Dissimilarity and the Classification of Female Singing

Voices: A Preliminary Study

Molly L. Erickson
Knoxville, Tennessee

Summary: Traditionally, timbre has been defined as that perceptual attribute

that differentiates two sounds when pitch and loudness are equal and thus is
a measure of dissimilarity. By such a definition, each voice possesses a set
of timbres, and the identity of any voice or voice category across different pitch–
loudness–vowel combinations must be due to an abstraction of the pattern of
timbre transformation. Using stimuli produced across the singing range by
singers from different voice categories, this study sought to examine how
timbre and pitch interact in the perception of dissimilarity. This study also
investigated whether listener experience affects the perception of timbre as
a function of pitch. The resulting multidimensional scaling (MDS) repre-
sentations showed that for all stimuli and listeners, dimension 1 correlated with
pitch, whereas dimension 2 correlated with spectral centroid and separated
vocal stimuli into the categories mezzo-soprano and soprano. Dimension 3
appeared highly idiosyncratic depending on the nature of the stimuli and on
the experience of the listener. Inexperienced listeners appeared to rely more
heavily on pitch in making dissimilarity judgments than did experienced
listeners. The resulting MDS representations of dissimilarity across pitch
provide a glimpse of the timbre transformation of voice categories across pitch.
Key Words: Voice classification—Perception—Timbre—Pitch—Listener
experience.

INTRODUCTION

Accepted for publication October 17, 2002.

This paper was presented at the 28th Annual Symposium:
Care of the Professional Voice, June 1999, Philadelphia, Penn-
sylvania.
From the Department of Audiology and Speech Pathology,
University of Tennessee, Knoxville, Tennessee.
Address correspondence and reprint requests to Molly Erick-
son, Department of Audiology and Speech Pathology, 457 South
Stadium Hall, University of Tennessee, Knoxville, TN 37996.
E-mail: merickso@utk.edu
Journal of Voice, Vol. 17, No. 2, pp. 195–206
쑕 2003 The Voice Foundation
0892-1997/2003 $30.00⫹0
doi: 10.1016/S0892-1997(03)00022-5
The accepted definition of timbre is as follows:
Two tones are of different timbre if they are judged
to be dissimilar and yet have the same loudness
and pitch.1 By this definition, when pitch and loud-
ness are equal, timbre becomes the primary auditory
cue used to identify differences and thus may be
used to compare sound-producing objects. Tradi-
tional timbre research has been guided by the notion
of timbre differences embodied in this technical
definition, and it has concentrated on the acoustic
factors that underlie the perception of dissimilarity
when pitch and loudness are constant or nearly so.2–5

