___________________________

FROM GRAINS TO FORMS

Curtis Roads
Media Arts and Technology
University of California, Santa Barbara
[email protected]
In Makis Solomos, (ed.), Proceedings of the international Symposium
Xenakis. La musique lectroacoustique / Xenakis. The electroacoustic
music (universit Paris 8, May 2012).
ABSTRACT
This paper explores strategies for organizing sound grains into larger
structures on the meso and macro time scales. We begin with a look at
clouds, streams, sprays, and Xenakiss original proposal based on a
sequence of screens. Then we examine tools for montage and micromontage in the studio, as well as instruments for gestural control. This
leads to a discussion of higher-order granulation and per-grain processing.
Dictionary-based pursuit is an analytical counterpart to granular synthesis,
enabling transformations that derive new sounds from analyzed sounds.
Next we look at proposals for generating grains based on physical and
biological models. The next section looks at abstract generative models
based on algorithms imported from mathematics. This leads into a general
discussion on the limitations of formalism in music composition. The core
of the paper concerns the problem of multiscale organization, which can

resolve the tension between formal and informal approaches through a

combination of heuristic algorithms and direct interventions.

ENCOUNTER WITH XENAKIS

I first met Maestro Iannis Xenakis at his final course Formalized and
Automated Music at Indiana University in 1972 (Figure 1).

Figure 1. Seminar in Formalized and Automated Music, 1972.

I was deeply inspired by Xenakiss original musical discourse and free

imagination. (1) Yet it was evident that Xenakis was not happy in
Bloomington (Logan 2000). We heard that the administration was not
maintaining the expensive hardware needed for computer sound synthesis

experiments. Not surprisingly, soon Xenakis returned permanently to

Paris.
In the following months, I studied Formalized Music and developed
computer programs to test the formulas in the book (Roads 1973). In this
period, sound synthesis by computer required specialized facilities. I found
these at the University of California, San Diego (UCSD). There I
developed a variant of Xenakiss theory of granular synthesis, based on a
model of scattering grains into sound clouds (Roads 1978). I did not expect
that twenty years later I would be working in Paris at Les Ateliers UPIC
(later renamed CCMIX and finally CIX) alongside Xenakis and his quipe
at the CEMaMu, developing another program for granular synthesis called
Cloud Generator (Roads and Alexander 1995) (Figure 2).

Figure 2. The Cloud Generator app in action 1995.

Since this time, the granular paradigm has proven to be one of the most
powerful methods of synthesis and sound transformation, implemented in
dozens of incarnations and used by innumerable musicians.
Granular synthesis requires an algorithmic model of grain generation and
organization. In this paper, I explore various granular models and how they
lead to higher-level musical structures. Of course, as Jean-Claude Risset
(2005) observed, a characteristic of the granular paradigm is that it blurs
the border between microstructure and macrostructural organization:
By bridging gaps between traditionally disconnected spheres like material
and structure, or vocabulary and grammar, software creates a continuum
between microstructure and macrostructure. It is no longer necessary to
maintain traditional distinctions between an area exclusive to sound
production and another devoted to structural manipulation on a larger
temporal level. The choice of granulation, or of the fragmenting of sound
elements, is a way of avoiding mishaps on a slippery continuum: it permits the
sorting of elements within a scale while it allows individual elements to be
grasped. The formal concern extends right into the microstructure, lodging
itself within the sound grain.

SCREEN, CLOUD, STREAM, AND SPRAY MODELS

Xenakis first demonstrated granular synthesis in his composition
Analogique B (1959). This was realized by recording sine tones on tape,
cutting the tapes into hundreds of tiny pieces, and then recombining them
by manual splicing according to a stochastically-generated score.

Xenakiss design for digital granular synthesis, as articulated in the chapter

Markovian Stochastic MusicTheory was based on generating a
sequence of time-frequency screens, akin to frames of a film, running at a
fixed rate (Xenakis 1971). He proposed to control the sonic energy
distributed on each screen through set theory operations such as this
screen is the set union of two preexisting screens. The sequence of
screens was to be controlled by a Markov chain process. In this design, the
composer would control the synthesis by means of operations that are not
directly tied to acoustical parameters such as pitch, duration, amplitude,
spectrum, and spatial position (Xenakis 1971, chapter IX). In this sense,
Xenakiss design for granular synthesis was conceptually similar to his
Dynamic Stochastic Synthesis or GENDY developed in the early 1990s, in
which sound production is an epiphenomenon of an abstract generative
process taking place on the subsymbolic time scale of individual samples.
By contrast, software-based granulators today implement a flowing spray
jet of sound particles. Spray tools inevitably shape higher levels of a
composition; meso and macro forms tend toward patterns of stream and
cloud formations on the time-frequency plane (Roads 1978; Truax 1986;
Solomos 2006).
This is made visually explicit in programs like Metasynth (Wenger and
Spiegel 2004) where the user literally sprays grains onto a time-frequency
grid onscreen (Figure 3). In this sense, the cloud, stream, and spray models
align with the spirit of Xenakiss UPIC paradigm, in which users directly
draw sounds in the time-frequency plane (Xenakis 1992).

Figure 3. Screen image of MetaSynth granular sprays.

