The Composition of "Vox-5"

Author(s): Trevor Wishart

Source: Computer Music Journal, Vol. 12, No. 4 (Winter, 1988), pp. 21-27
Published by: The MIT Press
Stable URL: http://www.jstor.org/stable/3680150
Accessed: 19/01/2009 14:10

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=mitpress.

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

JSTOR is a not-for-profit organization founded in 1995 to build trusted digital archives for scholarship. We work with the
scholarly community to preserve their work and the materials they rely upon, and to build a common research platform that
promotes the discovery and use of these resources. For more information about JSTOR, please contact support@jstor.org.

The MIT Press is collaborating with JSTOR to digitize, preserve and extend access to Computer Music
Journal.

http://www.jstor.org
TrevorWishart
Realspace Musics
83 Heslington Rd.
YorkY01 5AX
United Kingdom
Introduction
Vox-5 is the fifth in a series of six pieces exploring
the musical possibilities of the human voice. All
the other pieces in the series are for live perfor-
mance by four amplified vocalists, and most of
them also use sound-on-tape as the accompanying
medium, and simultaneously the environment in
which the vocalization takes place. Each piece in
the series is concerned with different ideas of mu-
sical structure, different performancetechniques,
and different aspects of human experience. Prior
to the composition of this series I had developed
a thorough system of notation for vocal sounds
together with a systematic classification of vocal
techniques. Vox-5 forms a bridgingrecitative into
the finale of this series (Vox-6), and is also a "poetic
summary" of the preceding material.
The primaryauralfocus of Vox-5is a (super)human
voice that metamorphoses into many recognizable
sonic images, such as the sounds of crowds, bees,
a horse, or bells. This voice also employs various
extended vocal techniques (such as ingressive com-
plex sounds, subharmonics, ululation, and special
consonant production).These are further extended
by spectral manipulation of the material by the
computer. In all the spectral transformationsone
aesthetic aim has been to retain the "source credi-
bility" of the resulting sounds; that is, they must
always be believably vocal or naturalistic, even if
startlingly unexpected in their spectral develop-
ment. I have been working with ideas of sound
transformationsince about 1970, but not having
had access to computing facilities, I developed my
previous techniques in a classical tape studio, using
only editing, mixing, and speed changing. The op-
portunity to realize these processes in a much more
This article refersto a set of sound examples. These examples
will appearon a soundsheet published in the next issue, Volume
13, Number 1, (Spring1989).
ComputerMusic Journal,Vol. 12, No. 4, Winter 1988,
? 1988 Massachusetts Institute of Technology.
The
Composition
of Vox-5
precise manner was one of the principal reasons for
coming to the Institut de Recherche et Coordina-
tion Acoustique/Musique (IRCAM)in Paris to do
a project.
Techniques
Compositional
The two majortechniques used in the composition
of Vox-5 are those of spectral manipulation, and
spectral interpolation. In both cases, the source
sound or sounds are first analyzed using the phase
vocoder program(Dolson 1985; Gordon and Strawn
1985).
The phase vocoder (PV)effectively divides the
source sound into a number of channels that are
equally spaced through the range of frequencies.
In each channel it stores the amplitude of the sig-
nal component, and its frequency (usually known
as "phase"in this context).It is now possibleto
manipulate these analysis data. For example, to
shift a spectralcomponentin channeln up one
octave,we need to move the amplitudeandphase
in this channelinto channel2n (whichis centered
aroundtwice the frequency)andalso multiplythe
phaseby two. (Phasecorrespondsto frequency,
so an octaveshift meansmultiplyingfrequency
by two.)
It is possibleto deviseprogramsthat will system-
aticallytransformthe completespectrumgenerat-
ing a secondanalysis-datafile, which can then be
resynthesizedusing the phasevocodersynthesisop-
tion. These transformations can be time-dependent
(the spectralshiftingcan follow an envelope,or
reada value-timefunctionto ascertainits valueat
any particulartime).The programsI wroteto do
this arenow availableon the systemat IRCAMand
as partof the ComposersDesktopProjectsoftware
(Atkinset al. 1987;formoreinformation,contact