195
196 MOLLY L. ERICKSON
Many sound-producing objects are capable of
large variations in pitch and loudness. In spite of this
great variation, not only can listeners distinguish
between sound-producing objects, but listeners can
also identify sounds of varying pitch and loudness as
having been produced by the same sound-producing
object. Thus, it could be argued that some element
of timbre also serves the important function of main-
taining the coherence of sound-producing objects
in an ever-changing auditory world. From this per-
spective, every sound source can be thought to have
a timbre transformation, a set of timbres across dif-
ferent pitches, loudness levels, vowels, and so on.
It is the listener’s task to abstract the timbre transfor-
mation, creating coherence of the source in the face
of timbre variation across conditions.
In the vocal pedagogy literature, the term timbre
is often used in a manner contrary to the technical
definition. It is generally stated that along with pitch
range and tessitura, timbre is a major factor in the
determination of a singer’s voice category.6
Cleveland7 states that an individual singer has a
characteristic timbre that is a function of the laryn-
geal source and vocal tract resonances. The pedagog-
ical use of the term timbre implies that it is a property
of the singer, not of each note sung by the singer.
This interpretation of the term timbre as used by
vocal pedagogues has historically been reflected
in vocal timbre research where timbre is assumed to
be constant throughout a singer’s range, or at least
throughout each of a singer’s registers.7,8 Given the
technical definition of timbre, it would be more accu-
rate to refer to a singer as having a set of timbres, or
a timbre transformation, rather than simply as having
a timbre. By this logic, singers with similar timbre
transformations constitute members of the same
voice timbre type or voice category. Given the possi-
ble changes in timbre across the singing range, it is
possible that any two voices may be perceived
as having similar timbre on one pitch–loudness–
vowel combination and dissimilar timbre on another.
Thus, the categories must be elastic and may not be
reducible to one unchanging acoustic metric.
To summarize this argument, one could hypothe-
size that each voice possesses a set of timbres, and
the identity of any voice across different pitch–
loudness–vowel combinations must be due to an
Journal of Voice, Vol. 17, No. 2, 2003
abstraction of the pattern of timbre transforma-
tion. Singers in the same category would be charac-
terized by similar transformations, whereas singers
in different categories would be differentiated by
dissimilar timbre transformations, not necessarily
by dissimilarity at specific pitches. This conceptual-
ization allows us to bridge the technical definition
of timbre as a difference between sources to the
conceptualization of the timbre transformation as
the “signature” of a source.
This experiment represents a preliminary attempt
to test the above hypotheses. Previously, the percep-
tion of voice classification has been examined using
forced-choice paradigms based on the traditional
voice classification system of bass, baritone, tenor,
alto, mezzo-soprano, and soprano.7 Such perceptual
experiments provide information as to how listeners
place vocal stimuli when provided with arbitrary
classification categories. They do not provide infor-
mation concerning how timbre is transformed across
pitch within categories. The question that begs an-
swering is this, if provided no classification system
a priori, how do listeners tend to group vocal stimuli
across pitch? Do they group vocal stimuli in a
manner that supports current classification systems
based on the presumption that vocal timbre is an
invariant perceptual cue across the singing range,
or do they group them in a manner that suggests that
vocal timbre undergoes a systematic transformation
across the singing range, with singers in the same
vocal category displaying similar timbre transforma-
tions? Additionally, this experiment tested whether
listeners with experience in the classification of sing-
ers differ from untrained listeners in the way they
organize vocal stimuli? Previous studies of timbre
at one pitch have not revealed differences between
trained and untrained listeners.2,9 The question re-
mains, does experience affect the perception of
the timbre transformation across pitch?
To answer these questions, this study used a
dissimilarity paradigm in which listeners judged
the timbre similarity between singers at different
pitches. If there is a timbre transformation shared
by singers of the same category, then the dissimilar-
ity judgments between different notes for the same
singer ought to follow a pattern that is similar to
other singers in the same category, but different from
singers in a different category. If listener experience
 CLASSIFICATION OF FEMALE SINGING VOICES
affects the perception of timbre transformations, then
the dissimilarity judgments should differ between
experienced and inexperienced listeners.
METHOD
Stimuli
Master’s level singers from the School of Music
at the University of Tennessee, Knoxville, provided
stimuli for the experiment. All singers were mem-
bers of the University of Tennessee Opera Studio
program, a training program jointly operated by the
University of Tennessee and the Knoxville Opera
Company. All subjects provided informed consent
using a procedure that was previously approved
by the Institutional Review Board of the Univers-
ity of Tennessee, Knoxville. These subjects met