Within the stream and cloud models, the distinction between synchronous
and asynchronous granular synthesis is compositionally pertinent (Roads
2001b). Synchronous granular synthesis (SGS) emits one or more streams
of grains where the grains follow each other at regular intervals. A prime
use for SGS is to generate metric rhythms keeping the grain emissions
sparse per unit of time.
One of the most important parameters of granular synthesis is grain
densitythe number of grains per second. In the case of SGS, this
corresponds to a regular frequency of grain emission. For example, a
density of 2 grain/second indicates that a grain is produced every half
second a repeating beep. Synchronous densities in the range of about
0.1 and 20 grains per second generate metrical rhythms. When the
densities change over time, we experience precise accelerandi/rallentandi
effects. At higher densities, long grains fuse into continuous tones. Here is
the sweeter side of granular synthesis, since these tones tend to have a

strong fundamental frequency. Depending on the grain envelope and

duration, these tones will also manifest sidebands.
Any granulator can generate a huge amount of derived sound material
from a given sound file by manipulating the position, speed, and direction
of the read pointer. A slow backwards-scan granulation of a sound file
changes the identity of the original, especially when this is combined with
pitch shifting, filtering, and spatialization, all randomized on a grain-bygrain basis. (See the discussion of per-grain transformations in the next
section.) Reducing the duration of the grains has the inevitable effect of
churning any source into broadband noise.
Asynchronous granular synthesis (AGS) abandons the concept of perfectly
sequential streams of grains. Instead, it scatters the grains over a specified
duration within regions inscribed on the time-frequency plane. These
regions are clouds the units with which a composer works. The scattering
of the grains is irregular in time, being controlled by a stochastic algorithm
(Roads 1991). In the case of AGS, grain density corresponds to the degree
of transparency or opacity of the sonic fabric.

STUDIO-BASED MONTAGE AND MICRO-MONTAGE

The output of the Cloud Generator app is a single cloud of grains. The idea
is that the composer can generate a number of such clouds in the studio
and then organize them using a timeline-based mixing app like Pro Tools.

This approach is extremely free in terms of compositional options,

considering that material at any time scale can be processed by plugins and
placed anywhere in the time line. Detached from real-time constraints,
ideas can be tested, edited, submixed, or deleted at will.
One of the strategies that this enables is a thematic approach, i.e.,
manipulating sound material so as to create repetitions and variations. For
example, a copy of a given sonic entity can be pitch-shifted, time-scaled,
ring modulated, filtered, reversed, etc. Parts of the piece, if not the entire
work, can be organized as a montage of variations of a finite number of
elements. This approach forms the basis of my compositions Never (2010)
and Always (in progress) and all of the electroacoustic works of Horacio
Vaggione. It appears in his concept of micro-figures and was crystallized
by the IRIN app developed by Vaggiones student Carlos Caires (Figure
4).

Figure 4. IRIN figure editor.

GESTURAL CONTROL OF GRANULAR INSTRUMENTS

The first generation of granulators operated in non-real time (Roads 1978).

Truax (1986) developed real-time implementations of granular synthesis,
but these required specialized hardware. By contrast, I conceived the realtime Creatovox instrument (1999-2000) as having a software-based engine
written in SuperCollider with common MIDI controllers as the
performance interface (Roads 2001b).
In the early 1990s, I began a design notebook containing schemes for a
real-time granular synthesis instrument, including both scheduling
algorithms as well as protocols for gestural interaction (Roads 1992-1997,
1998). This instrument would look like a typical keyboard instrument, with
the additional of a few special controllers. It could be compatible with the
MIDI protocol. As a guiding principle, I wanted to convey an illusion of
simplicity to the performing musician by hiding unnecessary technical
details.
By 1998, common microprocessors became fast enough to handle multiple
voices of granular synthesis in real time. At the same time, James
McCartney released the SuperCollider 2 language (McCartney 1998).
SuperCollider 2 combined musical interaction with synthesis and sound
processing, so it seemed ideal for prototyping this project. On the basis of
these developments, I launched the Creatovox research project at the
Center for Research in Electronic Art Technology (CREATE) at UCSB.
The goal of the Creatovox project was to invent a prototype instrument
optimized for expressive virtuoso performance of granular synthesis based
on the cloud/stream paradigm.
The Creatovox (Figure 5) emitted its birth cries on 13 July 1999 in the
presence of a project team consisting of Alberto de Campo, Ching-Wei

Chen, and me. As of January 2000, it had gone through another round of
development and produced octophonic output.

Figure 5. The Creatovox instrument, January 2000. Here I play two simultaneous
clouds: one in the bass register (the bass pedal) and one in the high register
(keyboard). The left hand manipulates three parameters via a 3D joystick: grain
density, grain duration, and amount of reverberation for example, while the right
foot controls volume.

A funny thing happened at this point. Every time I went out in public to
demonstrate the Creatovox, the results disappointed me. It was not the
fault of our design, it had to do with the fact that I had not invested a great
deal of time to practice playing this instrument. What a surprise: a virtuoso
instrument requires a virtuoso performer who practices the instrument

10

every day! Ultimately I decided that becoming a virtuoso performer was

not central to my interests in composition. One composer who did use the
Creatovox was the wonderful Bebe Barron (the soundtrack of Forbidden
Planet), who made her final piece Mixed Emotions in the summer of 2000.
However, in this case the Creatovox was used to generate fragments that
were ultimately assembled in Pro Tools. (2)
Since then there have been many real-time implementations of granular
synthesis, some with more performance potential than others. A current
trend favors a graphical scrubbing model, after the tape scrubbing
methods of the analog era (Figure 6).

Figure 6. Borderlands app for iPad. The user slides (scrubs) a circle along the
waveform to select segments to granulate. The attached satellite circles function as
potentiometers controlling granular synthesis parameters. Another recent scrubbing
app is MegaCurtis.

11

Of course, the challenge of making a satisfying piece with a rich multiscale

architecture purely from a live performance on a granular instrument
remains a daunting one. It is not impossible, but will require considerable
practice

ENVELOPE CONTROL OF PULSAR PATTERNS

In contrast to the Creatovox, which was designed for virtuoso
performance, our PulsarGenerator app (Figure 7), coded by Alberto de
Campo, was designed to allow a composer plan detailed pulsar patterns
(pulsars are similar to grains) in the studio by means of a set of 12
envelopes with novel envelope-based operations like envelope mixing,
inversion, scaling, reversal, and smoothing, (Roads 2001a, 2001b).

