Klingbeil Dissertation Web

Spectral Analysis, Editing, and Resynthesis:
Methods and Applications
Michael Kateley Klingbeil
Submitted in partial fulfillment of the

requirements for the degree
of Doctor of Musical Arts
in the Graduate School of Arts and Sciences
COLUMBIA UNIVERSITY
2009
© 2008
All Rights Reserved
ABSTRACT
Spectral Analysis, Editing, and Resynthesis: Methods and Applications
This document describes the design, development, and applications of cross-platform
software for audio analysis, editing, and resynthesis. Analysis is accomplished using a
variation of the McAulay-Quatieri technique of peak interpolation and partial tracking.
Linear prediction of the partial amplitudes and frequencies is used to determine the
best continuations for sinusoidal tracks. A high performance user interface supports
flexible selection and immediate manipulation of analysis data, cut and paste, and un-
limited undo/redo. Hundreds of simultaneous partials can be synthesized in real-time
and documents may contain thousands of individual partials dispersed in time without
degrading performance. A variety of standard file formats, including the Sound Descrip-
tion Interchange Format (SDIF), are supported for import and export of analysis data.
Specific musical and compositional applications, including those in works by the author,
are discussed.
Contents
List of Figures iv
List of Tables vi
Acknowledgements vii
1 Introduction 1
1.1 Timbre and the Computer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Synthesis Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Theory Versus Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.1 Implementation History . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.2 Present Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2 Spectral Modeling 13
2.1 The Frequency Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.1 Sinusoids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.2 The Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.3 The Discrete Fourier Transform . . . . . . . . . . . . . . . . . . . . . 16
2.1.4 The Short-Time Fourier Transform . . . . . . . . . . . . . . . . . . . . 18
2.2 STFT Analysis and the Phase Vocoder . . . . . . . . . . . . . . . . . . . . . . 22
2.2.1 Putting Phase to Good Use . . . . . . . . . . . . . . . . . . . . . . . . 24
2.2.2 Resynthesis and Transformation . . . . . . . . . . . . . . . . . . . . . 25
2.2.3 Time Expansion/Compression . . . . . . . . . . . . . . . . . . . . . . 26
2.2.4 Transposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.2.5 Refinements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.3 Spectral Envelopes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
i
2.4 Sinusoidal Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4.1 STFT Frequency Interpolation . . . . . . . . . . . . . . . . . . . . . . 32
2.4.2 Partial Tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.4.3 Improved Partial Tracking . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.4.4 Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.4.5 Limitations and Extensions . . . . . . . . . . . . . . . . . . . . . . . . 41
3 Sinusoidal Analysis and Resynthesis 42
3.1 Analysis Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.1.1 Analysis Window . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.1.2 Amplitude Thresholds . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.2 Partial Tracking Using Linear Prediction . . . . . . . . . . . . . . . . . . . . 46
3.3 Resynthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.3.1 Inverse FFT Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.3.2 IFFT Synthesis Implementation . . . . . . . . . . . . . . . . . . . . . . 52
3.3.3 Oscillator Bank Resynthesis . . . . . . . . . . . . . . . . . . . . . . . . 53
3.3.4 Phase Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.4 Data Storage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.4.1 Time-span Frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.5 Transient Sharpening . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.6 Noise Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.6.1 Bandwidth Enhanced Sinusoids . . . . . . . . . . . . . . . . . . . . . 62
4 Editing and Transformation 64
4.1 Graphical User Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.2 Selection Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.2.1 Rule-based Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.3 Editing Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.4 Time Expansion and Contraction . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.5 Frequency Modifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
ii
4.6 Data Exchange . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5 Compositional Applications 78
5.1 Spectral Composition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.2 Sonic Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.2.1 Timbre Hybridization . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.2.2 Spectral Tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.3 Software Composition Environments . . . . . . . . . . . . . . . . . . . . . . 85
5.3.1 OpenMusic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.3.2 Max/MSP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.3.3 Common Music . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.4 Compositional Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.5 Performance Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
6 Conclusions 97
References 101
A Tear of the Clouds score 107
A.1 Program Note . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
A.2 Score . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
iii
List of Figures
N
2.1 DFT magnitude spectrum of a sinusoid with period 4 . . . . . . . . . . . . 18
N
2.2 DFT magnitude spectrum of a sinusoid with period 4.21 . . . . . . . . . . . 19
2.3 Sinusoids of 4.21 and 4.0 cycles per N samples . . . . . . . . . . . . . . . . . 19
2.4 DFT magnitude spectrum and the interpolated underlying magnitude spec-
N
trum of a sinusoid with period 4.21 . . . . . . . . . . . . . . . . . . . . . . . . 21
2.5 DFT magnitude spectrum and the interpolated underlying magnitude spec-
N
trum of a sinusoid with period 4 . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.6 Rectangular window . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.7 Triangular window . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.8 Hann window . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.9 Blackman window . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.10 Overlap-adding the product of Blackman analysis and synthesis windows
with varying overlap factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.11 Spectral envelope of a soprano voice . . . . . . . . . . . . . . . . . . . . . . . 30
2.12 Phase vocoder analysis of a frequency sweep . . . . . . . . . . . . . . . . . . 32
2.13 Sinusoidal modeling analysis of a frequency sweep . . . . . . . . . . . . . . 32
2.14 Estimating a peak in the spectrum based on a parabolic curve fitting . . . . 33
2.15 Connecting STFT peaks into sinusoidal tracks . . . . . . . . . . . . . . . . . 35
2.16 Partial tracking conflicts due to glissandi . . . . . . . . . . . . . . . . . . . . 37
3.1 Magnitude spectrum of the main lobes of two Blackman windows with
complete separation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.2 Magnitude spectrum of the main lobes of two Blackman windows with the
minimum allowable frequency separation . . . . . . . . . . . . . . . . . . . . 44
iv
3.3 Frequency dependent threshold curve with the default values . . . . . . . . 46
3.4 Comparison of sinusoids generated via IFFT synthesis . . . . . . . . . . . . 52
3.5 Family of cubic phase interpolation functions with different values of M . . 55
3.6 Data structure for division of partials into time-span frames . . . . . . . . . 57
3.7 Breakpoints from the analysis of a castanet click using the standard model 60
3.8 Breakpoints from the analysis of a castanet click using time reassignment . 60
4.1 Zoomed-out display . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2 Zoomed-in display . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.3 The SPEAR tool palette . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.4 Playback control sliders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.5 Different breakpoint storage models . . . . . . . . . . . . . . . . . . . . . . . 69
4.6 Original partials prior to time expansion . . . . . . . . . . . . . . . . . . . . 71
4.7 Partials time expanded in independent mode . . . . . . . . . . . . . . . . . . 71
4.8 Partials time expanded in proportional mode . . . . . . . . . . . . . . . . . . 71
4.9 Dialog boxes for transposition and frequency shifting . . . . . . . . . . . . . 72
4.10 Frequency flipping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.11 The par-text-frame-format specification . . . . . . . . . . . . . . . . . . . . 76
4.12 Sample data as par-text-frame-format . . . . . . . . . . . . . . . . . . . . . 76
4.13 The par-text-partials-format specification . . . . . . . . . . . . . . . . . . 77
4.14 Sample data as par-text-partials-format . . . . . . . . . . . . . . . . . . . 77
5.1 OpenMusic patch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.2 OpenMusic notation editor . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.3 OpenMusic patch that reads and writes SPEAR data . . . . . . . . . . . . . 88
5.4 Max/MSP patch from the CNMAT Spectral Synthesis Tutorials . . . . . . . 89
5.5 Dynamic retuning with crossing partials . . . . . . . . . . . . . . . . . . . . 95
v
List of Tables
1.1 Digital synthesis techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
4.1 Some of the standard SDIF matrix types . . . . . . . . . . . . . . . . . . . . . 74
4.2 Supported file formats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.1 Transformation functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.2 Frequency adjustment functions . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.3 Spectral envelope functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
vi
Acknowledgements
I wish to thank my advisor Tristan Murail and committee members Brad Garton and Fred
Lerdahl for their insight and support throughout my time at Columbia University (and
beyond). Thanks also goes to James Beauchamp for introducing me to the details of signal
processing, to Rick Taube for revealing the realm of the metalevel, and to Karl Klingbeil
for opening my ears to the world of contemporary music. Special gratitude goes to Anita
Buckmaster for her unwavering encouragement and assistance.
This dissertation is dedicated to my parents, Laura Kateley and Jack Klingbeil.
September 17, 2008
vii
1
1. Introduction
1.1 Timbre and the Computer
Considerable energy and effort has been applied to further the understanding of timbre,
one of music’s most elusive features. The American National Standards Institute defines
timbre in a purely exclusionary manner: “. . . that attribute of sensation in terms of which
a listener can judge that two sounds having the same loudness and pitch are dissimilar.”
(ANSI S1.1-1960(R1976)-12.9) This rather unsatisfying description essentially relegates all
aspects of sound other than pitch and loudness to a single perceptual category. Since at
least the beginning of the twentieth-century, the conception of timbre has taken on an
increasingly important role in Western compositional practice. In 1922 Arnold Schoenberg
asserted, “The evaluation of tone color, the second dimension of tone, is in a much less
cultivated, much less organized state than is the aesthetic evaluation of pitch. Nevertheless,
we go right on boldly connecting sounds with one another, contrasting them with one
another, simply by feeling. . . ” (Schoenberg 1983). More than eight decades later, this
statement still seems regrettably accurate. Schoenberg goes on to call for a theory of timbre,
analogous to pitch theories, that will enable the composition of “tone color melodies.”
While falling far short of a unified theory, scientific and artistic efforts of recent decades
have achieved a greatly expanded understanding of timbre. For example, while Schoenberg
describes tone color as a singular “second dimension” of tone, there is now strong evidence
suggesting that timbre is in some sense multi-dimensional (Grey 1977). It is now theorized
that sensations of pitch and timbre are closely intertwined, and that we likely experience
different classes of pitch sensation (spectral pitches and virtual pitches) (Terhardt et al.
1. Introduction 2
1982). What is undeniable, whether a composer grounds her work in theory, or spirituality,
or chooses to work “simply by feeling,” is that technological resources have had a profound
impact on recent musical practice, and that an understanding of timbre is essential to
effectively harnessing these resources.
The application of the digital computer, in particular, has done much to expand the
practical range of sonic possibilities and also to enable the theoretical and empirical study
of timbre. In a prophetic lecture given in 1936, Edgard Varèse proposed a vision of music
made possible by machine:
. . . the new musical apparatus I envisage, able to emit sounds of any number
of frequencies, will extend the limits of the lowest and highest registers . . .
Not only will the harmonic possibilities of the overtones be revealed in all
their splendor, but the use of certain interferences created by the partials
will represent an appreciable contribution. . . . The time will come when
the composer, after he has graphically realized his score, will see this score
automatically put on a machine that will faithfully transmit the musical content
to the listener. (Varèse and Chou 1966, 12)
Varèse was hardly the first, nor last, to make such predictions, and yet the strength of
his conviction and the precision of his forecast remain striking. By 1953, when Varèse
would begin his first compositional experiments with magnetic tape, numerous composers,
technicians, and researchers were already at work in studios throughout the world ex-
ploring the musical applications of electronic instruments. By 1957, with the first musical
applications of general purpose digital computers at Bell Labs, the stage was set for the
eventual realization of Varèse’s predictions.
It was in fact the advent of digital sound synthesis that made this vision possible.
Computer music pioneer Max V. Matthews understood the power and generality of digital
synthesis which he described in his 1963 article in Science: “With the aid of suitable output
equipment, the numbers which a modern digital computer generates can be directly
converted to sound waves. The process is completely general, any perceivable sound can be so
produced.1 ” (Matthews 1963, 553). For the composer seeking to explore a world of timbral
possibilities, the notion is seductive. And yet the computer music composer, whether
1 My emphasis
1. Introduction 3
working in 1967 or forty years later, still faces a number of formidable challenges. We can
summarize these particular problems as follows: the time/cost problem, the synthesis
problem, the control problem, and the compositional problem. Let us briefly consider each
of these issues in turn.
Digital music production distinguished itself from all previous musical-mechanical
approaches in that sound creation could occur on a time scale completely distinct from the
temporal evolution of the sound itself. At first these computations were ponderous (many
orders of magnitude slower than real-time), but as computational power increased over the
decades, reasonably complex and credible sounds could be computed faster than real-time.
The implication of this temporal decoupling is that sound complexity is no longer bound
by space constraints (the number of musicians one can fit on a stage, or the number of
inputs on a mixing board, for example) but by limitations of time (and/or expense). The
composer now has the capability, as Varèse envisioned, to “emit sounds of any number
of frequencies,” and yet at the same time now grapples with a continual “compromise
between interest, cost, and work” (Matthews 1963, 555). One could argue that this problem
is being tackled with continual efficiency improvements in computer hardware. Compared
to the IBM 704 of 1957, the laptop computer of the early twenty-first century provides
a five-hundred-thousand fold increase in processing power at at fifteen-hundred times
less cost (Sandowsky 2001). Nevertheless, computer musicians seem to have a particular
penchant for using spare processing power.
With decreased costs and staggering increases in computing speed, significant addi-
tional challenges still remain. Although “any conceivable sound” is certainly possible, in
reality, the range of potential sounds is constrained by the development of appropriate
synthesis algorithms. The ability to formulate these procedures is severely limited by our
understanding of timbre and psychoacoustics (Matthews 1963). This can be classified as
“the synthesis problem.”
The control problem is a closely associated concern. Although digital synthesis “allows
composers freedom of expression at the timbral level” it also entails the “additional
burden of specifying the microstructure of sound, which was formerly the domain of the
1. Introduction 4
performing musician” (Loy 1989, 326). Because interesting musical sounds are rarely static,
nor entirely stochastic, the precise control parameters must be specified at every instant.
If this control information is not a natural result of the synthesis procedure in use, the
result is an unwieldy data explosion with separate detailed control streams for each sonic
parameter of interest.
Finally, the composer is faced with all of the traditional problems of composition
itself — matters of intention and aesthetics, as well as possible questions of melody, coun-
terpoint, rhythm, harmony, and/or form.
The following list summarizes the challenges and some possible approaches:
Challenge Possible Approaches