the following criteria: (1) bilateral hearing within
normal limits as determined by a 20-dB hearing
screening at 500 Hz, 1000 Hz, 2000 Hz, and 4000
Hz10; (2) voice study at the Master’s degree level
or higher; and (3) no voice problems at the time of
taping as determined by a certified speech-language
pathologist. Additionally, all subjects must have
been consistently categorized as a soprano or mezzo-
soprano for more than five years. Such categoriza-
tion must also have been agreed on by 100% of
the voice professionals associated with the Knoxville
Opera Company and the University of Tennessee
School of Music.
Two singers from each voice classification,
mezzo-soprano and soprano, were recorded singing
the vowel /ɑ/ on six pitches, A3, C4, G4, B4, F5,
and A5, at a comfortable loudness level. Each singer
produced a sustained /ɑ/ for approximately four
seconds. Both sopranos were characterized as lyric
sopranos, whereas one mezzo-soprano was char-
acterized as a lyric mezzo-soprano and the other as
a dramatic mezzo-soprano.
Recordings were made in a single-walled sound
booth (Acoustic Systems RE-144-S, Austin, TX).
Subjects were recorded using a digital audio tape
recorder (Sony PCMR500, Park Ridge, NJ) and a
Sennheiser MD 441-U (Old Lyme, CT) microphone.
Subjects stood in the center of the booth. Lip to
microphone distance was 12 inches. A keyboard was
used to present pitches. Prior to taping, subjects were
197
allowed to vocalize freely and become comfort-
able with the recording environment.
One-second digital samples were constructed for
each sung vowel. Each stimulus was low-pass fil-
tered at 20 kHz using a Tucker-Davis-Technologies
(TDT, Gainesville, FL) FT6 anti-aliasing filter, and
then digitized at 48 kHz using a Turtle Beach Pinna-
cle (Yonkers, NY) sound card. The software program
Cool Edit Pro (Syntrillium Software Corporation,
Phoenix, AZ) was used to extract one-second sam-
ples from the midpoint of the sung /ɑ/. This was
accomplished by measuring the exact duration of
the sung vowel, calculating its midpoint, and then
extracting a one-second segment that originated one
half-second before the midpoint and terminated
one half-second after the midpoint. Spline curve
amplitude shaping functions were applied to each
sample to provide ramped onsets and offsets. Be-
cause equal loudness contours in the frequency
range of these stimuli are relatively flat, the overall
amplitude of each stimulus was adjusted so that
all were of approximately equal amplitude (⫾3.5-
dB RMS). The small variations in dB RMS were
randomly distributed throughout the stimuli so that
any perceivable variations in loudness likely did
not systematically affect the outcomes.
Listeners
All listeners provided informed consent using a
procedure previously approved by the Institutional
Review Board of the University of Tennessee, Knox-
ville. Two groups of listeners were recruited for
the dissimilarity experiment: inexperienced listeners
and experienced listeners.
Inexperienced listeners were recruited from stu-
dents enrolled in introductory psychology courses
at the University of Tennessee, Knoxville. Eighteen
inexperienced listeners were recruited that met
the following criteria: (1) bilateral hearing within
normal limits as determined by a 20-dB hearing
screening at 500 Hz, 1000 Hz, 2000 Hz, and 4000
Hz10; (2) no history of choral singing or vocal train-
ing; and (3) no interest in classical vocal music
or opera.
Experienced listeners were recruited from the
Knoxville Choral Society and the Knoxville Opera
Company. Twelve listeners were recruited that met
the following criteria: (1) bilateral hearing within
Journal of Voice, Vol. 17, No. 2, 2003
198 MOLLY L. ERICKSON
normal limits as determined by a 20-dB hearing
screening at 500 Hz, 1000 Hz, 2000 Hz, and 4000
Hz10; and (2) bachelor’s degree or higher in a vocal
arts related discipline (e.g., pedagogy, performance,
or choral conducting) or five years professional ex-
perience in a vocal arts discipline.
Procedure
The 24 vocal stimuli (6 pitches × 2 classifica-
tions × 2 singers) were combined in all possible
pairs, for a total of 276 paired stimuli. Within the 276
pairs, there were three types of timbre comparisons.
In the first, participants judged the similarity be-
tween two different pitches produced by the same
voice. There were 15 comparisons for the 6 pitches
produced by each singer with 60 judgments in total.
In the second, participants judged the similarity be-
tween stimuli of the same pitch for two different
singers. There were six pitches produced by all
four singers, resulting in a total of 36 judgments. In
the third, participants judged the similarity be-
tween stimuli of different pitches produced by differ-
ent singers, for 180 judgments in total. On every
trial, each of the paired stimuli was presented once,
separated by 500 ms of silence. For each pairing
of stimuli, A and B, order of pairing was randomized
across listeners, resulting in nearly equal numbers
of instances when stimulus A was followed by stim-
ulus B and when stimulus B was followed by stimu-
lus A, thus eliminating ordering effects.
The 276 paired stimuli were presented to the two
groups of listeners, inexperienced and experienced.
Prior to the listening experiment, listeners were in-
structed to rate the difference in quality of the two
sounds. Quality was defined as that part of a sound
that distinguishes one person’s voice from another’s.
Listeners rated quality differences using a dissimilar-
ity scale (1 to 10). They were instructed not to use
pitch in their judgments. Each listener then com-
pleted a practice session consisting of ten pairs of