12

Figure 7. The PulsarGenerator screen focused on envelope control of particle

synthesis.

HIGHER-ORDER GRANULATION
AND PER-GRAIN PROCESSING
Recycling sounds by means of higher-order granulation is a method of
spawning new granular mesostructures out of old ones. In effect, we regranulate one or more existing granulated sound files. Depending on the
capabilities of the granulation algorithm, a wide range of variations can be
generated. The resulting sounds can be many times the duration of the
original input sound. For example, a single stream of granulation using
large grains and a sharp attack envelope breaks a continuous stream into
discrete chunks. If this granulation stream has a wide range of amplitude
variations, each chunk will have its own dynamic articulation, creating
articulated phrases.
The technical capability of regranulation was available in my 1988
Granulate program, which could granulate up to 64 sound files at a time to
create new hybrid sounds out of existing sounds (Roads 2001b). However,
I did not begin to experiment with higher-order granulation until 2003 with
the realization of Now (2004), a regranulation of my composition Volt air
(2003). In turn, Never (2010) was a third-order granulation of Now. I am
currently working on a fourth-order granulation of Now in the work-inprogress Always.
These experiments rely on my Constant-Q Granulator (Roads 1998,
2001b) and the EmissionControl app (Thall 2004a,b; Roads 2006)
software. (3). EmissionControl (Figure 8) is particularly interesting as it

13

implements a matrix modulation scheme for control of synthesis

parameters. In particular, parameters can be controlled by low-frequency
oscillators (LFO) and random functions as well as by interactive sliders
directly manipulated by the composer. LFO control lends itself to the
generation of phrase structures based on cyclical patterns.

Figure 8. Screen of EmissionControl. The faders on the left determine the amount
of modulation of the parameters in the right side. Modulation sources include LFOs
and random generators. The joystick controllers at the top let the user manipulate
two parameters simultaneously.

I should mention here an essential feature that is characteristic of all my

granulators since 1988, and which amazingly seems to be missing in other
implementations: per-grain effects processing. Starting in 1988 I wrote
programs that scattered each grain emitted to an individualized point in
virtual space. I later I extended this so that every grain passes through a

14

separate constant-Q filter. Each filter has its own center frequency and
bandwidth, selected randomly within limits stipulated by the user. The
number of filters in operation corresponds to the density of grains per
second, which can be hundreds per second. In a similar per-grain fashion
we can pitch-shift, ring-modulate, etc. each grain individually with
different parameter settings for each grain. The resulting heterogeneity of
sound is the signature of truly granular signal processing. By comparison,
many granulators that feed the entire grain stream through the same effects
channel tend to sound flat and one-dimensional.
DICTIONARY-BASED PURSUIT:
ANALYTICAL COUNTERPART TO GRANULAR SYNTHESIS
The granular paradigm is, of course, not limited to stream, cloud, and
spray models. Indeed, it applies to the analysis or synthesis of any sound,
which Gabors pioneering papers demonstrated (1946, 1947, 1952).
Indeed, as Xenakis (1960) wrote:
All sound is an integration of corpuscules, of elementary acoustic particles, of
sound quanta.

Thus we can extend the granular paradigm to the realm of sound analysis
by means of an analytic counterpart to granular synthesis: dictionary-based
pursuit (DBP), also known as the family of matching pursuit algorithms
(Sturm, et al. 2009). DBP seeks by iterative search to match the energy in a
signal with a vast dictionary of millions of sound atoms or grains,
proceeding step-by-step from the strongest unit of time-frequency energy
to the weakest.

15

The result of a DBP analysis is a time-frequency representation of grains

that is highly malleable. This opens up a largely unexplored universe of
potential sound transformations based on granular processes, including
pluriphonic spatial processing at the granular level, where a sounds
elemental particles scatter in evanescent spatial patterns, or cavitation
processes carve new granular patterns out of existing ones.

Figure 9. Screen image of Scatter, a graphic interface for manipulating granular

time-frequency representations. The vertical axis represents frequency and the
horizonal axis represents time. The area in red has been picked up and moved from
its original location on the time-frequency plane.

The two primary code libraries for the analysis stage of DBP, Matching
Pursuit Toolkit (MPTK) (Krstulovic and Gribonval 2006) and LastWave
(Bacry 2008) were not designed for musical purposes. Our Scatter app
(Figure 9) was designed precisely to take the analysis data generated by
MPTK and make it artistically usable. Scatter provides a graphic user
interface for performing dictionary-based pursuit analysis (McLeran et al.

16

2008). The app displays the Wigner-Ville distribution and energy

contribution of each atom or grain of sound. (See Preis and Georgopoulos
1999 for more on the Wigner-Ville distribution.) The program offers
multiple ways of transforming the decomposition: by direct interaction via
mouse, by parametric filtering, and by application of stochastic algorithms.
Scatter also enables real-time performance through a variety of playback
and scrubbing techniques. For a video demonstration see note (4).
PHYSICAL AND BIOLOGICAL MODELS
A potentially compelling paradigm for granular organization derives from
the world of physics. Physical modeling synthesis starts from a
mathematical description of a mechanical process (Roads 1996; Fletcher
and Rossing 1991). By now, a large body of scientific literature centers on
the micromechanics of granular processes, such as grain vibration patterns,
mixing, flow, and grain/fluid interactions (Aronson and Tsimring 2009).
The basic premise of these simulations is that the rules governing the
behavior of one thing are not the same as those governing a million.
An early example of physical model control of granular synthesis was a
simulation of the sounds of shaken and scraped percussion: maracas,
sekere, cabasa, bamboo windchime, tambourines, sleighbells, and the guiro
(Cook 1996, 1997, 2007). Keller and Truax (1998) created granular
models of natural physical processes such as liquid streams and a bouncing
metallic ball. Natasha Barrett modeled processes of self-organized critical
systems such as grain avalanches in her work The Utility of Space (2000),
which was just one facet of a multilayered acousmatic design.