the ComposersDesktopProject,11 KilburnRd.,
YorkY014DF,U.K.).(Iwouldlike to acknowledge
the adviceandassistanceof XavierRodet,Yves
Potard,ErikViara,Dan Timis, MillerPuckette,and
Wishart 21
others in helping to clarify aspects of the phase
vocoder, ironing out my C programmingtechnique,
and making helpful suggestions about directions
worth exploring!)
Special Shifting
Shifting and Stretching the Spectrum
In Vox-5, the spectral shifting techniques were
confined to global expansion or contraction of the
spectral data. It is useful to be able to stretch just a
part of the spectrum, so the programcalled specsh
divides the spectrum into two parts about a par-
ticular PV channel (referredto in the programsas
fdcno, or frequency-dividing channel number)
and then shifts the upper or lower portion (or both).
In this way, for example, subharmonics and other
kinds of (apparent)double-productioncan be in-
troduced into the voice, and (through the use of
the time-varying options) can be made to move
independently of one another. At the beginning
of Sound Example 1 (see the soundsheet in the
next issue, Volume 13, Number 1, Spring 1989),
the source sound is a voice producinga (real)sub-
harmonic at one octave and a fifth below the fun-
damental. The specsh programhas been used to
extend this subharmonic down another octave and
then cause it to slide upwardtoward its original
register in the course of the sound.
There are two fundamentally different ways in
which the spectral data can be expandedor com-
pressed. In the first case, the harmonic relation-
ships between (moved)components are preserved
(this I refer to as shifting of the spectrum). In the
other case, all the (moved) components move in the
same direction, but the amount they move depends
on their positions in the spectrum (this I refer to
as stretching of the spectrum). As a rule-of-thumb
for preservingvocal "credibility"in the resulting
sound, shifting is best applied to clearly pitched
sources, and stretching to noisy sources, but either
technique can be applied to either type of source. In
practice, experimentation on the particularsound
source is essential. The programspecsh proved to
be particularlyuseful for a subtle type of perceptual
masking discussed later under "Interpolation."
22
Preserving Formants
Moving spectral components of the voice a long
way distorts the original spectral envelope or for-
mant structure of the source. In some cases this is
undesirable, so the programspecf makes an approx-
imate calculation of the spectral envelope of the
source, which is retained on shifting the spectrum.
Different Kinds of Stretching
The stretching of the spectrum in specsh is lin-
ear; that is, for upwardstretching above fdcno, the
stretch at fdcno is zero. Intermediate spectral com-
ponents are shifted by a factor (less than or equal to
one) before being applied to any specific channel.
The value of this factor varies (linearly)from zero
to one as we progressfrom channel fdcno to the
topmost channel (Fig. 1).
One transformationI wanted to experiment with
was the stretching of a vocal source to produce an
inharmonic spectrum like a bell sound (as this is
one of the few sounds the voice cannot imitate).
The results with linear stretching were disappoint-
ing, so I wrote a modified version of the stretch part
of specsh, called spece, which allowed me to vary
the type of interpolation used in the stretching pro-
cess. For example, if the multiplying factor (which
is always less than one) is multiplied by itself, be-
fore being applied to stretch, we produce a stretch-
ing curve that is concave. In this case, the spectral
components are shifted less at the bottom of the
spectrum, and more at the top (Fig. 2). Conversely,
taking the square root of the multiplying factor, we
produce a convex curve, and the spectrum is shifted
more at the bottom than at the top of the spectrum
(Fig. 3).
The spece programpermits the multiplying fac-
tor to be raised to any power, and that power factor
can be specified by the composer. In Sound Ex-
ample 2 the original vocal source has been stretched
using a power factor of 0.4. Through the sequence
the maximum shift moves from 1.05 to 6.0.
After this spectral shifting I also applied time
stretching to the sources, taking care not to time-
stretch the attack portion of the sounds. The pro-
cess of spectral stretching is complemented by
Computer Music Journal
Fig. 1. Linear spectral Fig. 2. Concave spectral Fig. 3. Convex spectral
shifting from fdcno to the shift. shift.
top of the spectrum.