I. Time and Cost faster computers, higher processing density, more
efficient algorithms
II. Synthesis problem new synthesis methods, flexible software systems
III. Control problem new user interfaces (visual, gestural), computer
assisted composition, machine learning/listening
IV. Compositional problem continued exploration, theory
It is furthermore recognized that these problems are inherently interrelated. A given
composition aesthetic, one that involves indeterminacy, for example, would invite a
particular approach to the specification of control information. Particularly with regard to
the compositional “problem,” it is certain that no ultimate “solution” exists. If it did, such
a discovery would spell the end of creative artistic production and individual expression.
The act of composition must be seen as a continual search, for which each artist may be
free to define their own goals and strategies.
The chapters that follow are concerned in large part with particular approaches to the
synthesis problem, with a discussion of the resulting implications for the cost and control
problems. The motivation will always be oriented towards musical goals, which will be
addressed in the penultimate chapter.

1. Introduction 5
1.2 Synthesis Methods
Regardless of compositional approach the practitioner of computer music always faces
the same question: What sounds should I produce? Although current computer music
systems offer tremendous flexibility, the choices are constrained by current theories and
their implementations. In order to contextualize the work described in subsequent chapters,
it may prove instructive to consider a particular taxonomy of digital synthesis techniques
developed by Smith (1992). Table 1.1 maintains the original categories while adding some
more recent methods. A significant development in the past decade is the emergence of
hybrid approaches.
Time Domain Spectral Models Physical Models Abstract Methods

Methods (Frequency Domain)
Sampling Additive Waveguide FM
Concrète Phase Vocoder Finite Element AM
Granular Sinusoidal (MQ/PARSHL) (CORDIS-ANIMA) Waveshaping
Scanned Synthesis SMS (Smith & Serra) Karplus-Strong Analog simulation
Subtractive Modal (VCO, VCA, VCF)
Source-filter (LPC) Chaotic systems
UPIC (Xenakis)
FOF/FOG/PSOLA
Wavetable
Particles (Roads)
Concatenative
Audio Mosaicing (MPEG-7)
Physically Informed Sonic Modeling (PhISM)
Table 1.1: Taxonomy of current digital synthesis techniques (after Smith 1992).
Let us first consider some of the most commonly used synthesis methods. Since the
early 1990s, sampling synthesis has been the predominant sound production method
used in commodity hardware and software. Sampling offers very high fidelity results
(limited only by the quality of the input), but is restricted in terms of expressive and
transformational possibilities. A typical sampling synthesis architecture limits transfor-
mations to amplitude envelopes, filtering, and loop control. Commercial developers are
attacking the limitations of sampling in a brute force manner. It is not uncommon to
find multi-gigabyte sample libraries containing complex mappings, cross-faded overlays,

1. Introduction 6
and keyswitch programs. For example, a complex violin bank might include samples
for legato, marcato, spiccato, staccato, and tremolo bowings all recorded at a variety of
different dynamic levels. Fast secondary storage allows these massive sample banks to be
streamed from disk in real-time.
Granular synthesis may be viewed as a particular specialization of sampling synthesis.
The granular approach overlaps many short sound segments (generally less than 50 msec.
duration) which may be derived from recorded sources or produced via other synthetic
means. Usually, each sound segment is modulated by a corresponding amplitude envelope
that quickly fades-in, reaches a steady state, and then fades out. Granular synthesis offers
the possibility to decouple the time and frequency evolution of sound, as well as impart
particular characteristics modulating between rough, smooth, or stochastic textures.
Wavetable synthesis, one of the oldest techniques, can be viewed as a type of hybrid
approach. The time domain view concentrates on the particular wave shape of the table,
which may be either derived mathematically, drawn interactively, or (most commonly)
extracted from a recorded sample. In the frequency domain perspective, the contents of
the wavetable are specified in terms of the amplitude, frequency, and phase of a set of
harmonically related partials. With both approaches, wavetable interpolation is necessary
to produce lively and dynamic sounds.
Particle synthesis may be viewed as a hybrid extension of granular synthesis that
attempts to build up specific frequency domain content using a variety of elementary sound
types which might include grains, impulses (trainlet or pulsar synthesis), or chirps (glisson
synthesis). Convolution may be used to impart specific frequency domain characteristics
to the “pulslet” texture. For a more detailed discussion of these methods, consult Roads’s
Microsound (2001, ch. 4).
Frequency domain models are concerned first and foremost with the spectral content
of sounds. This is a particularly powerful technique because the perception of timbre
is closely related to spectral content. Auditory research shows that the ear behaves as a
sort of spectrum analyzer with precise physical positions along the basilar membrane of
the inner ear attuned to specific frequencies. Frequency domain methods may thus be
1. Introduction 7
seen as way to model synthesis on the processes of audition. Frequency domain models
usually follow an analysis/synthesis paradigm in which control parameters are derived by
analyzing pre-existing recordings.
Additive synthesis is one of the oldest and still most flexible of these frequency domain
methods. Oscillators tuned to specific frequencies, each with time varying amplitudes, are
mixed (summed) together to form a composite timbre. Typically, each oscillator produces
a sine-like waveform. In some applications the frequency and/or phase2 of each oscillator
may also be time variant. While theoretically capable of generating almost any type of
sound, additive synthesis is particularly unattractive in light of the control problem. A
complex timbre may easily require 100 individual oscillators that, in turn, each require
their own amplitude and frequency functions. This problem is typically solved by deriving
synthesis data from an analysis stage. The major advantage of additive synthesis, and
frequency domain methods in general, is that they are highly amenable to transformations.
The phase vocoder has been particularly favored for its support for independent time
and frequency modification. Andy Moorer’s Lions Are Growing, a work from 1978 that
time stretches, compresses, and transposes speech sounds,3 offers a clear demonstration
of the artistic potential of this approach. Beginning in the early 1990s, the technique
became possible on commodity computing platforms and received widespread use in
both commercial sound design and music composition. Spectral models, with particular
emphasis on sinusoidal models, will be taken up in further detail in subsequent chapters.
Concatenative synthesis, a category that encompasses a variety of specific techniques,
may be viewed as a hybrid of spectral and time domain methods. For example, attack
transients might be most effectively handled with sampling, while steady states could
be executed with a spectral model. By carefully concatenating4 sound segments derived
from a large database, more convincing and dynamic note-to-note transitions and timbral
modulations can be achieved than with sampling or spectral synthesis alone. A number of
2 The phase of any non-zero frequency oscillator is by definition time varying — what is emphasized here is
that phase may be modulated directly rather than being a result of a fixed or time varying frequency.
3 The source sound is a reading of Richard Brautigan’s poem of the same name. The work employs
LPC-style resynthesis coupled with phase vocoding.