17

Such research could be taken much further. In recent years, physicists have
built sophisticated physical models of a broad range of granular
phenomena. These include ordered patterns such as ripples, avalanches,
and bands of segregated materials. Another class of patterns emerge out of
disordered excitations such as sifting, shaking, and scattering (Bideau and
Hansen 1993). Many other phenomena are characterized by clustering on a
variety of scales (Rivier 1993; Reynolds 1993).
In similar manner, scientists study self-organizing patterns and behavior in
the biological world of social insects, schools of fish, and swarms of birds
(Camarzine et al. 2001). Today the scientific modeling paradigm seems
ripe for exploration by both artists and scientists. These simulations can be
arbitrarily complex, and extend into virtual worlds in which events that
would be impossible in the real world can be simulated as easily as actual
events.
Unfortunately, the problem with physical/biological modeling has always
been the same: a model of an instrument/system is worthless without an
expert player, whether real or virtual. Virtuosity demands daily practice
over a period of years. Obviously, to develop a software model of an
expert player poses a daunting challenge. This goes beyond the question of
how a performer interprets a score. In physical/biological models there is
the need to control dozens or hundreds of low-level parameters in real time
in order to simulate a sequence of related gestures in a natural sounding
way. How does one play a physical/biological simulation? Of course,
one could give up the natural sounding gesture constraint, but then we
are left with a robot player whose gestures may not make much sense to
human beings.

18

ABSTRACT GENERATIVE MODELS

Beyond the physical and biological lies the abstract realm of experimental
generative algorithms for granular synthesis, not related to a model of a
real-world phenomenon. They are sometimes derived from formulae or
algorithms that are en vogue in a given period, such as recursive
substitution, cellular automata, fractal or chaotic functions. For example,
Alberto de Campo (1998) proposed a method of grain scattering in time
based on a recursive substitution algorithm. Others have employed chaotic
functions to scatter the grains in time (Moon 1987). Chaotic functions
vacillate between stable and unstable states, between intermittent transients
and turbulence (Di Scipio 1990; Gogins 1991, 1995; Miranda 1998;
Wolfram 2002).
These generally low-level algorithms emit granular sound as a byproduct
of a formal process that is imported from an abstract domain. A
fundamental problem with such bottom-up approaches is the fact they do
not incorporate a multiscale notion of structure. Like the low-level rules in
Wolframs (2002) experiments, whether larger scale structures emerge out
of such rules is mainly a matter of guesswork; in many cases they do not.
As a compositional strategy, reliance on enumeration of a formula is
constraining, a classic inside the box approach. In effect, the composer
invites us to listen to her clockworks. However, the human fascination for
monitoring a mechanism is limited. Abstract algorithms are not tethered to
human action or perception. To the detached observer who is not invested
in the algorithm itself, these processes tend to either sound either
predictable or merely random. They lack the meaningful sense of causality
and effort that we associate with human gestures, the opposition between

19

tension and resolution to which we are psychologically disposed. Human

beings are highly sensitive to the virtuosic play of lively human gestures,
whether live on stage or practiced behind the scenes in a studio, and to the
unfolding of a sonic narrative played out by recognizable sonic characters
or themes.
LIMITATIONS OF FORMALISM IN MUSIC COMPOSITION
We see here a broader philosophical issue opening up: the limitations of
formalism in music composition. As Jean-Claude Risset (2007) observed,
the opposition between music-as-formal-concept versus music-asperceived-sound is as ancient as Pythagoras (the formalist) and
Aristoxenus (the listener). Applied at different strata of compositional
organization, formal algorithms can be a potent means of invention. For
example, granular synthesis proves that the micro layers of sound
generation can be controlled effectively by powerful algorithms that spawn
masses of fine detail.
Compositional algorithms can also enumerate a collection of variations
quickly, offering the composer a wide range of selections from which to
choose. Interactive performance systems try to balance preprogrammed
automation with spontaneous decisions.
However, at the same time, we see many examples of generative systems
that produce unlimited quantities of unremarkable music. (5)
While formal algorithms enable interaction with a machine, absolute
formalism in composition means imposing constraints on one's self. The
strictly formalist composer follows a predetermined conceptual plan from

20

beginning to end. The plan must ultimately be translated into the real
world: acoustics, psychoacoustics, and emotional response. It is in this
translation that the game is often lost.
Musical formalism is related to but not identical to the conceptual art
tradition. In a famous essay on conceptual art, the artist Sol Lewitt (1967)
wrote:
In conceptual art the idea or concept is the most important aspect of the work.
When an artist uses a conceptual form of art, it means that all of the planning
and decisions are made beforehand and the execution is a perfunctory affair.
The idea becomes a machine that makes the art.

Rather than critically evaluating the result on a perceptual basis, another

set of aesthetic criteria govern such art: is the concept politically,
culturally, and socially meaningful? How does it reflect on, comment on,
or critique society? Many such installations take in audience feedback and
thus function like polls, surveys, or marketing exercises.
Music has a long conceptual tradition, but composition tends to gravitate
toward mathematical and algorithmic strategies rather than cultural probes.
Many algorithmic composers are motivated by formalist ideals of
conceptual purity, logical rigor, and algorithmic beauty. Generative
strategies are conceptually attractive. They offer the possibility of
exploring novel musical processes and the creation of new musical
structures from freely invented axioms. Another often stated rationale for
generative techniques is that they allow composers to reach beyond
themselves. These romantic ideals favor self-transcendence, purity of

21

method, and often incorporate seductive metaphors of natural growth and

evolution, artificial life, and holistic ecosystems.
However, elegant rules do not necessarily make elegant music. As a
skeptical Debussy observed (quoted in Risset 2004):
Works of art make rules, rules do not make works of art.