Fig. 1 Top of spectrum

Top of spectrum

)-4

c)

a)
(3
6

fdcno !

fdcno 0 1
Shift multiplier
0 Shift multiplier 1

(component-shift = input-shift-value + shift-multiplier)

cence of the sound image as "bell,"while remain-
ing (at the start of the sequence) "voices."
Fig. 2
Top of spectrum Spectral Envelopes

All the previously mentioned programs allow one

u
a)
"S
z)
U
<^l
d4
fdcno
0 Shift multiplier 1 change is interpolated between the original and
the increased duration of the resulting sounds
(enhancing the perception of a transition toward
"bell-like"). The results are heard in Sound Ex-
ample 3. The time-stretching process introduces
distortions (appearingas subharmonics or ampli-
tude-modulation-type effects), which I decided to
retain as part of the final mixed sounds. In the final
composition of the piece, several such sequences
were synchronized, component-by-component, to
produce the sequence heard in Sound Example 3.
Here the ability to exactly align the perceived vocal
attacks (which differby between 0.06 and 0.1 sec
from onset of the sound) is crucial to the coales-
to define an envelope on the evolution of the spec-
tral shifting. This is similar to the traditional
attack-decay-sustain-release (ADSR) envelope on
analog synthesizers, but is defined as attack, diver-
gence, on, convergence, end. During the attack and
end portions there is no spectral change, and dur-
ing the on portion the spectral change is as pro-
grammed. In the other two portions the spectral
the changed versions.
The program specsh goes one step further, in that
it allows the parameters of the spectral shifting
(stretching) to be input as a time-varying function.
This program works on specifically defined bands
of the spectrum.
Spectral Interpolation
The other major technical procedure used in the
composition of Vox-5 is that of interpolating,
through time, between two different spectra to
generate a seamless transformation between the
two original sounds. Mathematically, and compu-