4 Generally trying to preserve phase continuity
1. Introduction 8
recently released commercial products such as Synful (orchestral instrument synthesis)
(Lindemann 2007) and Vocaloid (singing voice synthesis) (Bonada and Serra 2007) have
demonstrated the viability of this technique. A more detailed overview of concatenative
methods may be found in Schwarz (2005).
In the early 1990s, Julius O. Smith predicted “abstract-algorithm synthesis seems
destined to diminish in importance,” (Smith 1992) and although this proved true to some
extent, a resurgent interest in classic analog and digital hardware has demonstrated that
these methods retain a strong following. As a result, considerable efforts in both the
commercial and academic sectors have gone toward the software emulation of classical
voltage controlled oscillators and filters.5 In recent years the “vintage emulation” trend
has extended to early commercial digital synthesizers, including the venerable Yamaha DX
series which has been replicated in a number of software packages.6
It seems certain that no particular technique will ever disappear entirely from the
toolbox, rather the collection will expand with the exigencies of fashion dictating the
most popular methods of the time. The flexibility and modularity of current software
systems suggests that hybrid synthesis methods will become increasingly important. The
somewhat arbitrary boundaries between approaches will blur as practitioners explore
possible combinations and concatenations. After all, sampling, wavetable, granular, and
pulsar synthesis differ solely on their operative time scales (and approach to periodicity)
and thus could be subsumed under a single methodology.
Physical modeling represents a considerably different approach to sound production.
While sampling, spectral, and abstract models are all understood in terms of their effect
on the sound consumer (the ear), physical modeling is concerned with the details of the
sound producer (the instrument) (Smith 1992). The physics of musical instruments serve
as the starting point. Generally a set of differential equations are used to describe the
possible motion (vibrations) in response to physical input (such as the motion of a bow or
the stroke of a mallet). A virtual sound pressure wave can then be computed. A significant
5 The interested reader might wish to consult Välimäki and Huovilainen (2006) as well as Stilson (2006).
6 Most notably with Native Instruments FM-7 series (now FM-8 in its 2007 incarnation)
1. Introduction 9
benefit is that a very rich and sonically convincing output can result from simple control
inputs.
Despite many compelling advantages, physical modeling has yet to become perva-
sive in computer music composition. The lack of generality in many physical modeling
implementations may be one impediment to their acceptance. Even with a pre-designed
algorithm, the parameter space may be undesirably abstract and difficult to control. For
example, in a woodwind model the output pitch may be a result of multiple inputs. Partic-
ular physical modeling methods (for example waveguides) may be limited to modeling
certain restricted classes of instruments (for more detail see Smith 2004). Finite-difference
time-domain methods (Cadoz et al. 1993) represent a more general approach to physical
modeling, although the computational demands are still somewhat prohibitive for many
systems of musical interest. It is certain, however, that with continued research and
development and inevitable increases in processing speed, both this and other physical
modeling methods will find a following. After all, one must consider that even in the
relatively short history of computer music, these methods are still fairly new.
1.3 Theory Versus Implementation
Regardless of the technique of interest, the computer music practitioner is always limited
by the tools and implementations at hand. In some cases the constraints of the tools
may prove compositionally beneficial in their circumscription of the available musical
materials. At other times, the realities of the tools (missing features, bugs, awkward
interface, sluggish operation) may prove merely frustrating. Too often one finds a lack of
available implementations of some of the most compelling synthesis and signal processing
techniques described in the literature. In such situations computer music composers may
find it necessary to build their own tools — be it in software or hardware. Computer
music is an inherently multi-disciplinary endeavor involving researchers, mathematicians,
programmers, engineers, theoreticians, composers, improvisers, and performers, and in
most cases one individual may be closely tied to a number of these disciplines.
1. Introduction 10
Over the past several decades, many such individuals have devoted considerable
attention to the area of spectral modeling. The proliferation of various spectral modeling
implementations attests to its appeal. Some of the technique’s advantages (noted above)
are especially attractive to composers — in particular the possibility to explore timbre
in a manner that is directly related to real-world experience (unlike abstract algorithm
synthesis). The transformative potential of spectral modeling synthesis (largely unavailable
with sampling) opens up a vast space between purely mimetic or purely synthetic sounds.
Sinusoidal modeling has a proven track record of high quality resynthesis while offering
numerous possibilities for novel sonic transformations (Pampin 2004). The technique
(which will be described in detail in section 2.4) derives a detailed additive synthesis
model from an input sampled sound signal. Something that closely resembles the original
input sound (a resynthesis) can be generated by adding together a sufficient number of
individual sinusoidal waves. In most cases the resynthesis will not be identical to the
original sound, although it is possible to get very close and, with certain extensions to the
basic technique, it is possible to recreate the original exactly.
1.3.1 Implementation History
A complete implementation of sinusoidal modeling was first detailed in paper by Robert J.
McAulay and Thomas F. Quatieri (1986), who were working at MIT on the analysis
and synthesis of speech signals. The basic technique is often referred to as the MQ
method. This work was closely followed by the development of the PARSHL system for
analysis-resynthesis of musical tones by Julius O. Smith and Xavier Serra (1987) at Stanford
University. Serra extended this into the “Spectral Modeling Synthesis” (SMS) method
which added residual noise modeling to the basic sinusoidal analysis (Serra 1989). At the
same time Robert C. Maher (1989), working with James Beauchamp at the University of
Illinois, was using sinusoidal modeling for polyphonic voice separation. Much of Maher’s
code was subsequently integrated into the SNDAN suite (Beauchamp 1993).
1. Introduction 11
Maher’s work also served as the basis for Lemur by Kelly Fitz, Lippold Haken, and
Bryan Holloway (Fitz et al. 1995). Lemur, which ran on the Macintosh, was one of the
first graphical user interfaces (GUIs) available on commodity hardware for interactively
viewing and editing sinusoidal models. Its release was closely followed by the availability
of other GUI editors including SMSTools from Xavier Serra (running on Windows), InSpect
and ReSpect from Sylvain Marchand and Roberty Strandh (running on UNIX/X Windows),
and a number of tools from IRCAM including Diphone (Rodet and Lefèvre 1997), Sview,
and XSpect. Sadly, Lemur was never updated for newer MacOS versions. The original
version of SMSTools was discontinued and then relaunched as part of the cross-platform
CLAM project (Amatriain et al. 2002). Some of the IRCAM tools were made commercially
available as part of the IRCAM Forum software suite. In 1999, Juan Pampin developed the
ATS system which implemented sinusoidal and noise modeling in Common Lisp. A GUI
editor, ATSH by Oscar Pablo Di Liscia and Pete Moss soon followed.
1.3.2 Present Work
These packages are important predecessors of the software that will be detailed in this
document: SPEAR (Sinusoidal Partial Editing Analysis and Resynthesis). SPEAR was
created out of a desire for software that was faster, more flexible, and easier to install and
use than other options.
The following goals were kept in mind throughout the design and development phases:
editing should be as fast and as easy to understand as in a time domain waveform ed-
itor; listening to transformations should be possible with no intermediate synthesis or
processing stage; and high quality analyses should require only a few parameter set-
tings. Additionally, it was desired to have interoperability with other analysis-resynthesis
software by providing both SDIF (Wright et al. 1999a) and native format data exchange.
In order to offer a familiar and comfortable interface, SPEAR was written to run using
the native graphics of the host operating system. Development was founded on the
principle that the interface should be as close as possible to that of a first-class native
1. Introduction 12
application, with little or no evidence of compatibility or GUI emulation layers. Portability
is made possible using wxWidgets (http://www.wxwidgets.org), a C++ library that allows
the software to be compiled for MacOS, Windows, and GTK Linux (Zeitlin 2001). SPEAR
was first demonstrated at the Columbia University Computer Music Center on February
23, 2004. As of this writing, builds have been created MacOS 9, MacOS X, and Windows.
The following chapters will detail the design and evolution of this software. Chapter
2 will discuss the basic techniques of spectral and sinusoidal modeling, while chapter 3
will focus specifically on sinusoidal modeling and its implementation in SPEAR. Chapter
4 details the user interface design and implementation, and chapter 5 explores some
composition applications of sinusoidal modeling.

13
2. Spectral Modeling
2.1 The Frequency Domain
Spectral modeling is concerned primarily with the frequency domain content of audio
signals, as opposed to the time domain content. In the time domain, an audio signal may
be viewed as a function of time, x (t), with instantaneous amplitude as the function’s
range. Typically such a function is graphed with time on the abscissa (horizontal axis)
and amplitude on the ordinate (vertical axis), which results in the familiar waveform
representation. In the frequency domain, the same signal is represented as a new function,
X ( f ), where f corresponds to frequency. Given any frequency value f k , the value X ( f k )
can be used to determine relative strength of that frequency in the given signal.1 The
function X ( f ) defines the spectrum of the sound.
2.1.1 Sinusoids
In order to study frequency domain representations in more detail, the notion of frequency
must be made more explicit. Any function x (t) that has a precisely repeating shape is said
to be periodic. In musical applications, and signal processing applications in general, the
periodic functions of greatest interest are the trigonometric functions sine and cosine. A
sine wave signal with a frequency of f hertz is given by the following: x (t) = sin(2π f t)
(where t is in seconds).
Because of their relation to the unit circle, the sine and cosine functions have a period-
icity of 2π. Furthermore, the cosine function may be viewed as a phase shifted form of the
1 As we will define it, X ( f k ) is a complex value, so it does not directly give the amplitude of component
with frequency f k . However it is straightforward to derive it from the complex value.
2. Spectral Modeling 14
sine function. We use the term sinusoid to denote any any signal that can be described by a
function x (t) of the following form:
x (t) = A sin(2π f t + φ) (2.1)
where A = amplitude, f = frequency (in hertz), t = time (in seconds), and φ = initial phase
1
(in radians). Note that this function is periodic with a period of seconds.
f
2.1.2 The Fourier Transform
The Fourier transform provides an explicit method for converting from the time domain
to the frequency domain. The following defines the Fourier transform of x (t):
Z∞
X( f ) = x (t)e−i2π f t dt (2.2)
−∞
A complete examination of the Fourier transform is beyond the scope of this document,
but we shall consider the transform with an eye toward an intuitive and pragmatic
understanding.
Recall that x (t) represents a given time domain signal, and that evaluating the integral
for any desired frequency value f allows us to find the contribution of frequency value f in
x (t). In order to understand how frequency is defined, we must explore more closely the
inner portion of the integral, e−i2π f t . By definition, e is the base of the natural logarithm
√
(= 2.7182818 . . .) and i is the imaginary number −1. Euler’s relation is a remarkable
mathematical identity that relates these constants to the sine and cosine functions:
eiθ = cos θ + i sin θ (2.3)
Note that cos θ + i sin θ is a complex value (it has both real and imaginary parts). The
expression e−i2π f t is a complex exponential that defines a complex sinusoid with frequency f .
More generally, a complex sinusoid is given by the following (compare with the definition
of a real sinusoid in equation 2.1):
x (t) = Aei(2π f t+φ) (2.4)

Complex sinusoids are particularly useful because they are able to encode phase shifts in
a manner that is compact and easy to manipulate (Puckette 2007, 189). For example, given
the complex sinusoid of equation 2.4, the properties of exponents allow us to encode both
the amplitude and the initial phase shift φ as a single complex amplitude (also known as a
phasor). Any complex amplitude A0 can take the following form (for some constant phase
value φ):
A0 = Aeiφ (2.5)
and our complex sinusoid now simplifies as follows:
x (t) = Aei(2π f t+φ)
= Aei2π f t eiφ
= Aei(2π f t+φ)
= A0 ei2π f t
Other important results of Euler’s formula are the following relations:
eiθ − e−iθ eiθ + e−iθ

sin θ = cos θ =
2i 2
This tells us that real valued sinusoids are in fact a sum of two complex sinusoids, one
with positive frequency and the other with negative frequency. Complex sinusoids are
therefore simpler to manipulate because they have only a single frequency (either positive
or negative).
Given Euler’s identity, we may now re-express the Fourier transform as:
Z∞
X( f ) = x (t)(cos 2π f t − i sin 2π f t)dt (2.6)
−∞
To gain some intuitive sense of what this implies, consider the evaluation of X ( f ) for some
value f k . Informally, we may consider cos 2π f k t and sin 2π f k t to be “reference” or “basis”
sinusoids (each with freqeuncy f k ) that are multiplied by the time domain signal. The
multiplication of x (t) with one of the reference function results in a new function of time,
which can be plotted as a curve. We can think of the integral as giving the total area under
the curve. The area indicates the amount of the reference sinusoid present in the signal.
X ( f k ) is in fact a complex amplitude that encodes amplitude and initial phase (as shown
in equation 2.5).
2.1.3 The Discrete Fourier Transform
In the given definition of the Fourier transform, the limits of integration are ±∞, implying
signals of infinite duration. In real-world situations, we work with audio signals that, by
necessity, are of finite duration and are discretely sampled. Rather than working with the
continuous Fourier transform, we can calculate a discretely sampled frequency spectrum
using the discrete Fourier transform (DFT). Given a sampled signal x (n) with a duration
of N samples, the DFT of x (n) is:

N −1
∑
k
X (k) = x (n)e−i2π N n k = 0, 1, 2, . . . , N − 1 (2.7)
n =0
N −1
k k
= ∑ x (n) cos 2π n − i sin 2π n
N N
(2.8)
n =0
Note that frequency is defined in terms of k, which takes on values from 0 to N − 1. We
can think of k as measuring the number of sinusoidal cycles per N samples. As a concrete
example, consider a signal that is sampled 44100 times per second. Let N (the length of
our signal) be 1024. If k = 1, then we have 1 cycle per 1024 samples. Converting to cycles
per second we have:
1 cycle 44100 cycles

f1 = × ≈ 43.066 Hertz (2.9)
1024 samples 1 second
Successive values of k correspond to 0 Hz, 43.07 Hz, 86.13 Hz, 129.20 Hz, etc. Thus, the
frequency spectrum is sampled every 43.07 Hz. Each value of k is said to refer to a specific
frequency bin. Note that the DFT gives us signal X (k ) that, like the time domain signal
x (n), is N samples long.
One very useful property of the Fourier transform and the DFT is that the transform
is invertible. If we have the DFT X (k ), we may recover the time domain signal x (n) as
follows:
N −1
1
∑
k
x (n) = X (k )ei2π N n n = 0, 1, 2, . . . , N − 1 (2.10)
N k =0
Note that the inverse DFT is very similar to the DFT, differing only in a change of sign
1
and a scaling factor of .
N
As noted above, X (k ) is complex valued. In general, the time domain and frequency
domain signals may be either real or complex valued. In audio applications the time
domain signal is typically real valued only, which yields complex values in the frequency
domain. Each complex amplitude X (k ) = Ak encodes both the amplitude and initial
phase shift φk . From our definition of complex amplitude in equation 2.5 and from Euler’s
identity (equation 2.3), we have the following:
X (k ) = Ak eiφk
= Ak cos φk + iAk sin φk
Since Ak and φk are constants, we can represent our complex amplitude as a complex
number of the form a + bi where a = Ak cos φk and b = Ak sin φk . We can visualize this as
a point ( a, b) on the complex plane. The amplitude Ak of the sinusoid is the distance from
the origin to the point ( a, b), and the phase φk is the angle of rotation about the origin. The
amplitude Ak is defined as the magnitude of the complex point.
p
Ak = | X (k )| = a2 + b2 (2.11)
The phase angle φk is given by the inverse tangent function.
b
φk = tan−1 (2.12)
a
Equations 2.11 and 2.12 give an explicit method to convert from a complex frequency bin
value to magnitude (amplitude) and phase. Note that for a real valued signal, phase is
measured relative to a cosine function. Phases may be expressed in sine phase via a phase
shift of − π2 radians.
2.1.4 The Short-Time Fourier Transform
While almost all musical signals have spectra that vary significantly over time, many
sounds exhibit spectra that are locally stable. Such signals may be well-represented by a
series of spectral “snapshots” or “frames” that capture the average spectral content and, in
sequence, represent the time varying spectrum of the sound. Such an analysis is called a
short-time Fourier transform (STFT). Each spectral frame is computed by means of a DFT
of length N and is centered on a particular sample n. Successive frames are computed by
advancing some number of samples H (the hop size) and computing another DFT.
As noted above, the DFT is a discrete sampling of the frequency spectrum. When a
signal consists of a sinusoid with a period that is an integer factor of the DFT length N,
then the sinusoid’s frequency is precisely aligned with one of the DFT frequency bins.
Figure 2.1 shows the magnitude spectrum of a DFT (length N = 16) of a sinusoid with
N
period 4. Note that the magnitude spectrum consists of a single spike at bin 4.
N
Figure 2.1: DFT magnitude spectrum of a sinusoid with period 4, ( N = 32).
Since such precisely constructed signals are unlikely to be found in real-world digital
N
recordings, let us now consider the case of a sinusoid with a period of 4.21 samples, and
its corresponding DFT magnitude spectrum (figure 2.2). In this case, there are significant
components present in bins 0 through 16. Since the signal is N samples long, but has a
sinusoidal period of 4.21 cycles, it cuts off abruptly creating a significant discontinuity.
N
Figure 2.2: DFT magnitude spectrum of a sinusoid with period 4.21 , ( N = 32).
N N
Figure 2.3 compares two sinusoids, one with period 4 and the other with period 4.21 . Note
N
that the sinusoid with period of 4 arrives smoothly at 0 at sample index 32.
Figure 2.3: Sinusoids of 4.21 cycles and 4.0 cycles per N samples, ( N = 32).
Intuitively, one can interpret such discontinuities as additional high frequency com-
ponents. These additional components are often described as spectral distortion or spectral
leakage. These distortions are a significant issue when one is attempting to detect multiple
sinusoids at different frequencies: it may be impossible to differentiate spectral leakage
from sinusoids. By applying a smoothing window function, h(m), in the time domain, these
edge discontinuities can be reduced and the spectral leakage attenuated. The STFT of
length N beginning at sample n and with smoothing window h(m) is defined as follows:
N −1
∑
k
Xn ( k ) = x (m + n)h(m)e−i2π N m (2.13)
m =0
A number of different window functions are commonly used in audio applications.
The most effective are generally bell-shaped curves. Note that the apparent absence of a
smooth window function in fact implies a rectangular window function. The rectangular
window function is defined to be 1 on the interval 0 . . . N − 1 and 0 otherwise. See Harris
(1978) and Nuttall (1981) for a complete discussion of window functions, their definitions,
and properties.
Before continuing further with our discussion of windowing, we must introduce one
other important signal processing concept. The convolution operation ∗ for two signals
x (n) and y(n) is defined as follows:

N −1
f (n) = x (n) ∗ y(n) = ∑ x (m)y(n − m) (2.14)
m =0
The convolution of x (n) and y(n) can be thought of as a new signal, f (n), where each
sample of x is multiplied by a time shifted copy of all the samples in y and the results
summed. Thus, f (n) consists of the scaled shifted sums of copies of y(n).
The convolution theorem (which we will not prove here) states that the spectrum of the
convolution of two signals is equivalent to the multiplication of the signals each expressed
in the frequency domain:
DFT [ x (n) ∗ y(n)] = DFT [ x (n)] DFT [y(n)]
= X ( n )Y ( n )
The converse is also true: the spectrum of the multiplication of two signals in the time
domain is equivalent to the convolution of the signals each expressed in the frequency
domain:
DFT [ x (n)y(n)] = DFT [ x (n)] ∗ DFT [y(n)]
= X (n) ∗ Y (n)
By the convolution theorem, the multiplication of a signal x (n) and a window h(n) is
equivalent to their convolution in the frequency domain: DFT [ x (n)h(n)] = X (n) ∗ H (n).
Since the spectrum of a sinusoid is a single impulse, it follows that the spectrum of a
windowed sinusoid will be a scaled copy of the spectrum of the window function, centered
on the frequency of the sinusoid. Returning to the example of a sinusoid with a period of
4.21 cycles per window, we can get a sense of the effects of windowing by inspecting the
underlying continuous spectrum.2
Figure 2.4: DFT magnitude spectrum and the interpolated underlying magnitude spectrum
N
of a sinusoid with period 4.21 , ( N = 32).
The dotted curve in figure 2.4 shows the continuous underlying spectrum which
consists of a main lobe centered at the sinusoid frequency and many side lobes of lower
amplitude arranged around the main lobe. Different window functions exhibit unique
main lobe shapes and side lobe levels. The case of a sinusoid with an integer period
and a rectangular window is quite special. Figure 2.5 shows that the although there are
significant side lobes in the underlying spectrum, the DFT samples align precisely with
nulls in the continuous spectrum. The result is a single impulse in frequecy bin 4.
Examining the spectra of various window functions can provide a good sense of their
practical capabilities. Figures 2.6 through 2.9 show various window functions in both
the time domain and frequency domains. Note that as the center area of the window
2 The continuous spectrum can be approximated by zero-padding the signal and then performing the DFT.
The result is a spectrum that is oversampled by the zero-padding factor.

Figure 2.5: DFT magnitude spectrum and the interpolated underlying magnitude spectrum
of a sinusoid with period N4 , ( N = 32).
function becomes narrower in the time domain, the main lobe of the window spectrum
becomes wider. This demonstrates an important tradeoff in the design and use of window
functions.
By using the STFT and appropriate windowing, we can successfully model many types
of musical signals and more significantly, we can use these models to effect a number of
different sonic transformations.
2.2 STFT Analysis and the Phase Vocoder
The phase vocoder is one of the most widely utilized STFT techniques for time and
frequency scale modification. The technique makes use of phase information in successive
analysis frames to make a precise estimate of the sinusoidal frequency present at a
particular analysis bin. By default, the frequency resolution of the STFT is limited by
the bin spacing, which in turn is controlled by the DFT length N. For example, as was
observed in equation 2.9, a DFT of length 1024 at a sampling rate of 44.1 kHz results in
a bin resolution of 43.07 Hz. Practical DFTs are computed by means of the fast Fourier
transform which limits N to powers of 2. Therefore, there are some significant restrictions
on the frequency resolution. Moreover, the center frequencies of each STFT bin will be
limited to harmonics of the center frequency of bin 1.

Figure 2.6: Time domain plot and magnitude spectrum of the rectangular window.
Figure 2.7: Time domain plot and magnitude spectrum of the triangular window (N = 31).
Figure 2.8: Time domain plot and magnitude spectrum of the Hann window.
Figure 2.9: Time domain plot and magnitude spectrum of the Blackman window.
Purely harmonic sounds may be analyzed by resampling the signal in the time domain
so that the length of the fundamental period is an integer factor of the DFT length N
(Beauchamp 1993). This assures that each harmonic in the input will be aligned with the
center of an analysis bin. However, this does not accommodate inharmonic sounds, nor
does it allow the measurement of frequency deviations (including vibrato) that are present
in almost all real-world acoustic signals.
2.2.1 Putting Phase to Good Use
Thus far we have been looking at the magnitude spectrum, but each analysis frame also
has a phase spectrum and this information is useful. Since frequency is defined as the
amount of phase change per time unit, the change in phase between successive analysis
frames can be used directly to estimate frequency. We define the phase deviation for
frequency bin k and analysis frame n as follows:

2πk∆ f
∆φk,n = φk,n − φk,n−1 + H (2.15)
fs
where φk,n is the phase for bin k, frame n, H is the analysis hop size in samples, ∆ f is the
frequency bin spacing in hertz, and f s is the sampling rate. Equation 2.15 expresses the
2πk∆ f
idea of phase unwrapping, where the term φk,n−1 + H fs represents the unwrapped
phase of a sinusoid with frequency centered precisely on analysis bin k. Positive phase
deviation ∆φk,n indicates a sinusoid with a frequency higher than the bin center frequency,
negative phase deviation indicates lower frequency.
In practice, the phase deviations must be re-wrapped to the half-open interval [−π, π )
which constrains the allowable frequency deviation and locates the estimated frequency
around analysis bin center frequency. If Θφk,n represents the re-wrapped phase deviation,
then the estimated frequency (in hertz) for bin k and analysis frame n is
Θφk,n

fs
f k,n = + k∆ f (2.16)
2π H
Note that decreasing the hop size increases the range of possible frequency deviation. Small
hop sizes will allow the phase vocoder to more closely track the rapid phase deviations of
non-stationary (noisy) spectral components.
For quasi-stationary (sinusoidal) components, we note that there is an important
relationship between the choice of window function and hop size. As mentioned earlier, a
single sinusoidal component will result in a spectrum that is a scaled translation of the
spectrum of the analysis window. As such, we desire that all bins in the main lobe of
the window should contribute to the energy of the sinusoid, i.e. their frequencies should
“lock” to the frequency at the center of the main lobe. For a main lobe width of W bins and
M
window size of M, a hop size H such that H ≤ W will be sufficient (Puckette 1995). Thus,
the wider the main lobe, the smaller the allowable hop size. This also makes intuitive sense
from a time domain perspective. Since wider main lobes result from narrower bell-like
window shapes in the time domain, the decreased width of the window shape must be
compensated for by more dense sampling in the time domain.
2.2.2 Resynthesis and Transformation
Following analysis, the phase vocoder data can be used to perform a resynthesis of the
original sound. Depending on the desired result, the analysis data may be modified to
effect different sonic transformations. Two different resynthesis methods are in common
use.
The first is the STFT overlap-add technique. With this method, the inverse DFT is
applied to the data in each analysis frame. Since the analysis hop size Ha is typically less
than the DFT length N, a time domain overlapping of each DFT window results. Provided
that the analysis window function sums to unity upon overlap (which can be assured
by choosing the appropriate window function and normalization factor), the original
signal can be recovered exactly. Setting the hop size so that Hs 6= Ha gives a time warped
resynthesis.
The second resynthesis method is the summing oscillator bank. Each bin of the
phase vocoder data is associated with a sinusoidal oscillator and each phase vocoder
frame provides frequency, amplitude, and phase information that can be used to drive
the oscillators. Since each analysis frame is separated by a hop of Ha samples, linear
interpolation of frequency and amplitude is often used to fill in intermediate samples. With
linearly interpolated frequency, the phase function becomes the integral of instantaneous
interpolated frequency. As such, the resynthesized phases cannot be guaranteed to match
the original phases, and thus resynthesis with the oscillator bank will not recover the
original signal. In practice, different frequency interpolation strategies can be used to
achieve fairly high fidelity results. For example, the cubic phase interpolation method
commonly used in sinusoidal modeling can be used to match phase at each frame (see
section 3.3.4 for details).
Typically one is interested in transforming the analysis data in order to achieve one or
more of the following results:
1. frequency filtering (boosting or cutting certain frequency bands)
2. time expansion or compression
3. transposition (frequency scaling)
Of these three, frequency filtering is the most straightforward, and can easily be imple-
mented by selectively scaling the amplitudes in the desired analysis bins. Time varying
filtering may be used in applications such as broadband noise reduction.
2.2.3 Time Expansion/Compression
Time expansion/compression may be achieved in the overlap add case by adjusting Hs

Hs
and/or Ha . The expansion ratio is given by α = . When performing time expansion or
Ha
compression, care must be taken to properly update the phases in each synthesis frame.
Let φk0 (n) denote the resynthesis phase for bin k in frame n. Phase is propagated forward
from each previous frame according to the analysis frequencies as follows:
φk0 (0) = αφk (0) (2.17)

f k (n)2π
φk0 (n) = φk0 (n − 1) + Hs (2.18)
fs
Note that the initial phase is multiplied by the expansion ratio α. In their 1997 study,
Laroche and Dolson show that scaling the initial phases will preserve the original phase
relationships when expanding by an integer factor.
Because time expansion/compression no longer guarantees the same phase relation-
ships as the original sound, an additional smoothing synthesis window function is gen-
erally applied after each inverse DFT. This serves to reduce the effect of any phase or
amplitude discontinuities that result from frequency domain transformations. Since the
application of a synthesis window amplitude modulates the signal, proper care must be
taken to assure enough overlap to reduce amplitude modulation artifacts.
To explore this issue in more detail, it will prove useful to examine the amplitude
behavior of several analysis and synthesis windows under different overlap-add conditions.
N
First let us define the overlap factor β = . When β = 1 there is no overlap, and the
Hs
output consists of a signal that is amplitude modulated by the product of the synthesis
window with the analysis window. (The original analysis window was recovered from
the inverse DFT although the shape may be distorted due to operations performed in
the frequency domain.) Figure 2.10 shows the overlap-adding of Blackman analysis and
synthesis windows with β values of 1, 3, 4, and 6. Note that un-modulated gain is achieved
when β = 6. In general, windows with a narrower time domain peak will require more
overlap.
Additional steps must be taken to assure proper amplitude normalization. For a
synthesis window of w0 of length N, the normalization factor γ is

N −1
γ = Hs ∑ w0 (i ) (2.19)
i =0
Figure 2.10: Overlap-adding the product of Blackman analysis and synthesis windows
with varying overlap factors (β).
Time varying expansion and contraction factors imply a variable synthesis hop size
with a corresponding variable normalization factor. Due to the possibility of amplitude
modulation artifacts, care must be taken not to increase Hs beyond the point of constant
gain. In practice it is best to fix Hs and vary Ha (or linearly interpolate frequency and
amplitude from the analysis spectrum) to achieve the desired expansion ratio.
2.2.4 Transposition
Transposition (frequency scaling) with the phase vocoder may be handled in a number of
different ways. A simple, though computationally expensive, approach synthesizes each
phase vocoder band with a sinusoidal oscillator. Frequency and amplitude may be linearly
interpolated from frame to frame. If F̂k (n) and Âk (n) are the piecewise interpolated
frequency and amplitude functions for bin k, then the resynthesis is

N/2−1
2π
x (n) = ∑ Âk (n) cos nP F̂k (n)
fs
+ φk (2.20)
k =0
where P is the transposition ratio, and φk is the initial phase of bin k.

A less computationally expensive approach is to resynthesize with the overlap-add
method with a time expansion/compression ratio that is the inverse of the transposition
ratio. The transposition is then achieved via resampling in the time domain. Another
approach is to scale the spectrum in the frequency domain as described in Laroche and
Dolson (1999). The spectral magnitude peaks are scaled according to the transposition
ratio, and the phases are adjusted to reflect the new transposed frequencies.
2.2.5 Refinements
A number of additional refinements can be added to the basic phase vocoder technique.
One drawback that has been observed with time expansion/compression is the problem of
“phasiness.” This can best be described as a loss of presence in the synthesized signal; the
result is less crisp and there is sense of greater reverberation. Much of this can be attributed
to a lack of phase coherence between adjacent frequency bins. Although the standard
phase vocoder technique interpolates coherent phase between adjacent analysis frames
(“horizontal” phase), it does nothing to maintain “vertical” phase alignment. Puckette
(1995) proposed a method of phase locking that proceeds by first finding amplitude peaks
in each frame. Under the assumption that frequency bins adjacent to the peak contribute
to the same sinusoid, the phases of adjacent bins are set equal to the phase of the peak bin.
Laroche and Dolson (1997) describe a number of different phase maintenance strategies (all
based on this peak detection method) that can provide significant improvements in sound
quality. Phase relationships are particularly important for preserving transient and micro-
transient events. Robel (2003) describes a technique for reinitializing phases at transients
that significantly reduces phasiness and transient smearing under time scale modifications.
Although the phase vocoder has its limitations, when carefully implemented it can offer
excellent fidelity.
2.3 Spectral Envelopes
Up to this point we have been discussing the precise frequency domain content of audio
signals. It has been shown that the ear functions similarly to Fourier analysis, with specific
locations along the basilar membrane of the inner-ear corresponding to specific frequencies.
However, in addition to frequency content, we are also simultaneously sensitive to the
overall general shape of the spectrum. We call the curve that defines the shape of the
magnitude spectrum the spectral envelope. The dotted curve in figure 2.11 shows one
possible spectral envelope for the given spectrum.
Figure 2.11: Spectral envelope of a soprano voice.
Many vocal and instrumental sounds can be understood in terms of a source-filter model,
where the resonant body of the instrument or of the vocal tract shape (filter) the input (the
source). In voiced human speech, the source is provided by the periodic vibration of the
vocal folds. The spectral envelope can be taken as a model of the resonances of the vocal
tract. For vowels, the peaks in the spectral envelope trace the formants. For any particular
speaker and vowel, the spectral envelope stays fixed regardless of pitch.
If a sound is transposed using the phase vocoder, for example, the spectral envelope
shape will change along with the transposed spectrum. For an instrumental or vocal
sound, the perceived result is a change in size of the resonating body. Thus there is
a change in timbre as well as in pitch. The timbre can be preserved by dividing out
the spectral envelope prior to transposition (this process of flattening the spectrum is
sometimes referred to as whitening), and then filtering the transposed magnitude spectrum
by the original spectral envelope.
Common methods for determining the spectral envelope include linear predictive
coding (LPC), cepstrum, discrete cepstrum, and true envelope methods (Robel and Rodet
2005). For a thorough discussion of spectral envelope estimation methods and applications
see Schwarz (1998).
2.4 Sinusoidal Modeling
Like the phase vocoder, sinusoidal modeling is an STFT based analysis/resynthesis
technique. It attempts to represent an audio signal as a generalized collection of sinusoids
with time varying frequency, amplitude, and phase. Unlike the phase vocoder, the number
of sinusoids present at a given time is not fixed by the DFT size. As a consequence, the
sinusoids need not necessarily be centered around a fixed frequency band, and sinusoids
may vary arbitrarily in frequency. The following plots of a frequency sweep (chirp) first
analyzed with the phase vocoder (figure 2.12) and then with sinusoidal modeling (figure
2.13), clearly illustrate the difference between the two approaches.
The phase vocoder represents the chirp with many frequency bands that each lock
to the chirp frequency as it sweeps upward. The result is highly overspecified, with
many bands per sinusoidal feature. Manipulation of the chirp (transposition for example)
requires the adjustment of many frequency bands. As a corollary, a local adjustment to
any one frequency band will degrade the integrity of the frequency sweep.
In contrast, a sinusoidal model represents the chirp as a single breakpoint function
with time varying frequency. The chirp can be freely transposed and reshaped by the
manipulation of a single frequency/amplitude trajectory. Although in most cases perfect

Figure 2.12: Phase vocoder analysis of a frequency sweep from 100 Hz to 500 Hz.
Figure 2.13: Sinusoidal modeling analysis of a frequency sweep from 100 Hz to 500 Hz.
reconstruction is not possible with a purely sinusoidal model, the benefits of the technique
may outweigh this drawback.
2.4.1 STFT Frequency Interpolation
The basic method of sinusoidal modeling begins with a STFT. As observed in section
2.1.4, the short-time spectrum of a signal is the convolution of the spectrum of the window
function with the spectrum of signal x (n). If the signal in question is in fact a single
sinusoid at frequency f k , then its magnitude spectrum is a single impulse peak at frequency
f k , and its windowed magnitude spectrum is simply the spectrum of the window function
with its main-lobe centered at frequency f k . In the DFT magnitude spectrum, a sinusoid
appears as a local maximum near the true maximum of the underlying continuous
spectrum. Parabolic interpolation may then be used to approximate the position of the
peak of the main lobe.
estimated peak
β γ
α
Figure 2.14: Estimating a peak in the spectrum based on a parabolic curve fitting.
Given the DFT X (k ), the magnitude spectrum can be searched for bins that are local
maxima. Bin n is considered a local maximum if its magnitude is greater than its two
neighboring bins, or more precisely | X (k n−1 )| < | X (k n )| > | X (k n+1 )|. We can then fit
a parabola that passes through the magnitude values of bins n − 1, n, and n + 1. The
parabola takes the following form:
y ( x ) = a ( x − p )2 + b
The known points on the curve are y(−1), y(0), and y(1) and are given by the the bin
magnitudes (measured in dB). According to Smith and Serra (1987), frequency estimates
are found to be more accurate when interpolating using dB rather than linear magnitude.
y(−1) = α = 20 log10 | X (k n−1 )|
y(0) = β = 20 log10 | X (k n )|
y(1) = γ = 20 log10 | X (k n+1 )|