In certain masterworks, the composer escapes from the cage of selfimposed rule systems. For example, Xenakis treated the output of his
composition programs flexibly. In particular, he edited, rearranged, and
refined the raw data emitted by his Free Stochastic Music program (Roads
1973).
When I used programs to produce music like ST/4, ST/10, or ST/48, the output
sometimes lacked interest. So I had to change [it]. I reserved that freedom for
myself. Other composers, like Barbaud, have acted differently. He did some
programs using serial principles and declared: The machine gave me that so
I have to respect it. This is totally wrong, because it was he who gave the
machine the rule! Iannis Xenakis (Varga 1996).
Formalized music does not sound free, but it is. I wanted to achieve a
general musical landscape with many elements, not all of which were formally
derived from one another. Iannis Xenakis (1996).

In his later years, Xenakis no longer relied on computers for instrumental

composition; he had absorbed algorithmic strategies into his intuition
(Varga 1996; Harley 2004).
THE PROBLEM OF MULTISCALE ORGANIZATION

22

We return now to the main theme of this paper: the notion of organizing
sound grains into larger structures. Cloud, stream, and spray models have
been effective in composition because they agglomerate grains into
multiple levels of musical structure, specifically the object (100 ms-8 sec)
and meso (> 8 sec) time scales (Roads 2001b).
By contrast, physical/biological models and abstract algorithms of grain
generation pose puzzling challenges. How can one create coherent
multiscale structures with these techniques? As Wesley Smith (2011)
observed:
One of the major challenges in building a system that can increase in
complexity as it runs is figuring out how to transfer complex structures in a
lower level space into simple structures in a higher level space while still
maintaining the essential qualities that the complex lower level structure
represents.

This is one of the great unsolved problems in algorithmic composition.

The issue is not merely a question of scale, i.e., of creating larger sound
objects out of grains. It is a question of creating coherent multiscale
behavior extending all the way to the meso and macro time scales.
Multiscale behavior means that long-term high-level forces are as powerful
as short-term low-level processes. This is why simplistic bottom-up
strategies of emergent self-organization tend to fall short.
Xenakis speculated on abstract schemes for multiscale organization on
several occasions. In Formalized Music (1971) he conjectured on a kind of

23

multiscale clustering applicable to granular synthesis, where each point in

a cluster contains a sub-cluster (6):
Within human limits, using all sorts of manipulations of these grain clusters,
we can hope to produce not only the sound of the classical instruments and
elastic bodies, and those sounds generally preferred in concrete music but
also sonic perturbations with evolutions, unparalleled and unimaginable until
nowWe can even express a more general supposition. Suppose that each
point of these clusters represents not only a pure frequency and its satellite
intensity, but an already present structure of elementary grains, ordered a
priori. We believe that in this way a sonority of a second, third, or higher
order can be produced.

In Music composition treks (1985) Xenakis enumerated nine proposals;

six of these involved hierarchical schemes in which one stochastic process
determines macroform, while another low-level stochastic process
calculates the actual sounds.
It seems that a new kind of musician is necessary, that of the artist-conceiver
of free and abstract new forms, tending toward complications and
generalizations at several levels of sound organization. For example, a form,
a construction, an organization built on Markovian chains or on a complex of
interlocked probability functions may be carried over simultaneously onto
several levels of musical micro-, meso-, and macrocompositions.

His remaining proposals dealt with logical operations (Cartesian products

of sets, sieve theory, logical functions applied to sets of sound parameters),
which were previously used by the composer.

24

Of course, the problem of multiscale granular organization is not merely a

matter of building arbitrary clusters. For the music to be convincing, these
formations must articulate a compelling process that is essentially narrative
in function.
Multiscale composition with physical/biological models poses many
challenges. Most physical models describe a single instrument and are
gesture-driven. As we have already pointed out, a physical model is
worthless without an expert player of gestures. Even then, a player needs
to know what gestures to play. The challenge is to build game-like virtual
worlds

in

which

composer

can

interact

with

modeled

instruments/biosystems in such a way as to construct a meaningful

narrative, expressed as a pattern of gestures. For example, one can imagine
a kind of musical pinball machine supplied with thousands of grains,
where the levers are controlled by the composer. Certain levers could
model physical forces acting on the granular flows.
FORMAL/INFORMAL STRATEGIES
IN HEURISTIC ALGORITHMS
Music interacts in deep ways with the memory and expectations of
listeners (Huron 2006). Human beings respond intuitively to contextdependent cognitive impressions that are difficult to formalize, like wit,
irony, tension, surprise, virtuosity, humor, and clever twists and
transitions. One sound appears to cause another sound. Sounds converge
on points of attraction or scatter at points of repulsion. In general, formal
methods do not address these issues. How to codify these essentially
narrative functions based on human expectation? This is not obvious, but it
is clear that it is unlikely to emerge from a formula borrowed from an

25

arbitrary branch of physics or mathematics.

The effect of music is essentially magicalbeyond logical explanation. In
her fascinating book, Music, Science, and Natural Magic in Seventeenth
Century England, Penelope Gouk (1999) quotes the English philosopher
Francis Bacon (1561-1626), who defined magic as follows:
The science which applies the knowledge of hidden forms to the production of
wonderful operations.