stimuli that were not used in the experiment. The
276 trials were presented in a different random order
to each participant. Listeners were allowed to replay
each trial as many times as they needed. Listener
judgments were recorded via a computer interface
consisting of a monitor and a mouse.
To verify the perceived vocal categories of each
of the singers, upon completing the dissimilarity
Journal of Voice, Vol. 17, No. 2, 2003
experiment, all experienced listeners heard each of
the 24 vocal stimuli in isolation and were asked to
rate them on a visual analog scale (0 to 100) using
a scroll bar where the left anchor was mezzo-soprano
(assigned a value of 0) and the right anchor was
soprano (assigned a value of 100).
The listening experiment took place in a single-
walled sound booth (Acoustic Systems RE-144-S,
Austin, TX). Stimuli were presented binaurally via
a Turtle Beach Montego II (Yonkers, NY) sound
card and Sennheiser HD 545 (Old Lyme, CT)
headphones. The stimuli were presented at a com-
fortable listening level, approximately 65-dB SPL.
Acoustic Measures
As discussed by Bloothooft and Plomp,2 the
acoustic signal produced by the voice can be
described using a production-oriented or perception-
oriented point of view. From the production-oriented
point of view, the vocal instrument is generally con-
ceptualized using a source-filter model.11 Assuming
such a model, vocal acoustics could ideally be sepa-
rated into those characteristics that are produced by
the source, the vibrating vocal folds, and those pro-
duced by the filter, the vocal tract. Traditionally,
singing voice authorities have postulated that timbre
properties that distinguish singers of different
voice categories are related to properties of the filter,
particularly vocal tract length,7,12 whereas timbre
properties that distinguish vocal registers,13 brassy
versus fluty quality,14 pressed versus flow phona-
tion,15,16 or breathy versus clear tone17 could be
related to characteristics of the voice source. Thus,
acoustic measures such as long-term average spectra
(LTAS) or mean formant frequency are believed to
predict vocal categories, whereas measures such as
spectral slope and harmonics-to-noise ratio are be-
lieved to be unrelated to voice category and may help
distinguish voices within categories or to distinguish
registers within a particular voice.
From the perception-oriented point of view, the
output vocal signal is treated as a whole, not the
sum of two distinct parts, source and filter. Such an
approach makes some sense when considering high-
pitched voices where (1) the true resonance frequen-
cies of the vocal tract will not likely be reflected
in the vocal output signal due to the wide spacing
 CLASSIFICATION OF FEMALE SINGING VOICES
of harmonics and (2) there may be significant source-
filter interaction.18 Using this approach, Bloothooft
and Plomp2 analyzed the vocal signal using a set of
bandpass filters. Another possible perception-ori-
ented acoustic measure might be spectral centroid,
a measure that is frequently employed in studies of
instrumental timbre, but is not often applied to stud-
ies involving the vocal instrument.
The current experiment employed a variety of
acoustic measures in an attempt to identify those
spectral cues that lead to both categorization and
identification of specific voices. Additionally, be-
cause little is known concerning the effects of tempo-
ral parameters on the perception of vocal categories
or individual voices, two measures of vibrato were
calculated. In total, four acoustic measures were
computed: amplitude of the first harmonic minus
the amplitude of the second harmonic (H1-H2),
spectral centroid from 2 to 5 kHz, mean rate of
frequency vibrato, and mean extent of frequency
vibrato. Each measure was calculated from the
middle one second of the original digitized voice
sample. All samples were digitized at 48 kHz unless
otherwise noted.
H1-H2
This measure is believed to provide a relative
indication of glottal adduction and, therefore,
source spectral slope.13,19,20 Source slope is believed
to be unrelated to voice category and may help
distinguish voices within categories. Source slope
is often associated with terms such as “sharpness,”2
“brassy” versus “fluty” quality,14 or “pressed” versus
“flow” phonation.15
Spectral Centroid from 2 to 5 kHz
Singing voice researchers have suggested that
vocal tract length is one of the primary physiological
predictors of voice category.12,14 Although all for-
mants are affected by vocal tract length, formants
1 and 2 are also greatly affected by tongue and
jaw position. Thus, it has been postulated that the
location of the upper formants, particularly formants
3 and 4, provide the perceptual cue to voice cate-
gory.12,21 However, in female voices at high pitches,
wide spacing of harmonics makes it unlikely that
these resonance peaks will be represented precisely
in the acoustic output at fundamental frequencies
199
above 350 Hz.22 In such cases, formant frequen-
cies obtained through linear predictive coding (LPC)
might have little relationship to the actual acoustic
cues present in the signal. It might be possible to
measure the frequency of upper peaks directly from
the output spectrum; unfortunately, at high pitches,
such peaks are often not present. What is needed
is a method for quantifying the shape of the spectrum
in the regions of formants 3 and 4. One such method,
the spectral centroid, has been used extensively by
instrumental timbre researchers. Thus, for high-
pitched voices, spectral centroid in the frequency