Wishart 23
tationally speaking, the process of interpolation
is quite straightforward.All we need to decide is
when, for how long, and in what manner we inter-
polate between the phase and amplitude values in
the two analysis files. Any kind of interpolation
can be mathematically described. The problem is
primarily that of finding musically and acoustically
satisfactory solutions. There are four aspects to this
problem.
1. Matching of Sources
First, it goes without saying that the two source
sounds should have as many similarities as pos-
sible. In the now well-known examples of inter-
polation between pitched and instrumental sounds
made at StanfordUniversity in the 1970s, the prob-
lem was simplified by dealing with harmonic spec-
tra of stable tessitura. With more complex sounds,
a high level of acoustic/musical judgment is re-
quired to match the sources sufficiently before
embarking on what can be a long interpolation
computation. In this task, the spectral and mor-
phological malleability of the voice as a sound pro-
ducer is vital. The voice can be made to match
almost any other sound, apartfrom bell-like stable
inharmonic spectra. Knowledge of extended vocal
techniques is a great asset here.
2. Preprocessingof Sources
I brought with me a simple technique from the ana-
log studio. A complex source can be reduced in its
spectral detail by a process of spectral thickening,
which makes it much easier to match to another
source. Thus many of the analog sources I brought
with me to IRCAMhad been wedged or diswedged
as follows. To make a diswedge, the sound is re-
corded backward.It is then mixed with two copies
of itself, not synchronized, with the copies diverg-
ing slightly in tape speed from the first as the mix
proceeds. The two copies fade in imperceptibly
during the mix, thus progressivelythickening the
mix. This process is then applied to the resulting
mix (taking care to avoid echo effects between prom-
24
inent features in the source), producing a nine-
strand wedge. Three further repetitions generate a
243-strandwedge. This is then played backward,
and the original sound emerges out of the 243-times-
thickened spectrum. In this way the sound of a
crowd can be more easily matched with a plosive
unvoiced k.
3. Specifying the Interpolation
The programvocinte allows one to specify a start
and end time for the interpolation of amplitude, in-
dependent of the interpolation of phase, and vice
versa. The two interpolations could also be made to
proceed either linearly or by specifying a power fac-
tor (see the section "DifferentKinds of Stretching")
for amplitude, and another for phase. A great deal
of experimentation was requiredwith each pair of
sources to determine the values for start-and-end
times and power factor that gave the best result.
Other rules-of-thumb included the following:
Amplitude and phase should be interpolated over
the same time interval.
Amplitude interpolation should have power fac-
tor 0.8 (rapidat start, slowing at end).
Phase interpolation should have power factor 2
(slow at start, and rapid at end).
But these rules broke down in the case of Sound
Example 1 (an interpolation between a vocal voiced
zzz and the sound of bees). At the start, the curves
for amplitude and phase interpolation are exactly
the opposite of the rule (!),while phase interpo-
lation takes place in a much shorter time than
amplitude interpolation (Table 1). [Note again:
Sound examples are on a soundsheet included with
issue 13(1), Spring 1989.]
4. Perceptual Boundaries
The fourth and potentially most difficult area for
sound interpolation is the problem of categoric
recognition. No matter how smooth an interpola-
tion may be acoustically or mathematically, we are
perceptually prone to make sudden leaps in our per-
Computer Music Journal
Table 1. Sound Example 1 Interpolation Parameters
Interpolation Start Time End Time Power Factor
Amplitude 0.2 sec 7.8 sec 2.0
Phase 0.2 sec 3.0 sec 0.8
Front Right Behind Left
ception of unfamiliar objects. As suggested by ca-
tastrophe theory, when perceiving a continuously
changing object that we initially recognize, we con-
tinue to interpret it as the original object until a
certain threshold is reached. At this point there is
likely to be a sudden switch to the recognition of
a new object. To minimize this perceptual switch-
ing, I reprocessedthe interpolated sound, making
(usually the upper)part of the spectrum move in-
dependently through the moment of maximum
ambiguity. In this way the tendency to make a per-
ceptual switch is distracted by our following the
other movement in the spectrum (which is not di-
rectly related to the spectral interpolation that is
going on) and we are more likely to perceive the
transition as seamless. This process can be heard
at the beginning and end of Sound Example 1.
In Sound Example 4, on a first listening we hear
the sound of a human voice becoming a horse neigh.
However, on a second hearing (in isolation), be-
cause I knew it would become a horse, I heard it all
as a horse! This difficulty can be overcome within
the structure of the piece, however, by the choice of
an appropriatecontext. As (almost) every other
sound in the piece begins as a vocal source, the
context is established for the listener to interpret
the start of this sound as vocal (which it is!) so the
perceptual ambiguity does not arise.
Making transitions back to the voice proved to be
generally more problematic than divergences from
the voice. This often required that the sound being
interpolated from persist, and, as it faded away, fol-
low exactly the same amplitude contour and mo-
tion in space (see the next section) as the target
sound, to make a convincing convergence. Sound
Example 6 gives an approximateindication of the
process involved (the true version has sounds con-
verging from a quadraphonicspace).
The programsdescribedhere can never give au-
tomatically valuable results. They need to be fine
tuned to the particularsource sound and the sonic
result required.Musical decision-making is still
paramount, and a most important aspect of this
work has been that composition can now approach
a sculptural process in which the composer fash-
ions the sounds just as a sculptor might work with
physical materials. Each sonic object is unique, and
it requires a special sort of acoustically informed
open-minded approachto mold it to the desired
end. The mental attitude is quite distant from the
predetermininglogical thoroughness that we have