In the context of the 17th century, the knowledge of hidden forms

involved mastery of esoteric skillsanalogous to the knowledge of
programming languages todaybut also how to apply them to the
production of wondrous operations. Criteria for the production of wonder
have never been formalized. However, we observe that certain highly
talented people make inspired choices from myriad possibilities to create
fascinating designs. This remains the strong suit of human talent.
For grains to evolve into wondrous forms, it appears that what is needed is
a hybrid formal/informal approach, combining the computational power of
algorithmic control with the magical influence of heuristics. What is
heuristic influence? Heuristics is the art of experience-based strategies for
problem solving:
Heuristic knowledge is judgmental knowledge, the knowledge that comes from
experiencethe rules that make up the art of good guessing. (Cohen et al.
1984).

The master chess player uses heuristic insight, in contrast to a computer

26

chess player that must employ brute-force search on millions of

possibilities at each move without any knowledge of previous moves.
Heuristic methods include rules of thumb, educated guesses, intuitive
judgments, and common sense. Heuristic methods are inevitably
intertwined with an understanding of context, whether it be the state of a
game, or the state of a composition.
Heuristic methods are compatible with formalization. However, in practice
they implement tailor-made solutions that are domain-specific and contextdependent, rather than imported whole cloth from one area of study to
another. For example, the visual artist Harold Cohen has long applied
heuristic algorithms to aesthetic problems. Over a period of forty years,
Cohen has been developing a body of highly specific algorithms for
drawing and coloring shapes. These algorithms do not attempt to draw like
a human being would.
Most importantly, heuristic algorithms are tested by experiments and
refined by human perceptual judgments. Xenakis used stochastic processes
in a heuristic manner, sometimes modifying and rearranging the results to
better suit the piece. Poetic license is the ultimate heuristic.
For granular synthesis, one heuristic approach would be to borrow certain
concepts from scientific models of granular processes but then rework
them to serve more effectively in a musical context. In this sense the
physical model serves as a kind of metaphor for granular organization in
music, rather than a strict model. The main point is that we can design
heuristic algorithms to implement methods of mesostructural formation
and multiscale behavior. These algorithms need to work according to
testing and expert judgment.

27

The principle of economy of selection

Hand-in-hand with the use of heuristic algorithms is one of the most
important issues in composition: the principle of economy of selection
(Roads forthcoming). Economy of selection means choosing one or a few
perceptually and aesthetically optimal or salient choices from a vast desert
of unremarkable possibilities. This choice relies on the powerful aesthetic
perception of an expert practitioner.
Making the inspired, intuitive choice from myriad possibilities remains the
exclusive domain of human talent. As Stuckenschmidt (1970) observed:
Bach was as well versed in the possible uses of the three mirror forms of a
melody as any Netherlands polyphonist of the fifteenth or sixteen century. He
did not omit to use one or another of them out of forgetfulness or a defective
grasp of the full range of possibilities. He knew that a two-part invention can
occupy only a limited amount of space. The ability to make the right choice
from the million or more possible forms is a creative secret that cannot be
uncovered by science or technology. Here, too, is where the astonishing
capabilities of computers prove to have limitations.

Long seen as a gift of the gods, inspired choice seems difficult to teach to
human beings and even more so to computers. Indeed, what makes a
choice inspired is hard to generalize, as it is particular to its context.
Sometimes it is the surprising or atypical choice, but other times it is
simply satisfying, optimal, or salient in a way that is not easy to
formalize. As Vaggione (2003) observed, reliance on formulas is not
adequate; direct action (non-formulaic singularities) are also needed:

28

To articulate a highly stratified musical flux is unthinkable using operations

based on statistical means. On the contrary, this requires an approach based
on singularities of discontinuity, contrast, and detail. This is why causal
formulas are problematic in composition if their automation is not
compensated by other modes of articulation, i.e., unique compositional
choicessingularitiesglobal as well as local, integrated explicitly in the
composition strategy.

In discussing inspired choice, it is not a question of idealizing either the

composer or the selection. Here optimal does not imply perfection; it is
simply a particularly satisfactory choice given the constraints. Indeed, it
would be hard to prove by scientific argument that a specific solution to a
musical problem is inspired, satisfying, or optimal. It may simply
satisfice, to use Herbert Simons (1969) term for sufficiently
satisfactory. Indeed, in many compositional decisions, more than one
choice would be equally effective, but the composer simply had to pick
one. Caprice is part of the composition process.
Economy of selection is an important concept because it emphasizes the
role of intuitive choice in all compositional strategies. Even in generative
composition, the algorithms are chosen according to subjective
preferences. The rules are inevitably loosely constrained or incorporate
randomness so as to allow many possible correct solutions. Computer
programs can solve for and enumerate many of these solutions, but
carefully picking the best or optimal solution is a human talent.
Moment-to-moment choices create our lives. I would go so far as to say
that the talent of a composer lies primarily in his or her ability to listen

29

and understand deeply enough to make optimal choices. This begins with
choosing the right compositional problems to solvea question of
strategy, tactics, tools, and materials.
CONCLUSION: THE GRANULAR PARADIGM REFUSES TO DIE
The granular paradigm refuses to die. As Gabor showed, it is a universal
representation for sound. The challenge for the composer has always been:
what to do with granular materials? How do we build meaningful
multiscale musical structures? What is the role of algorithms, and what is
the role of gestural control?
The cloud, stream and spray paradigms have served as effective tools
when combined with studio-based montage and micro-montage. Live
performance with granular instruments is a more open question. We have
not yet seen the Steinway of granular instruments nor the Cecil Taylor of
granulation. The key to success in performing with granular instruments is
the same as any other instrument: the development of virtuosity.
In the studio, graphical envelope control as in PulsarGenerator is an
excellent compromise between gestural interaction and the kind of detailed
micro control that can only come from scripts or code. Analysis-based
granular processing is still in the beginning stages, with much territory to
explore. Physical, biological, and abstract models of granular processes
have potential, but probably more as metaphors for heuristic algorithms
suited to specific compositional problems, rather than as full-fledged
scientific models of reality. After all, the power of software is its ability to
model not only known realities, but also fantastic imaginary worlds.