With three equations and three unknowns (a, p, and b) we can solve for the parabola peak
location p.
α−γ

1
p= (2.21)
2 α − 2β + γ
The center frequency (measured in bins) will be n + p. The parabola peak location p can
then be used to solve y( p) for both the real and imaginary parts of the complex spectrum.
For the real part, we set α< , β < , and γ< to the real values at bins n − 1, n, and n + 1,
α< = <e[ X (k n−1 )]
β < = <e[ X (k n )]
γ< = <e[ X (k n+1 )]
and for the imaginary part
α= = =m[ X (k n−1 )]
β = = =m[ X (k n )]
γ= = =m[ X (k n+1 )]
Solving y( p) for the real part a and imaginary part b gives
1
a = y ( p ) < = β < − p ( α < − γ< )
4
1
b = y ( p ) = = β = − p ( α = − γ= )
4
√
According to equations 2.11 and 2.12 the magnitude wil be a2 + b2 and the phase will be
b
tan−1 .
a
This frequency, phase, and magnitude interpolation method means that the analysis
procedure can accurately detect sinusoidal frequencies that are non-integer multiples
of the analysis bin spacing frequency. This makes it ideal for analyzing sounds that
have inharmonic components, have variable frequency content, or exhibit limited non-
stationary behavior (a sum of different harmonic complex tones, sounds with wide vibrato
or glissandi, etc.)
2.4.2 Partial Tracking
Once all of the sinusoidal peaks in a particular STFT frame have been detected, the
peaks must be organized such that they form continuous, time-varying sinusoidal tracks
(partials). This is accomplished by matching the peaks in the current frame with peaks
present in previous STFT frames. During resynthesis the frequency, amplitude, and phase
can be smoothly interpolated from a peak to its matched successor in the next frame.
k0 k1 k2 k3 k4 k5
unmatched peak
frequency
track end (death)
track start (birth) frequency threshold
time
Figure 2.15: Connecting STFT peaks into sinusoidal tracks.
Typically, some peaks that fall below a certain amplitude threshold will be discarded
prior to matching. This means that each frame may contain different numbers of peaks and
that, over time, individual partials may vanish or appear depending on their amplitude.
Amplitude thresholding techniques and tradeoffs will be discussed in more detail in the
next chapter. The major implication for this type of sinusoidal model is that the number
of sinusoids present at any one instant is variable. By eliminating redundancies and low
amplitude components, the result is a model that is efficient to store and manipulate.
Several different techniques have been proposed for partial tracking. The simplest
method is a locally optimal greedy algorithm that matches each peak to the peak in
the next frame that is closest in frequency. Typically there is a distance threshold that
constrains the maximum allowable frequency jump. Each peak is defined in terms of the
following data structure:

structure Peak
Freq
Amp
Phase
Forwardmatch
Backmatch
end structure
Note that peaks keep track of both their successor (Forwardmatch) and predecessor
(Backmatch). Forwardmatch and Backmatch are initialized to Nil, indicating the absence
of a matching peak. Listing 2.1 illustrates the greedy peak matching algorithm.
1: for prevpeak ∈ Pk−1 do

2: for curpeak ∈ Pk do
3: distance ← | prevpeak.Freq − curpeak.Freq|
4: if distance < f req_threshold then
5: if curpeak.Backmatch 6= Nil then
6: existing_distance ← |curpeak.Backmatch.Freq − curpeak.Freq|
7: else
8: existing_distance ← f req_threshold
9: end if
10: if distance < existing_distance then
11: curpeak.Backmatch.Forwardmatch ← Nil
12: curpeak.Backmatch ← prevpeak
13: prevpeak.Forwardmatch ← curpeak
14: end if
15: end if
16: end for
17: end for
Algorithm 2.1: Greedy peak matching from frame k − 1 to frame k.
Pk is the collection of peaks in frame k (the current frame) and Pk−1 is the collection of
peaks in the previous frame. The outer loop iterates through each of the peaks in frame
k − 1. The inner loop looks for the best match in the current frame by iterating through
each of frame k’s peaks. If the frequency distance between any two peaks is less than
distance_threshold (line 4), then it is a candidate for matching. Before connecting the peaks,
the algorithm checks to see if the candidate peak in frame k + 1 has already been claimed
(line 5). If so, the new distance must be less than distance of the current match (line 10).
Following matching, the Forwardmatch and Backmatch fields indicate the complete
sinusoidal tracks. If Backmatch = Nil, a new partial is starting up. If Forwardmatch = Nil,
a partial is ending. When a partial begins or ends, a short amplitude fade is usually added
in order to avoid a discontinuity.
Given that the greedy algorithm has no explicit model of the underlying sound
structure, nor any context beyond the peaks of the current and previous frames, it actually
performs reasonably well in practice. However, the greedy algorithm can be easily confused
by spurious peaks or significant frequency deviations such as vibrato or glissandi (see
figure 2.16). A number of minor refinements can be made, such as incorporating amplitude
into the measure of peak proximity or measuring frequency distance on a logarithmic
scale.
k0 k1 k2 k3 k4 k5
partial tracking conflicts

frequency
time
Figure 2.16: Partial tracking conflicts due to glissandi.
2.4.3 Improved Partial Tracking
A number of other improved partial tracking strategies have been proposed. Under the
assumption that the sound being analyzed contains relatively stable frequency trajectories
(such as stable harmonic or inharmonic partials of a single fundamental), peaks can be
matched to a set of frequency guides (Serra 1989). Guide based partial tracking proceeds
as follows: at frame k we have a set of N peaks Pk with frequencies f 0 , f 1 , f 2 . . . f N −1 ,
and a set of M frequency guides G with frequencies f 00 , f 10 , f 20 . . . f M

0
−1 . As in the greedy
algorithm, each guide attempts to claim the peak in Pk that is closest in frequency to the
guide frequency (| f m0 − f n | is minimized for all combinations, subject to constraints on
the maximum allowable frequency jump). When a match is found between partial n and
guide m, the partial trajectory is continued and a new guide frequency fˆm is computed
according to
fˆm = α( f n − f m0 ) + f m0 for α ∈ [0, 1]
The parameter α is a simple lowpass filter on the guide frequency. When α is zero, the
guide frequencies remain constant. As α approaches 1, the guide frequencies more closely
track the frequency of the evolving partial. Guides that fail to find a match are set to a
dormant state. They can “wake up” in future frames if they find a peak appropriately
proximate in frequency. If a guide remains dormant for too long, it is removed from G.
Peaks that are not matched to a guide may (optionally) spawn new guides. The number of
maximum allowable guides may be limited, if desired.
For harmonic sound, the guide frequencies may be fixed to multiples of the fundamen-
tal frequency f 0 . In this case the number of guides can remain constant throughout the
evolution of the sound. Note that the tracking method works well only if the fundamental
frequency estimation is reliable. Guide based tracking has two distinct advantages: it can
fill temporal gaps in the frequency trajectory, and it can provide a clean representation
for harmonic sounds (with one partial per harmonic). Guide based tracking has been
implemented in a number of sinusoidal modeling packages including SMS, Loris, and
ATS.
For sounds with non-stationary behavior (noise, microtransients, rapid frequency
fluctuations, etc.) guide based tracking offers few advantages in comparison to the basic
greedy algorithm. Depalle, García, and Rodet proposed a tracking method based on
a combinatorial Hidden Markov Model (HMM) that provides more globally optimal
frequency trajectories (Depalle et al. 1993). A Hidden Markov Model consists of a set
of related states, observations, and parameters. In a typical HMM problem, the goal is
to determine a temporal sequence of states based solely on observations that have some
probabilistic connection to the states. The states are the “hidden” part of the model,
and the parameters determine the probability of a state (or state transition) given certain
observations.
In the application of HMMs to partial tracking, a state consists of a set of connections
from frame k − 1 to k. The goal is to find the most likely sequence of states (connections)
given the observations. In this case the observation is simply the number of peaks present
in frame k − 1 and in frame k. If we let Sk model the set of peak connections from frame
k − 1 to k then next state, Sk+1 will specify the connections from frame k to k + 1. Thus,
a state transition involves the peak trajectories of three frames: k − 1, k, and k + 1. The
parameters of the HMM specify the probability of a transition from state to state. The
parameters are based on a cost function that favors peak connections with continuous
frequency slopes. Note that actual frequency distance is not considered. For this reason
the HMM method can be used to track crossing partials. Given the parameters of the
model (the probability of a particular state transition), the Viterbi algorithm is used to find
the most likely sequence of states. The optimal sequence of transitions will be that which
minimizes frequency discontinuities.
The computational cost of the HMM method is high since it must consider all com-
binations of peak connections across all frames. In practice, the number of frames (and
therefore the number of state transitions) is limited to some window of T frames. The
number of possible peak connections under consideration is further constrained by fixing
the number of partials alive during any interval of T. Other options could further constrain
the combinations, such as limiting maximum allowable frequency slopes or limiting the
number of partial crossings.
A particular weakness of the HMM method described by Depalle et al. is that it cannot
fill temporal gaps in the partial trajectories. Gaps may occur (especially in low amplitude
partials) due to bin contamination from noise, or if one is analyzing material that has
undergone lossy compression (a situation that may be common given current audio
distribution methods). Two new promising partial tracking approaches include future
trajectory exploration (Lagrange et al. 2004) and tracking using linear prediction (Lagrange
et al. 2003). SPEAR has implemented the linear prediction method which will be detailed
in section 3.2.
2.4.4 Synthesis
The standard resynthesis method for a sinusoidal model is the summing oscillator bank.
Each frequency trajectory computed via partial tracking is represented as a breakpoint
envelope where where point k of envelope i is represented by the time, frequency, ampli-
tude, and phase values (tik , f ik , aik , φik ). Linear interpolation of amplitude and frequency
can be used to resynthesize each breakpoint segment. Fik (n) and Aik (n) are the linear
interpolation frequency and amplitude functions for the segment from breakpoint i to
breakpoint i + 1.
n k
!
f s − ti

Fik (n) = f ik + f ik+1 − f ik for f s tik ≤ n < f s tik+1 (2.22)
tik+1 − tik
n k
!
f s − ti

Aik (n) = aik + aik+1 − aik for f s tik ≤ n < f s tik+1 (2.23)
tik+1 − tik
where n is the output sample index, f s is the sampling rate, and tik is breakpoint time in
seconds. If we let F̂i (n) and Âi (n) be the piecewise amplitude and frequency functions for
partial i defined for all n, then the final resynthesis is given by

N −1
2π
x (n) = ∑ Âi (n) cos n F̂i (n)
fs
+ φi (2.24)
i =0
where φi is the initial phase of partial i.
Note that the instantaneous phase is the integral of the frequency function. For linear
frequency interpolation this implies a quadratic phase function. The phase values at all
but the first breakpoint are ignored. If we desire a resynthesis that is phase accurate, a
different interpolation strategy must be used. McAulay and Quatieri (1986) developed a
cubic phase interpolation method that matches frequency and phase at each breakpoint.
This will be discussed in detail in section 3.3.4.