regions of 2000 to 5000 Hz might provide a better
measure of vocal tract acoustics in the regions of
F3 and F4. Spectral centroid was calculated as fol-
lows. The average spectrum was calculated from 0
Hz to 8000 Hz using a fast Fourier transform (FFT)
with a bandwidth of 50 Hz. One hundred sixty-four
data points were obtained representing the spectrum
from 2000 to 5000 Hz. Using these data points,
the spectral centroid from 2000 to 5000 Hz was
calculated after Sandell23 using the formula:
164
兺 ek fk
k⫽1
164
兺 ek
k⫽1
where e is the vector of 164 spectral amplitude data
points between 2000 and 5000 Hz and f is the vector
of 164 spectral frequency data points between 2000
and 5000 Hz. The resulting number is essentially a
measure of central tendency and would be highly
correlated with the average of F3 and F4 when strong
peaks are present in the output signal. When no peaks
are present, the measure is affected by spectral slope.
Frequency Vibrato Rate and Extent
Temporal characteristics of voice are typically
overlooked in studies of vocal timbre. Yet, the fre-
quency modulation underlying vocal vibrato likely
has perceptual consequences.2,24 Mean frequency
vibrato rate was calculated after Prame.24 Each
digitized stimulus was subjected to an autocorrela-
tion analysis via the Corr routine of Soundswell
(Hitech Development) using a fixed window of 10
ms. The resulting waveform representing frequency
as a function of time was segmented via cursors
to determine the period of each vibrato cycle. The
Journal of Voice, Vol. 17, No. 2, 2003
200 MOLLY L. ERICKSON
FIGURE 1. Experienced listeners’ ratings of voice category for each of the 24 vocal
stimuli.
inverse of each period was calculated and averaged
to produce the mean frequency vibrato rate. Fre-
quency vibrato extent was calculated by using cur-
sors to measure maximum and minimum frequencies
at each vibrato half-cycle. Using these data points,
mean vibrato extent in cents was calculated after
Prame.25
RESULTS
Analysis of the experienced listener judgments
regarding the voice category of each stimulus sug-
gests that although listeners were heavily influenced
by pitch in judging these categories, they also
used other perceptual cues to categorize each stimu-
lus (Figure 1). Stimuli were perceived as increas-
ingly soprano-like with increasing pitch; however,
at all pitches but A5, the stimuli produced by the
two sopranos were heard as more soprano-like than
those produced by the two mezzo-sopranos. These
results support the notion that all of the singers in
the study are in fact perceived to be members of
the vocal category assigned to them by the vocal
pedagogues most familiar with their voices.
To test the assumption that the MDS solution
obtained from the inexperienced listeners differed
Journal of Voice, Vol. 17, No. 2, 2003
from that obtained from the experienced listeners,
an INDSCAL analysis (SPSS 10.0, Chicago, IL)
using an ordinal distance measurement and a Euclid-
ian metric was performed where each group was
treated as a separate subject. Group weights gener-
ated by INDSCAL were used to determine the
importance of each dimension to each group of
listeners. Group weights for each group are pre-
sented in Table 1. The first dimension was slightly
more important to the inexperienced listeners than
to the experienced listeners, whereas the second di-
mension was equally important to both groups. The
third dimension was more important to experienced
listeners than to inexperienced listeners. Given these
differences in the MDS solutions for the two groups,
separate MDS ALSCAL analyses (SPSS 10.0) were
performed. These analyses used ordinal distance
measurements and Euclidian metrics. Throughout
the following results, all stress values reported were
calculated using Young’s S-stress formula 1. Also
throughout the following results, correlation alpha
levels for each data set were adjusted using a Bonfer-
roni correction (p ⬍ 0.005).
The inexperienced listeners generated an optimal
MDS solution in three dimensions (S-stress ⫽
0.133, R2 ⫽ 0.91). The experienced listeners gener-
ated an optimal MDS solution in three dimensions
 CLASSIFICATION OF FEMALE SINGING VOICES
(S-stress ⫽ 0.201, R2 ⫽ 0.83). The three-dimen-
sional solution for both groups is presented graphi-
cally in Figure 2.
Inexperienced Listeners
Results of the INDSCAL analysis indicate that
dimensions 1 and 2 were of primary perceptual im-
portance to the inexperienced listeners (Figure 3A).
Dimension 1 appeared most closely associated with
pitch. This dimension was monotonically related
to pitch for all four voices, and pitch, as measured in
semitones, was highly correlated with dimension 1
(R ⫽ ⫺0.99, p ⬍ 0.001). Dimension 2 separated the
two voice categories, soprano and mezzo-soprano,
at all pitches but the highest pitch, A5. Throughout
the entire pitch range, dimension 2 scale values were
most highly correlated with the acoustic measure of
spectral centroid from 2 to 5 kHz (R ⫽ 0.660,
p ⬍ 0.001). Visual comparison of dimension 2
values (Figure 3A) and spectral centroid values
(Figure 3B) reveals relatively isomorphic curves in
the pitch range A3 through F5. In this pitch range, the
correlation between dimension 2 and pitch was
quite high (R ⫽ 0.758, p ⬍ 0.001). Dimension 3
(Figure 3C), although contributing to an improve-
ment in overall model fit, did so less for the inexperi-
enced listeners than for the experienced listeners.
None of the acoustic measures calculated in this study
correlated significantly with dimension 3. Examina-
tion of dimension 3 scale values as a function of
pitch shows that this dimension failed to discrimi-