30

ACKNOWLEDGMENTS
I deeply thank Sharon Kanach (Forthcoming) for her detailed comments
on this manuscript. Thanks to Makis Solomos and Charles Turner for their
views on related historical matters. I would also like to deeply thank my
partners in granular research over the years: John Alexander, Alberto de
Campo, Garry Kling, Aaron McLeran, Bob Sturm, and David Thall. It has
been a privilege to collaborate with them.

31

NOTES
1. This was only the beginning of my inspiration. A key encounter was
experiencing the Polytope de Cluny in Paris eight times in the Fall of 1973.
2. Hear Bebe Barrons Mixed Emotions:
www.youtube.com/watch?v=Biqz1r2d_xY
3. The Constant-Q Granulator requires Mac OS9. The EmissionControl
prototype requires a non-Intel PowerPC processor running MacOSX 10.4
(Tiger).
4. See Scatter: www.youtube.com/watch?v=JgGR6VjTiaA&lr=1
5. I am an advocate of Stephen Wolframs important monograph A New
Kind of Science (2002). However, his WolframTones is a classic example
of a system that uses four billion cellular automata rules to produce
trillions of pieces of unremarkable music. See: tones.wolfram.com.
6. Xenakis GENDY system embodies the notion of chains of interlocked
probability functions, but he never applied this paradigm to granular
synthesis.

32

REFERENCES
Aronson, I., and L. Tsimring (2009) Granular Patterns. Oxford, Oxford
University Press.

Bacry, E. (2008) Lastwave: the signal processing command language.

Internet: www.cmap.polytechnique.fr/~bacry/LastWave.
Bideau, D., and A. Hansen (1993) Disorder and Granular Media.
Amsterdam, North-Holland.
Camarzine, S., J.-L. Deneubourg, N. Franks, J. Sneyd, G. Therauaz, and E.
Bonabeau (2001) Self-Organization in Biological Systems. Princeton,
Princeton University Press.
Cohen, H., B. Cohen, and P. Nii (1984) The First Artificial Intelligence
Coloring Book. Los Altos, William Kaufmann.
Cook, P (1996) "Physically informed sonic modeling (PhISM): percussive
synthesis." In L. Ayers and A. Horner, eds. Proceedings of the 1996
International Computer Music Conference. pp. 228-231.
Cook, P (1997) "Physically informed sonic modeling (PhISM): synthesis
of percussive sounds." Computer Music Journal 21(3): 38-49.
Cook, P (2007) Real Sound Synthesis for Interactive Applications.
Wellesley, A. K. Peters.

33

de Campo, A (1998) "Using recursive diminution for the synthesis of

micro time events." Unpublished manuscript.
Di Scipio, A. (1990) Composition by exploration of nonlinear dynamical
systems. Proceedings of the 1990 International Computer Music
Conference. San Francisco, International Computer Music Conference. pp.
324-27;
Di Scipio, A (1999) Synthesis of environmental sound textures by iterated
nonlinear functions. Proceedings of the 2nd COST G-6 Workshop on
Digital Audio Effects (DAFx99), NTNU, Trondheim, December 9-11,
1999. Internet:
www.iet.ntnu.no/groups/akustikk/meetings/DAFx99/discipio.pdf
Fletcher, N., and T. Rossing (1991) The Physics of Musical Instruments.
New York, Springer-Verlag.
Gabor, D (1946) "Theory of communication." Journal of the Institute of
Electrical Engineers Part 3, 93: 429-457.
Gabor, D (1947) "Acoustical quanta and the theory of hearing." Nature
159(1044): 591-594.
Gabor, D (1952) "Lectures on communication theory." Technical Report
238, Research Laboratory of Electronics. Cambridge, Massachusetts
Institute of Technology.
Gogins, M (1991) "Iterated function systems music." Computer Music
Journal 15(1): 40-48.

34

Gogins, M (1995) "Gabor synthesis of recurrent iterated function systems."

In R. Bidlack, ed. Proceedings of the 1995 International Computer Music
Conference. San Francisco, International Computer Music Association. pp.
349-350.
Gouk, P (1999) Music, Science, and Natural Magic in Seventeenth Century
England. New Haven, Yale University Press.
Harley, J (2004) Xenakis: His Life in Music. New York, Routledge.
Huron, D (2006) Sweet Anticipation: Music and the Psychology of
Expectation. Cambridge, The MIT Press.
Kanach, S. ed. (Forthcoming) Xenakis Matters. Pendragon: Hillsdale.
Keller, D., and B. Truax (1998) "Ecologically-based granular synthesis."
In M. Simoni, ed. Proceedings of the 1998 International Computer Music
Conference. San Francisco, International Computer Music Association. pp.
117-120.
Krstulovic, S., and R. Gribonval (2006) MPTK: Matching pursuit made
tractable. Proceedings of the IEEE International Conference on
Acoustics, Speech, and Signal Processing, Toulouse, France, Vol. 3, pp.
496499. New York, IEEE.
Lewitt, S (1967) "Paragraphs on conceptual art. Artforum 5(10): 79-83.

35

Logan, G (2000) The Indiana University School of Music: A History.

