Identifying Saxophonists from Their Playing Styles - Intelligent Music Information Systems: Tools and Methodologies - page 113

Figure 5. Prototypical Narmour structures

Figure 6. Narmour analysis of All of Me

tify a musician by his or her playing style is, how
is this task performed by a music expert? In the
case of Jazz saxophonists we believe that most of
the cues for interpreter identification come from
the timbre or “quality” of the notes performed
by the saxophonist. That is to say, while timing
information is certainly important and is useful
to identify a particular musician most of the in-
formation relevant for identifying an interpreter is
the timbre characteristics of the performed notes.
In this respect, the saxophone is similar to the
singing voice in which most of the information
relevant for identifying a singer is simply his or
her voice's timbre. Thus, the algorithm to identify
interpreters from their playing style reported in
this chapter aims to detect patterns of notes based
on their timbre content. Roughly, the algorithm
consists of generating a performance alphabet by
clustering similar (in terms of timbre) individual
notes, inducing for each interpreter a classifier
which maps a note and its musical context to a
symbol in the performance alphabet (i.e., a clus-
ter), and given an audio fragment identify the
interpreter as the one whose classifier predicts
best the performed fragment. We are ultimately
interested in obtaining a classifier MC mapping
melody fragments to particular performers (i.e.,
the identified saxophonist). We initially segment
all the recorded pieces into audio segments
representing musical phrases. Given an audio
fragment denoted by a list of notes [ N 1 ,…,N m ] and

a set of possible interpreters denoted by a list of
performers [ P 1 ,…,P n ], classifier MC identifies the
interpreter as follows:

MC ([ N 1 ,…,N m ], [ P 1 ,…,P n ])
for each interpreter P i
Score i = 0
for each note N k
PN k = perceptual_features( N k )
C N k = contextual_features( N k )
( X 1 ,…,X q ) = cluster_membership( PN k )
for each interpreter P i
Cluster(i,k) =CL i ( CN k )
Score i = Score i + X Cluster(i,k)
return P M such that Score M = max( Score 1 ,…
,Score n )

This is, for each note in the melody fragment
the classifier MC computes the set of its percep-
tual features, the set of its contextual features
and, based on the note's perceptual features, the
cluster membership of the note for each of the
clusters ( X 1 ,…,X q are the cluster membership for
clusters 1,…,q , respectively). Once this is done, for
each interpreter P i its trained classifier CL i ( PN )
predicts a cluster representing the expected type
of note the interpreter would have played in that
musical context. This prediction is based on the
note's contextual features. The score Score i for
each interpreter i is updated by taking into ac-
count the cluster membership of the predicted
cluster (i.e., the greater the cluster membership