nate categories of voices or individual voices,
throughout the pitch range. However, the dimension
did separate sopranos from mezzo-sopranos at the
lowest pitch, A3. The small variations in dB RMS
seen throughout the stimuli did not correlate signifi-
cantly with any dimension.
Experienced Listeners
Results of the INDSCAL analysis indicated that di-
mensions 1, 2, and 3 were of perceptual importance
to the experienced listeners. Plots of dimension 1
versus dimension 2 and dimension 1 versus dimen-
sion 3 are shown in Figure 4A and C, respectively.
As in the MDS analysis for inexperienced listen-
ers, dimension 1 appeared most closely associated
with pitch. Here, there was a monotonic relationship
201
TABLE 1. Group Weights Generated by INDSCAL
for Inexperienced and Experienced Listeners
Inexperienced Experienced
Dimension listeners listeners
1 0.7507 0.6260
2 0.5628 0.5438
3 0.1541 0.3273
between dimension 1 and pitch throughout the entire
pitch range for lyric soprano 1. For the remaining
three voices, a monotonic relationship between di-
mension 1 and pitch was seen for the pitch range
A3 through F5. Pitch, as measured in semitones,
was highly correlated with dimension 1 (R ⫽
⫺0.874, p ⬍ 0.001).
Again, dimension 2 separated the two voice cate-
gories, soprano and mezzo-soprano (Figure 4A), at
all pitches but the highest pitch, A5. Similar to the
data for the inexperienced listeners, the acoustic
measure of spectral centroid from 2 to 5 kHz was most
highly correlated with dimension 2 throughout the
pitch range (R ⫽ 0.687, p ⬍ 0.001) (Figure 4B).
This correlation was even stronger when the pitch
range was restricted to A3 through F5 (R ⫽
0.858, p ⬍ 0.001).
Dimension 3 appeared to separate individual sing-
ers within voice category (Figure 4C). For these
data, the strongest acoustic correlate to dimension
3 was vibrato rate (R = 0.651, p = 0.001). Visual
comparison of dimension 3 (Figure 4C) with vibrato
rate (Figure 4D) reveals that within each voice cate-
gory, the singer with the higher vibrato rate was also
the singer with the more positive scale values along
dimension 3. The small variations in dB RMS seen
throughout the stimuli did not correlate significantly
with any dimension.
DISCUSSION
This paper tested the hypothesis that rather than
having a characteristic “timbre,” a singer has a set
of timbres, or a timbre transformation. Therefore,
singers with similar timbre transformations would
constitute members of the same voice timbre type
or voice category. Based on this hypothesis, three
Journal of Voice, Vol. 17, No. 2, 2003
202 MOLLY L. ERICKSON
FIGURE 2. MDS perceptual space for singers as perceived by inexperienced listeners (a) and experienced listeners (b).
predictions can be made. First, listeners will per-
ceive dissimilarity across pitch within the same
singer. Second, any two voices, or more likely,
voice categories, may be perceived as having similar
timbre at one pitch and dissimilar timbre at an-
other. Third, singers of the same voice category
should have similar MDS representations. A second
hypothesis was that listener experience would affect
the perception of dissimilarity across pitch.
For both experienced and inexperienced listeners,
pitch correlated most highly with dimension 1, in
spite of instructions to disregard pitch. As there
are necessarily timbre changes associated with pitch
changes, it is impossible to determine to what extent
either of these perceptual cues affected listener judg-
ments. Regardless, it is important to note that these
listeners perceived greater dissimilarity across pitch
range within singers than they did between singers,
regardless of voice category. Yet, inspection of di-
mension 2 shows that listeners also perceived great
similarity within voice categories, although, as hy-
pothesized, sopranos and mezzo-sopranos were
heard to be similar to each other at some pitches
and dissimilar at others. Thus, the MDS solution
provides a visual cue to the pattern of timbre trans-
formation that constitutes the “signature” of each of
these vocal categories. These findings provide strong
Journal of Voice, Vol. 17, No. 2, 2003
perceptual support for the existence of these theoreti-
cal categories when conceptualized as timbre trans-
formations, rather than as static timbre categories.
In this study, for both inexperienced and experi-
enced listeners, dimension 2 separated singers based
on voice category and correlated moderately well
with spectral centroid from 2 to 5 kHz. Spectral
centroid values were higher for both sopranos than
they were for both mezzo-sopranos at all pitches
except A5, with the greatest difference between the
two groups in the mid-pitch range (G4, B5). Al-
though spectral centroid is affected primarily by the
location of peaks in the output spectrum, it is also
secondarily affected by spectral slope. Examination
of the spectra for each of the stimuli revealed at
least one visible peak in the upper frequencies for
all stimuli at pitches A3 through F5, but no visible
peaks in the upper frequencies for any stimuli at
the pitch A5. It is therefore likely that the spectral
centroid measurements at pitches A3 through F5
were primarily affected by location of spectral peaks,
whereas the spectral centroid measurements at the
pitch A5 were primarily affected by spectral slope.
As spectral centroid correlated well with dimension
2 in the pitch range A3 through F5, these results
lend support to the notion proposed by Dmitriev and
Kiselev12 and Sundberg21 that location of higher
spectral peaks provides a strong perceptual cue to
 CLASSIFICATION OF FEMALE SINGING VOICES 203

