
current audio analysis systems, we believe expres-
sive-content-based performer identification is a
promising research topic in music information
retrieval. After presenting the background to this
area and briefly discussing the limitations of this
approach to performer identification, we present
an algorithm for identifying Jazz saxophonists
using high-level semantic information obtained
from real performances. This work is based on
our previous work on expressive performance
modeling (Ramirez & Hazan , 2005; Ramirez,
Hazan, Maestre, & Serra , 2006). Finally, we
discuss the results from the case study and draw
some conclusions.
The data used in our investigations are audio
recordings of real performances by Jazz saxophon-
ists. The use of audio recordings, as opposed to
MIDI recordings where data analysis is simplified,
poses substantial difficulties for the extraction of
music performance information. However, the
obvious benefits of using real audio recordings
widely compensate the extra effort required for
the audio analysis. We use sound analysis tech-
niques based on spectral models (Serra & Smith,
1990) for extracting high-level symbolic features
from the recordings. The spectral model analysis
techniques are based on decomposing the origi-
nal signal into sinusoids plus a spectral residual.
From the sinusoids of a monophonic signal it is
possible to extract high-level semantic informa-
tion such as note pitch, onset, duration, attack and
energy among other information. In particular,
for characterizing expressivity in saxophone we
are interested in two types of features: intranote
or perceptual features representing perceptual
characteristics of the performance, and internote
or contextual features representing information
about the music context in which expressive
events occur. We use the software SMSTools

(SMS) which is an ideal tool for preprocessing the
signal and providing a high-level description of
the audio recordings. Once the relevant high-level
information is extracted we apply machine-learn-
ing techniques (Mitchell, 1997) to automatically
discover regularities and expressive patterns for
each performer. We use these regularities and
patterns in order to identify a particular performer
in a given audio recording. We discuss different
machine learning techniques for detecting the
performer's expressive patterns, as well as the
perspectives of using sound analysis techniques
on arbitrary polyphonic audio recordings.
The rest of the chapter is organized as follows:
Section 2 sets the background for the research
reported here. Section 3 describes how we pro-
cess the audio recordings in order to extract both
perceptual and contextual information. Section 4
presents our algorithm for identifying performers
from their playing styles, as well as some results.
Section 5 briefly discusses future trends in the
context of this research, and finally, Section 6
presents some conclusions and indicates some
areas of future research.

Background

Music performance plays a central role in our
musical culture today. Concert attendance and
recording sales often reflect people's preferences
for particular interpreters. The manipulation of
sound properties such as pitch, timing, amplitude
and timbre by different performers is clearly dis-
tinguishable by the listeners. Expressive music
performance studies the manipulation of these
sound properties in an attempt to understand why ,
what , how and when expression is introduced to a
performance. There has been much speculation as
to why performances contain expression. Hypoth-
esis include that musical expression communicates
emotions (Justin, 2001) and that it clarifies musical
structure (Kendall, 1990), that is, the performer
shapes the music according to her own inten-