been taught to bring to the writing of an instru-
mental score, and adds an exciting new dimension
to the sphere of compositional competence.
Spatialization
One further important structural feature of Vox-5
is the use of spatial placement and movement to
articulate the sonic development and the aural
imagery of the materials. I have used computer-
controlled, prescoredspatialization in the piece
Vox-1, where the live sounds of the four vocalists
are moved (independently)aroundthe quadraphonic
space by a custom-built device and software that I
developed, using ambisonic technology. I would
have preferredto spatialize the sounds in Vox-5 di-
rectly on the DEC VAXcomputer at IRCAM.It is
possible to do this, but there were only two digital-
to-analog converters available, so it would not have
been possible to hear the results of one's moves and
thus test the validity of one's compositional pre-
conceptions. The sound materials were therefore
spatialized during a mixdown on IRCAM's24-track
digital tape recorder,using a simple quadrapan
device.
I have written much about the significance of
"landscape"in the perception and composition of
electroacoustic music. The sonic space I am refer-
ring to is indicated in Fig. 4, where the numbers
represent spatial positions and also loudspeaker
placements. Thus the position 1/2 refers to the
place equidistant from the two places 1 and 2.
In a system with a limited number of loudspeak-
Wishart 25
Fig. 4. Layout of playback
channels. Position "1/2" is
between channels 1 and 2.
1 Front 2 used in Vox-5: (1) the motion of individual sounds
1/2
e. 1/3 Audience 2/4g:
3/4
3 Behind 4
ers our aural discrimination of direction and mo-
tion is limited. Moreover,spatial motion is best
perceived with certain types of sounds (e.g., high-
frequency noise bands and repeating sounds). Mo-
tion is thus confined primarily to the lines joining
the loudspeakerpairs (including the diagonals).In
addition, perception of spatial position tends to be
strongly distorted for audience members sitting a
long way off-center. This effect can be remedied sig-
nificantly by ambisonic encoding and projection of
the sounds; an ambisonic master of the finished
piece has also been completed.
The principal sonic image of the piece is a voice
located at 1/2. The emergence (genesis) of other
sounds out of vocal initiators is emphasized by
their spatial "ejection"from this location into the
quadraphonicspace aroundthe listeners. This is
most dramaticallyheard in the five rapidvocal
noise band attacks in the early part of the piece (re-
curringin curtailed form later). These are "ejected"
out into the corners of the space. In other cases
the process is enhanced by moving from a "single-
stream" vocal source located at 1/2, to a stereo
sound. Here it is necessary to separate the two
channels of the stereo source, analyze and interpo-
late them separately (using exactly the same pa-
rameters for each), and then recombine the results.
26
Two different kinds of motion organization are
within the space, like the three explosive unvoiced
du ... sounds behind the "choralbells" material,
which travel individually aroundthe space in differ-
ent arcs; and (2) "framemotions," such as that of
the vocal-ululation material (near the beginning of
the piece) and the ingressive complexes (near the
end), where three streams of similar (but separable)
materials move simultaneously aroundthe space,
redefining their relationship to one another. This
latter type of motion seems (perceptually)more
like a constant shifting of the aural space itself,
and is almost equally effective from all listening
positions.
Other Techniques
Other computer music techniques were employed
in the composition of Vox-5.One importantmethod
was the use of the mix programto control very pre-
cisely the envelopes of sound events. This was par-
ticularly crucial when joining end-to-end sections
of vocal material recordedindependently.In order
to create the illusion of a continuous vocal stream,
the finishing sound had to be enveloped (often over
a period of only 0.05 sec) to suggest a vocal closure
or the preparationin the vocal cavity of the next
sound. Conversely, the beginning of the next sound
often had to be enveloped to suggest the reopening
of the mouth. This kind of subtle detail should re-
main unperceived by the listener if it is correctly
done, but would be jarringlyobtrusive if not. I
enjoy this kind of "self-effacing"technique, be-
cause it contradicts the mythology of a composer
as a strongly projected ego in music-making!!
Coda
The creation of Vox-5 helped me to test out many
ideas about the control of musical articulation in a
continuum, about spectral interpolation, and about
the organization of sound in space. I am particu-
larly grateful to IRCAMfor inviting me to use its
facilities, as my ideas could not have been realized
Computer Music Journal
without these computer music resources (no
equivalent facilities exist in the U.K.).
The computer opens up areas of compositional
exploration that were previously inaccessible. The
precision with which sound materials can be speci-
fied implies two things: (1) Given an understanding
of acoustics, sounds can be transferreddirectly
from the composer's imagination to the perfor-
mance situation; (2) Areas of sonic organization
previously inaccessible to composers through the
existing media of notation can be explored, open-
ing up a new world of dreamedof, but unsung
possibilities.
In the future I would like to explore more
precisely the musical possibilities of moving and
"moving/static" aural events like forms in fluids.
Consider the turbulence aroundan obstruction that
appearssemistatic in space, but is constructed en-
tirely out of rapidly moving flows. I hope to relate
the sonic evolution to the global properties of fluid
flow (e.g., the onset of turbulence, granulationpro-
cesses) as part of a compositional research project.
References
Atkins, M., A. Bentley, T. Endrich,R. Fischman, D. Mal-
ham, R. Orton, and T. Wishart. 1987. "The Composers'
Desktop Project."In S. Tipei and J.Beauchamp,eds.
Proceedings of the 1987 International Computer
Music Conference. San Francisco:Computer Music
Association, pp. 146-150.
Dolson, M. 1986. "The Phase Vocoder:A Tutorial."
ComputerMusic Journal10(4): 14-27.
Gordon,J.,and J. Strawn. 1985. "An Introductionto the
Phase Vocoder."In J. Strawn, ed. Digital Audio Signal
Processing:An Anthology. Madison, Wisconsin: A-R
Editions, pp. 221-270.
Wishart 27