2.4.5 Limitations and Extensions
Sinusoidal modeling makes the assumption that the audio signal can be closely approxi-
mated as a sum of sinusoids. Although this may hold true for certain classes of signals
(harmonic sounds with clear tonal components and slowly varying frequency content)
many types of signals are not well represented by sinusoids — broadband noise or transient
events such a snare drum hit, for example. In many cases, sounds of interest will exhibit
both tonal and noisy components. For example, a flute tone will contain quite a bit of
broadband breath noise in addition to a clear fundamental.
Partial tracking becomes particularly difficult in the presence of noise, reverberation,
or dense polyphony. Recent research continues to suggest new approaches for partial
tracking, noise modeling, and transient preservation. A few of these methods have been
implemented in SPEAR and will be discussed in the next chapter.
Even with extensions to the basic sinusoidal modeling method, perfect signal recon-
struction is not guaranteed. However, when considering compositional goals, perfect
reconstruction is rarely the intent. If lossless reproduction is required, sampling synthesis
accomplishes this with far greater ease and efficiency. The power of a sinusoidal model
lies in its potential to allow independent transformations in both time and frequency. One
can begin with models that closely mimic real-world acoustic events and then, through
a series of transformations or interpolations, explore a vast space of synthetic variations.
Clearly this is fertile territory for creative musical exploration.

42
3. Sinusoidal Analysis and Resynthesis
SPEAR provides two different spectral analysis modes — the phase vocoder and sinusoidal
modeling. The phase vocoder follows the standard method described in the preceding chap-
ter. This chapter will focus on specific implementation of sinusoidal analysis/resynthesis.
Several extensions to the basic techniques, as described in the preceding chapter, will be
discussed.
3.1 Analysis Parameters
Sinusoidal analysis requires a number of input parameters including window type, window
size, DFT size, analysis hop size, amplitude thresholds, and partial tracking constraints.
All of these parameters can have a significant effect on the quality of the analysis. The
interaction of amplitude thresholds and partial tracking can be particularly complex. The
goal is to provide defaults for all of these parameters that work reasonably well for a
wide variety of musical signals. However, even with careful choices, some user-selected
parameter inputs are almost always required. Typically, the user might begin analysis
using the default settings and then adjust accordingly after viewing the analysis and
auditioning the resynthesis.
3.1.1 Analysis Window
One of the most crucial parameter choices is the length of the analysis window. This will
control width of the main lobe in the frequency domain and, as a result, will determine the
frequency resolution of the analysis. Rather than specifying the window length in samples,
the user inputs the frequency resolution f a in hertz. Adjacent sinusoids that are at least f a
3. Sinusoidal Analysis and Resynthesis 43
hertz apart can be properly resolved in the analysis. Additionally, no significant sinusoid
with a frequency less than f a hertz should be present in the input. For a harmonic sound,
f a should correspond to the frequency of the fundamental.
Given f a , we must determine the appropriate window size and DFT size. In order to
resolve sinusoids using the peak picking and quadratic interpolation method, the window
size must be large enough to insure separation of the window main lobes. The window
length M is given by

fs
M = ∆s (3.1)
fa
where f s is the sampling rate and ∆s is the desired window separation in bins.
SPEAR uses a Blackman window which has a main lobe width of 6 bins. If we require
complete separation of the main lobes as shown in figure 3.1, then ∆s = 6. Note that as
∆s increases, the window length grows. Since large windows introduce an undesirable
loss of temporal resolution, we would like to use a smaller value of ∆s that still allows
for the resolution of closely spaced sinusoids. Abe and Smith (2004) made an extensive
study of the minimum allowable frequency separation (MAFS) for various window types.
The MAFS is the smallest separation that allows for peak detection and does not introduce
significant bias in the peak interpolation due to interference from a neighboring peak. With
a spectral oversampling (zero padding) factor of 2, the MAFS for the Blackman window is
3.53 bins (Abe and Smith 2004, 8, table 9). This is shown in figure 3.2. SPEAR uses the
slightly more conservative separation of 4 bins. Given a window length M, the FFT size is
N = 2dlg Me+1 , which results in spectrum oversampling by a minimum factor of 2.
With a peak separation of 4 f a and spectrum oversampling by a minimum factor of
2, each main lobe peak will be sampled by at least 7 DFT bins. To aid the rejection of
spurious peaks we may conservatively impose the restriction that the three magnitudes
| X (k n−1 )|, | X (k n )|, | X (k n+1 )| of a parabolic peak must also all exceed the magnitude of
either neighboring bin: | X (k n−1 )| > | X (k n−2 )| or | X (k n+1 )| > | X (k n+2 )|.
Figure 3.1: Magnitude spectrum of the main lobes of two Blackman windows with a
frequency separation ∆s = 6.
Figure 3.2: Magnitude spectrum of the main lobes of two Blackman windows with the
minimum allowable frequency separation ∆s = 3.53.
3.1.2 Amplitude Thresholds
Amplitude thresholds are used to limit the number of peaks in each analysis frame. The
desire is to detect and track only the most perceptually significant partials. Because
many sounds of interest exhibit a high frequency roll-off, it is helpful to equalize the
spectrum. This results in an analysis that gives equal weight to both high and low
frequency components and avoids resynthesis that sounds dull and heavily lowpass
filtered.
Equalization can be achieved by applying a pre-emphasis high-pass filter in the time
domain (Maher 1989), or by applying an emphasis curve to the STFT data (Serra 1989).
One particular disadvantage of the applying a filter in the time domain is an alteration
of the magnitude and phase spectra, the effects of which can only be easily undone by
applying a complementary de-emphasis filter following resynthesis.
SPEAR achieves the effect of pre-emphasis by using a frequency dependent amplitude
threshold curve. SPEAR establishes two amplitude thresholds: Tb , which is a variable
threshold that must be exceeded to begin a new partial track, and Td , which is fixed lowest
threshold (default value is −90 dB). If peaks are found that exceed Td , then they may be
used to continue an existing partial track. If no appropriate track continuation is found,
the partial ends — so Td may be considered a “death” threshold for possible continuation
of a partial.
Tb is computed using a frequency dependent threshold curve, At ( f p ), in combination
with the maximum bin amplitude of the analysis frame, amax

n . Incorporating the maximum
bin amplitude allows Tb to track the overall signal level. The threshold curve is given by

aR aR
At ( f p ) = a T + a L + − b f p /20000 (3.2)
b−1 b−1
where f p is the frequency in hertz, b is a parameter controlling the shape of the curve, a T
is user controllable threshold in dB, a L is the amplitude offset at 0 kHz in dB, and a R is
the range of the curve in positive dB from 0 to 20 kHz. Figure 3.3 shows At ( f k ) with the
default values b = 0.0075, a T = −60, a L = 26, and a R = 32.
Given the maximum bin amplitude in dB for a frame, amax

n and the frequency f p in Hz
of the peak under consideration, the threshold Tb is given by
Tb = amax
n + At ( f p ) (3.3)
Tb constitutes a “birth” threshold for potential new partials. The birth threshold requires
low frequency peaks to exceed a greater threshold than high frequency ones. In addition
to helping find more high frequency partial tracks, this aids in the rejection of spurious low
frequency, low amplitude partial tracks. Although the high frequency emphasis afforded
Figure 3.3: Frequency dependent threshold curve with the default values b = 0.0075,
a T = −60, a L = 26, and a R = 32.
by the variable threshold provides increased resynthesis fidelity for many real world
sounds, it can result in many low amplitude, high frequency partials. In many cases, the
signal energy above 5 kHz could be more efficiently and robustly modeled as noise bands,
rather than clusters of low amplitude sinusoids. Possible approaches to modeling noise
will be discussed in section 3.6.
3.2 Partial Tracking Using Linear Prediction
Peaks that are retained following thresholding are joined into breakpoint functions rep-
resenting individual sinusoidal partials. The linear prediction (LP) method of Lagrange
et al. (2003) is used to connect the peaks into tracks. Linear prediction treats each evolving
sinusoidal track as a signal. The current signal sample, x (n), is approximated as a linear
combination of past samples:

K
x̂ (n) = ∑ a(k) x (n − k) (3.4)
k =1
There are K linear prediction coefficients, a(k ), that are calculated by minimizing the error
between the predicted value, x̂ (n), and the actual value x (n). The K coefficients can then
be used to predict successive values x̂ (n + 1), x̂ (n + 2), etc. Several different algorithms,
including the Burg method, autocorrelation, and covariance can be used to compute a(k ).
Lagrange et al. show that the Burg method is most suitable in this application.
LP can be used to compute possible future values for both frequency and amplitude.
The frequencies and amplitudes of newly detected peaks are compared to the predicted
values, and the closest matches are used to extend the sinusoidal tracks. The method is
implemented as follows: at frame k we have detected N candidate peaks and we have
M active sinusoidal tracks which extend to at most frame k − 1. For each of the M
active sinusoidal tracks, the Burg method is used to compute two sets of linear prediction
coefficients, one on frequency and one on amplitude. An LP model of order 6 computed
from a maximum of 64 previous values has worked well in practice. It is critical that
enough points are used to capture periodic features such as amplitude modulation or
vibrato. For a more thorough discussion of the choice of LP parameters, see Lagrange et al.
(2007).
pr pr
The LP coefficients for track m are used to predict frequency f m and amplitude am
values for frame k. When a new sinusoidal track starts and there are not enough values to
compute the LP coefficients, the mean frequency and amplitude are used as the predicted
values. The error between predicted values for track m and peak n are given by a measure
of Euclidean distance
s 2 2
f nobs aobs

n
Em,n = 12 lg pr + α 20 log10 pr (3.5)
fm am
where f nobs and aobs

n are the observed frequency and amplitude values for peak n. The first
term measures distance in semitones and the second in dB with a scale factor α. Informal
1
tests have shown α = 12 to offer a reasonable weighting between prediction errors in
frequency versus those in amplitude.
Each track m selects the best continuation peak n over all N peaks subject to constraints
on the maximum allowable difference in predicted versus actual frequency. More precisely,
pr
| f nobs − f n | < ∆ f max where ∆ f max is proportional to the analysis frequency f a . For harmonic
sounds, ∆ f max = 3
4 fa is a reasonable default value. ∆ f max is a user-controllable parameter
and can be reduced for sounds with very stable partials or increased for sounds with
partials that exhibit wide frequency excursions.
As with guide based tracking methods, linear prediction partial tracking can span
temporal gaps in the analysis data. Tracks that fail to find a match either become inactive
or lie dormant for several successive frames. In the case of a track that has been dormant
for j frames, linear prediction is used to predict the j + 1th values for matching to candidate
peaks.
3.3 Resynthesis
SPEAR offers several different synthesis methods. For real-time playback, the inverse
FFT method (abbreviated as IFFT or FFT−1 ) offers excellent efficiency and very good
quality (Rodet and Depalle 1992). Although most synthesis artifacts of the standard IFFT
method can be minimized with the use of appropriate overlap-add windows, extremely
rapid modulations of frequency or amplitude may not be adequately reproduced. For the
highest quality sound, SPEAR can perform oscillator bank synthesis with optional cubic
phase interpolation as in the classical McAulay-Quatieri method. SPEAR also supports the
resynthesis of Loris RBEP files which include a noise component for each breakpoint (Fitz
1999). These so-called bandwidth enhanced partials can be synthesized with either the
IFFT method or noise modulated oscillators.
3.3.1 Inverse FFT Synthesis
Since the DFT is invertible, it is possible to synthesize any desired time domain waveform
by constructing the complex spectrum of the desired signal and applying the inverse DFT.
In this case we are interested in the synthesis of sinusoids which, as detailed in section
2.1.4, have well understood spectra. A time domain windowed sinusoid can be synthesized
by the following procedure:
1. shift the complex spectrum of the window function so that it is centered on the bin
frequency of the desired sinusoidal frequency