FIGURE 3. Two-dimensional representations of the MDS perceptual space for singers as perceived by inexperienced listeners.
Dimension 2 as a function of dimension 1 (A), dimension 2’s acoustic correlate, spectral centroid, as a function of pitch (B),
dimension 3 as a function of dimension 1 (C). There was no acoustic correlate to dimension 3.

vocal category. It is interesting that for the pitch
A5, where no spectral peaks were noted, listeners
heard little difference between voices, regardless
of category.
This study found several differences between in-
experienced and experienced listeners in the percep-
tion of timbre as a function of pitch. First, although
both experienced and inexperienced listeners used
pitch as the primary cue in this dissimilarity task,
experienced listeners judgments were less highly
correlated with pitch than inexperienced listener
judgments, suggesting that experienced listeners
have developed listening strategies that use pitch
as a cue to a lesser degree. Second, inexperienced
listeners, although able to discriminate differences
between voice categories based on spectral centroid,
were not able to discriminate differences between
voices within a voice category. Experienced lis-
teners, on the other hand, were able to discriminate
between singers within voice categories seemingly
using temporal cues, as evidenced by the correlation
between extent of vibrato rate and dimension 3 for

Journal of Voice, Vol. 17, No. 2, 2003

204 MOLLY L. ERICKSON

FIGURE 4. Two-dimensional representations of the MDS perceptual space for singers as perceived by experienced listeners.
Dimension 2 as a function of dimension 1 (A), dimension 2’s acoustic correlate, spectral centroid, as a function of pitch (B),
dimension 3 as a function of dimension 1 (C), and dimension 3’s acoustic correlate, vibrato rate, as a function of pitch (D).