Bloomington, Indiana University.
McLeran, A., C. Roads, B. Sturm, and J. Shynk (2008) Granular methods
of sound spatialization using dictionary-based methods. In Proceedings of
the Sound and Music Computing Conference, Berlin, Germany. Internet:
smcnetwork.org/files/proceedings/2008/session2_number4_paper20.pdfhtt
p://smcnetwork.org/files/proceedings/2008/session2_number4_paper20.pd
f.
Miranda, E (1998) "Chaosynth: a cellular automata-based granular
synthesiser." Internet: website.lineone.net/~edandalex/chaosynt.htm.
Moon, F (1987) Chaotic Vibrations. New York, Wiley-Interscience.
Preis, D. and V. Georgopoulos (1999) Wigner distribution representation
and analysis of audio signals: an illustrated tutorial review. Journal of the
Audio Engineering Society 47(12): 1043-1053.
Reynolds, P., ed. (1993) On Clusters and Clustering: From Atoms to
Fractals. Amsterdam, North Holland.
Risset, J.-C (2004) The liberation of sound, art-science and the digital
domain: contacts with Edgard Varse. Contemporary Music Review
23(2): 27-54.
Risset, J.C (2005) Horacio Vaggione: toward a syntax of sound.
Contemporary Music Review 24(4-5): 287-294.

36

Risset, J.C (2007) Comments accompanying the lecture Fifty years of

digital sound for music. In C. Spyridis, et al. Proceedings of the SMC07
4th Sound and Music Computing Conference. Athens, National
Kapodistrian University of Athens. pp. 3-8.
Rivier, N (1993) Order and disorder in packings and froths. In D. Bideau
and A. Hansen, eds. Disorder and Granular Media. Amsterdam, NorthHolland.
Roads, C (1973) Analysis of the composition ST/10 and the computer
program Free Stochastic Music by Iannis Xenakis. Unpublished
manuscript.
Roads, C (1978) Automated granular synthesis of sound. Computer
Music Journal 2(2): 61-62. Revised and updated version printed as
"Granular synthesis of sound" in C. Roads and J. Strawn, eds. 1985.
Foundations of Computer Music. Cambridge, The MIT Press. pp. 145-159.
Roads, C (1991) Asynchronous Granular Synthesis. In G. DePoli, A.
Piccialli, and C. Roads, eds. 1991. Representations of Musical Signals.
Cambridge, The MIT Press. pp. 143-186.
Roads, C (1992) "Composition with machines." In J. Paynter, T. Howell,
R. Orton, and P. Seymour, eds. Companion to Contemporary Musical
Thought. London, Routledge. pp. 399-425.
Roads, C (1992-1997) "Design of a granular synthesizer." Unpublished
design documents.

37

Roads, C (1996) The Computer Music Tutorial. Cambridge, The MIT

Press.
Roads, C (1998) "The Creatovox synthesizer project." Unpublished
manuscript.
Roads, C (2001a) Sound composition with pulsars. Journal of the Audio
Engineering Society 49(3): 134-147.
Roads, C (2001b) Microsound. Cambridge, The MIT Press.
Roads, C. (2006) The evolution of granular synthesis: an overview of
current research. Presented at the International Symposium on The
Creative and Scientific Legacies of Iannis Xenakis. Toronto: University of
Guelph.
Roads, C (Forthcoming) Composing Electronic Music: A New Aesthetic.
New York, Oxford University Press.
Roads, C., and J. Alexander (1995) Cloud Generator Manual. Distributed
with the program Cloud Generator. Internet: www.create.ucsb.edu.
Simon, H (1969) The Sciences of the Artificial. Cambridge, The MIT
Press.
Smith, W (2011) Hierarchies of spaces: building from the bottom up.
Internet: moniker.name/worldmaking.
Solomos, M. (2006) The granular connection (Xenakis, Vaggione,

38

DiScipio). Presented at The Creative and Scientific Legacies of Iannis

Xenakis Symposium, Toronto. Unpublished.
Solomos, M., A. Soulez, and H. Vaggione, eds. (2003) Formel informel.
Paris. LHarmattan.
Stuckenschmidt, H (1970) Twentieth Century Music. New York. McGrawHill.
Sturm, B., C. Roads, A. McLeran, and J. J. Shynk (2009) Analysis,
visualization, and transformation of audio signals using dictionary-based
methods. Journal of New Music Research 38(4): 325-341.
Thall, D. 2004a. Experiments in sound granulation and spatialization for
immersive environments. Presented at the Interactive Digital Multimedia
IGERT Annual Research Review, sponsored by the National Science
Foundation. Santa Barbara: University of California.
Thall, D. 2004b. A Scheme for Generalized Control of Particle Synthesis.
MS thesis. Media Arts and Technology. Santa Barbara: University of
California.
Truax, B (1986) "Real-time granular synthesis with the DMX-1000." In P.
Berg, ed. 1987. Proceedings of the 1986 International Computer Music
Conference. San Francisco: Computer Music Association. pp. 138-145.
Vaggione, H (2003) Composition musicale et moyens informatiques :
questions dapproche. In Solomos, M., A. Soulez, and H. Vaggione, eds.
2003. Formel informel. Paris: LHarmattan. pp. 91-117.

39

Varga, B (1996) Conversations with Xenakis. London, Faber and Faber.

Wenger, E., and E. Spiegel (2004) MetaSynth 4.0 User Guide and
Reference. Redwood City: U&I Software LLC.
Wolfram, S (2002) A New Kind of Science. Champaign, Wolfram
Research.
Xenakis, I (1960) "Elements of stochastic music." Gravesaner Bltter 18:
84-105.
Xenakis, I (1971) Formalized Music. Bloomington, Indiana University
Press.
Xenakis, I (1985) Music composition treks. In C. Roads, ed. Composers
and the Computer. Madison, A-R Editions. pp. 169-192.
Xenakis, I (1992) Concerning time, space, and music. Chapter X of
Formalized Music, revised 1992 edition. Stuyvesant, Pendragon Press. pp.
255-267.

40