these listeners. Such results are not necessarily con-
trary to those obtained by Bloothooft and Plomp2
who found no difference in performance between
experienced and inexperienced listeners on dissimi-
larity tasks using vocal stimuli at one pitch with
temporal characteristics removed. For the experi-
enced listeners, the lack of a strong correlation be-
tween dimension 3 and any one acoustic measure
suggests that these listeners probably use multiple
idiosyncratic cues based on their own experiences
and the peculiarities of the data to make distinc-
tions between singers within a vocal category. The
strongest such cue in the case of these four singers
was vibrato rate.
These results show that when presented with a
pair of stimuli, it is impossible for listeners to judge
voice similarity independently of pitch differences.
If that was possible, the multidimensional represen-
tation would consist of discrete clusters of stimuli,
one for each voice or, more likely, one for each
category. Yet, category differences do emerge in the
spectral centroid dimension (dimension 2). In that
dimension, voices in the same category tend to group
at each pitch and tend to share dimension 2

Journal of Voice, Vol. 17, No. 2, 2003

 CLASSIFICATION OF FEMALE SINGING VOICES
coordinates of the same polarity (e.g., positive values
for the sopranos and negative values for the
mezzo-sopranos). Thus, although listeners could not
disregard pitch differences when comparing stimuli
within a singer across pitch, they were nonetheless
able to hear timbre differences between categories
across pitch.
It is interesting to note that experienced listeners
differed from inexperienced listeners in the manner
in which they weighted pitch in dimension 1. Stimuli
of the same pitch tended to be scaled similarly by
these listeners across all voices, as indicated by the
nearly vertical arrangement along this dimension
(see Figure 3). On the other hand, for the experienced
listeners in the vocal experiment, stimuli of the same
pitch were not scaled similarly across all voices
(see Figure 4). In fact, the experienced listeners heard
greater dissimilarity across pitch range for both so-
pranos than they did for both mezzo-sopranos, sug-
gesting that these listeners may have been less
influenced by pitch and harmonic spacing than were
the inexperienced listeners.
The fact that listeners were able to make consis-
tent judgments for each voice in the face of large
pitch differences suggests that although the timbre of
an individual singer is not invariant across pitch,
the singer’s timbre transformation may provide per-
ceptual cues to the identity (or at least, category) of
the singer. At the present time, it is unknown whether
listeners can abstract such timbre transformations in
an identification task. However, it could be specu-
lated that through normal experience and/or explicit
training, listeners come to build a model for how the
timbre of one singer changes across pitch. In other
words, listeners may build a unique timbre trans-
formation for each singer or perhaps for each singing
category. This is analogous to the uncanny ability
of some individuals to connect baby pictures to the
correct adult pictures. In both cases, the perceptual
invariant must be linked to mechanical constraints.
In the former case, the mechanical constraints are
based on the physical properties of the laryngeal
source and vocal tract.7 In the latter case, the me-
chanical constraints can be defined in terms of
growth patterns.26 Thus, listeners may judge two
vocal stimuli as sounding very different and yet
maintain that they were produced by the same singer
or the same category of singer. Listeners could use
205
their knowledge of a singer’s timbre transformation
to discriminate between individual singers or catego-
ries of singers, regardless of pitch differences. If this
explanation is true, then these results suggest that
a task comparing two pitches, although providing
sufficient information to distinguish between catego-
ries of voices, does not provide the necessary percep-
tual information to discern timbre transformations
across a wide range of pitches that can be used to
identify a voice across pitch. Moreover, it helps
to explain why previous work using voices at one
pitch did not find consistent categories.2
Studies such as this are limited due to the combi-
natorial nature of multidimensional scaling. It was
not possible to include more than four singers or
more than one vowel. For example, increasing the
number of singers in each category from 2 to 3
would have increased the number of paired stimuli
from 276 to 630. Thus, although these results do
not disprove the hypotheses postulated here, they
cannot necessarily be generalized to the larger popu-
lation of singers. Studies such as this one should be
replicated with other singers and singing categories.
If listeners use timbre transformations to help
identify sound sources such as voices over varying
pitch and loudness, then it is important to understand
how listeners construct such transformations. Using
a three-note oddball task where two stimuli were
produced by the same singer and a third stimulus
was produced by a different singer, Erickson et al.27
found that when the stimuli produced by the same
singer were separated by more than one octave, lis-
teners were unable to construct a timbre transforma-
tion sufficient for identification of the “oddball”
stimulus. Future research should focus on the ability
of listeners to build such transformations based on
the number of same-singer stimuli and the pitch
differences between same-singer stimuli.
SUMMARY AND CONCLUSIONS
Traditionally, timbre has been defined as that per-
ceptual attribute that differentiates two sounds when
pitch and loudness are equal. By this definition, each
voice or category of voices possesses a set of tim-
bres, and the identity of any source across different
pitch–loudness combinations, or in the case of singers,
Journal of Voice, Vol. 17, No. 2, 2003
206 MOLLY L. ERICKSON
across different pitch–loudness–vowel combina-
tions, must be due to an abstraction of the pattern
of timbre transformation. Using stimuli produced
across the singing range by singers from different
voice types, this study sought to examine how timbre
and pitch interact in the perception of dissimilarity.
The resulting MDS representations showed that al-
though listeners cannot ignore pitch differences in
the perception of dissimilarity, non–pitch-related
timbre differences can be perceived within the con-
text of pitch. Such representations of dissimilarity
across pitch provide a glimpse of the timbre transfor-
mations associated with voice categories. For these
data, timbre transformations in dimensions 2 were
nearly identical for singers of the same voice cat-
egory, providing strong evidence for the existence
of these theoretical categories.
Acknowledgements: The author would like to express
her sincere gratitude to Joshua Lacey for his assistance
with the experimental procedures and to Stephen Handel
for his support, encouragement, and comments.
REFERENCES
1. ANSI. Psychoacoustical Terminology. S3.20. New York:
American National Standards Institute; 1973.
2. Bloothooft G, Plomp R. The timbre of sung vowels. J
Acoust Soc Am. 1988;84:847–860.
3. Iverson P, Krumhansl CL. Isolating the dynamic attributes
of musical timbre. J Acoust Soc Am. 1993;94:2595–2603.
4. Plomp R. Timbre as a multidimensional attribute of com-
plex tones. In: Plomp R, Smoorenburg GF, eds. Frequency
Analysis and Periodicity Detection in Hearing. Leiden, The
Netherlands: Sijthoff; 1970:397–414.
5. Samson S, Zatorre RJ, Ramsay JO. Multidimensional scal-
ing of synthetic musical timbre: Perception of spectral and
temporal characteristics. Can J Exp Psych. 1997;51:307–
315.
6. Vennard W. Singing, the Mechanism and the Technique.
New York: Fisher; 1967.
7. Cleveland TF. Acoustic properties of voice timbre types
and their influence on voice classification. J Acoust Soc
Am. 1977;61:1622–1629.
8. Keidar A, Hurtig R, Titze I. The perceptual nature of vocal
register change. J Voice. 1987;1:223–233.
9. Lakatos S. A common perceptual space for harmonic and
percussive timbres. Percept Psychphys. 2000;62:1426–
1439.
Journal of Voice, Vol. 17, No. 2, 2003
10. ASHA. Guidelines for screening for hearing impairments
and middle ear disorders. ASHA. 1990;32:17–24.
11. Fant G. Acoustic Theory of Speech Production. The Hague:
Mouton; 1960.
12. Dmitriev L, Kiselev A. Relationship between the formant
structure of different types of singing voices and the dimen-
sions of the supraglottal cavities. Folia Phoniatr (Basel).
1979;31:238–241.
13. Sundberg J, Högset C. Voice source differences between
falsetto and modal registers in counter tenors, tenors and
baritones. Log Phon Vocol. 2001;26:26–36.
14. Titze IR. Principles of Voice Production. Englewood Cliffs,
NJ: Prentice Hall; 1994.
15. Sundberg J, Gauffin J. Waveform and spectrum of the glottal
source. In: Lindblom B, Öhman S, eds. Frontiers of Speech
Communication Research, Festschrift for Gunnar Fant.
London: Academic Press; 1979:301–320.
16. Gauffin J, Sundberg J. Spectral correlates of glottal voice
source waveform characteristics. J Speech Hear Res. 1989;
32:556–565.
17. Mori K, Blaugrund SM, Yu JD. The turbulent noise ratio:
An estimation of noise power of the breathy voice using
PARCOR analysis. Laryngoscope. 1994;104:153–158.
18. Hertegard S, Gauffin J. Voice source-vocal tract interaction
during high-pitched female singing. In: Friberg A, Iwarsson
J, Jansson E, Sundberg J, eds. Proceedings of the Stockholm
Music Acoustic Conference (SMAC93), 79. Stockholm:
Royal Swedish Academy of Music; 1993:177–182.
19. Holmberg EB, Hillman RE, Perkell JS, Guiod PC, Goldman
SL. Comparison among aerodynamic, electroglottographic,
and acoustic spectral measures of female voice. J Speech
Hear Res. 1995;38:1212–1223.
20. Klatt DH, Klatt LC. Analysis, synthesis, and perception of
voice quality variations among male and female talkers.
J Acoust Soc Am. 1990;87:820–857.
21. Sundberg J. Perceptual aspects of singing. J Voice. 1994;
8:106–122.
22. Monsen RB, Engebretson AM. The accuracy of formant
frequency measurements: A comparison of spectrographic
analysis and linear prediction. J Speech Hear Res. 1983;
26:87–97.
23. Sandell GJ. Roles for spectral centroid and other factors in
determining “blended” instrument pairings in orchestration.
Music Percept. 1995;13:209–246.
24. Prame E. Measurements of the vibrato rate of ten singers.
J Acoust Soc Am. 1994;96:1979–1984.
25. Prame E. Vibrato extent and intonation in professional
Western lyric singing. J Acoust Soc Am. 1997;102:616–621.
26. Pittenger JB, Shaw RE, Mark LS. Perceptual information
for the age level of faces as a higher-order invariant of
growth. J Exp Psych Human Percept Perform. 1979;5:
478–493.
27. Erickson ML, Perry S, Handel S. Discrimination functions:
Can they be used to classify singing voices? J Voice. 2001;
15:492–502.