2. scale the spectrum of the window function according to the desired complex ampli-
tude (magnitude and phase)
3. accumulate the scaled shifted window function into the DFT buffer
4. perform the inverse DFT
Steps 1–3 correspond to convolving the window spectrum with the spectrum of a sinusoid
(a line spectrum). Because convolution is a linear operation, steps 1–3 can be performed
for each sinusoid to be synthesized. The inverse DFT can then computed as a final step.
The final DFT only adds a constant amount of work regardless of the number of sinusoids.
Provided that steps 1–3 can be carried out efficiently, we have the basis for an algorithm to
quickly synthesize a sum of many sinusoids.
Step 1 can be performed efficiently by pre-computing an oversampled spectrum of the
window function. For a DFT/IDFT of length N, we first compute a time domain window
function of h(n) of length N. The window function is then zero padded (zeros appended)
so that it has length M = N × R, where R is the oversampling factor (R is typically chosen
to be a power of 2). Call this zero padded window h0 (n).


 h(n) for 0 ≤ n < N

0
h (n) = (3.6)
 0 for n < 0 or n ≥ N

The DFT of the zero padded window gives the oversampled interpolated spectrum of the
window H (k ).
N −1
∑
k
H (k) = h0 (n)e−i2π N n k = 0, 1, 2, . . . , N − 1 (3.7)
n =0
The window spectrum can then be shifted to any fractional bin position that is a multiple
1 1
of . For bin frequencies that are not a multiple of , the window value may be
R R
approximated by index truncation, index rounding, linear interpolation, or some higher
order interpolation scheme.
Steps 2–3 — scaling and accumulating the complex spectrum of the window func-
tion — can be performed with various degrees of efficiency depending on the precision
required. To precisely synthesize the desired sinusoid, all N values of the window spec-
trum must be scaled and accumulated into the DFT buffer. Equation 3.8 summarizes
the shifting, scaling, and accumulation of the window spectrum W (n) for a sinusoid of
amplitude An and bin frequency of bn . Note that to synthesize a real valued-sinusoid,
we must represent half of the amplitude as a complex sinusoid, and half as its complex
conjugate.
An An
X (k) = H (b(k − bn ) Rc)eiφ + H (b( N − k − bn ) Rc)e−iφ (3.8)
2 2
The symmetry of the DFT buffer means that X (k ) can be constructed by first accumulating
N/2 values in the positive frequency half (0 ≤ n ≤ N/2), and then reflecting their complex
conjugate in the negative frequency half (N > n > N/2).
The procedure implies one table lookup (for the window function), and one complex
N
multiply (four scaler multiplies) for each of the 2 + 1 values. Compared to synthesis via a
table lookup oscillator, this method is actually less efficient. The oscillator requires one
table lookup and one scaler multiply for each of the N samples. However, significant
savings can be achieved if we approximate the scaled and shifted window function
spectrum.
Although the complete window function spectrum has length N, it should be noted
that many of the values are close to zero. Therefore, we can approximate steps 2–3 by
scaling and shifting only J window samples immediately surrounding the main lobe
(typically between 7–15 samples). Laroche (2000) calls the restricted set of samples the
“spectral motif.” Accumulating only the spectral motif (rather than the entire window
spectrum) reduces the number of table lookups and complex multiplies to a constant J,
regardless of the DFT buffer size.
Although any window function can be used, for computational efficiency we desire a
window function with a reasonably narrow mainlobe and low sidelobes. When the DFT
buffer is filled, only the values of the main lobe and the first few side lobes need to be
computed.
The result of the inverse DFT is a sum of sinusoids (each with constant amplitude
and frequency) that have been multiplied by the time domain window function h(n).
To synthesize time varying amplitudes and frequencies, an overlap-add process can be

1
used. We begin by multiplying each sample in the DFT buffer by h(n)
. This effectively
removes the time domain effect of the window, leaving a constant amplitude sinusoid. We
then apply a triangular synthesis window s(n) which gives the desired linear amplitude
interpolation when overlap-added together.
Since frequency is constant across each DFT buffer, there will be some modulation
artifacts resulting from overlap-adding. These effects can be reduced by carefully matching
the phase of each sinusoid from one buffer to the next. For an overlap factor of 2, the
phase should be matched at the midpoint of the triangular slope (Rodet and Depalle 1992).
An important implementation detail of the IFFT method concerns the imprecision
introduced by the approximation of the window spectrum. By only using J samples
centered on the main lobe, the window spectrum has been band-limited. Effectively the
window spectrum has been multiplied by a rectangular window. Since multiplication in
the frequency domain results in convolution in the time domain, the time domain window
is no longer limited to N samples. The result is time domain aliasing which introduces
distortion in the resynthesized sinusoid. This distortion is particularly evident at the edges
of the time domain buffer where the window function approaches its minimum. When
1
the window function is divided out (multiplied by h(n)
), these distortions are dramatically
amplified. It is particularly evident for window functions that taper to zero at the edges
(windows in the Blackman-Harris family, for example). Figure 3.4 compares a sinusoid
with IFFT synthesized versions using different values of J. Note that even with large values
of J, significant error is still present. To avoid this problem, some number of samples D at
the beginning and end of the DFT buffer must be discarded. D is typically on the order of
N
, so the reduction in synthesis efficiency is negligible.
8
Inverse FFT synthesis appears to offer some open areas for continued research. More
detailed studies of signal distortion in inverse FFT synthesis would be welcome. It would
be helpful to have a more precise analysis of the effects of different synthesis windows,
window oversampling and interpolation strategies, the choice of J (controlling the size
Figure 3.4: Comparison of sinusoids generated via IFFT synthesis with different values of
J (the spectral motif length) showing the first 16 samples of the time domain signal where
N = 64, frequency = 1.5 cycles per N samples. Window type is Blackman-Harris.
of the spectral motif), and the effect of D on signal to noise ratio. Jean Laroche’s (2000)
investigations provide some important preliminary insights in this area.
3.3.2 IFFT Synthesis Implementation

N − 2D
With an overlap add factor of 2, each inverse DFT computes new output samples.
2
N
In the current implementation, SPEAR performs IFFT synthesis with N = 512, D = .
8
These parameters allowed for real-time resynthesis of hundreds of partials on relatively
modest hardware.1 Future versions of SPEAR may allow the user to fine tune the resyn-
thesis parameters. In particular, it may be desirable to reduce N. Because frequency and
amplitude are constant across each DFT buffer, rapid frequency changes may not be well
represented with larger N.
In the current implementation, the spectral motif length, J, is fixed at 9. Because of
the problem of distortion at edge of the DFT buffer, the Hamming window, as defined
in equation 3.9, was chosen for the synthesis window. Since we divide by the window
function, windows that taper to zero will amplify distortions rather dramatically as the
1 PowerPC G3 class processor at 400 MHz — circa 2003
inverse tends toward infinity. The minimum value of the Hamming window at the buffer
2
edge is .
23
25 21 2πn

 46 − 46 cos N − 1 for 0 ≤ n < N



h(n) = (3.9)


for n < 0 or n ≥ N

 0
Real time resynthesis proceeds according to a time counter t that indicates the current
location in the analysis data. Whenever a new output buffer is required, the synthesis
callback checks to value of t to determine the set of sinusoids that need to be synthesized.
The center of each DFT synthesis window is represented by t. Pn is the set of partials
active at time tn . The amplitude and frequency for each partial in Pn is determined by
linear interpolation. At time tn+1 , we have a new set of partials Pn+1 . All partials in
T
Pn Pn+1 must have phase continuity at the overlap-add midpoint. This is achieved by
maintaining a set of virtual oscillators that store the previous phases. Each oscillator
structure is associated with an active partial by means of a hash table (Cormen et al. 1990).
Thus, it is reasonably efficient to lookup phase as partials turn on and off. Moreover t
can proceed at any rate through the analysis data. This facilitates the ability to manually
“scrub“ forward and backward through the sound. The analysis data can also be freely
transformed concurrent with resynthesis. SPEAR does appropriate data locking to make
sure the analysis data structures remain consistent between the concurrent resynthesis and
user interface threads.
3.3.3 Oscillator Bank Resynthesis
SPEAR can optionally perform a non-realtime oscillator bank resynthesis with linear
interpolation of frequency and amplitude as shown in equation 2.24 (page 40). This
provides a much higher fidelity result than the IFFT method.

3.3.4 Phase Management
As noted in section 2.4.4, linear interpolation of frequency will, in general, only retain
the phase of the first breakpoint of a partial. Although in many situations the phase
information may not be perceptually important, there are cases where temporal phase
coherence between partials is quite important, for example in the reproduction of transients.
We also seem to be particularly attuned to the phase relationships in speech signals.
McAulay and Quatieri (1986) developed a frequency/phase interpolation method that
maintains phase and frequency at each breakpoint. Consider two successive breakpoints i
and i + 1 with phases φi and φi+1 and frequencies ωi and ωi+1 (here we are expressing
frequency in radians per sample rather than cycles per second). Since frequency is the
derivative of phase, we desire a phase interpolation function with endpoints equal to φi
and φi+1 and first derivatives at the endpoints equal to ωi and ωi+1 . Independent control
of the slope at two points requires a cubic function, so we begin by defining a cubic phase
of the form
θ (n) = ζ + γn + αn2 + βn3 for n = 0 . . . S (3.10)
where n is the sample index and S is the number of samples from breakpoint i to breakpoint
i + 1. The first derivative (the frequency function) is
θ 0 (n) = γ + 2αn + 3βn2 for n = 0 . . . S (3.11)
The endpoints at sample 0 and sample S are
θ (0) = ζ = φi (3.12)
θ 0 (0) = γ = ωi (3.13)
θ (S) = ζ + γS + αS2 + βS3 = φi+1 (3.14)
θ 0 (S) = γ + 2αS + 3βS2 = ωi+1 (3.15)
With four equations we can solve for the four unknowns. ζ is equal to the initial phase, and
γ is equal to the initial frequency, which leaves only to solve for α and β. One complication
9π 9π
φi+1 + 8π M=4
7π 7π
θ(n) = φi + ωin + α(M)n2 + β(M)n3 φi+1 + 6π M=3

5π 5π
phase
φi+1 + 4π M=2
3π 3π
φi+1 + 2π M=1
φi π π
0 slope = ωi+1
slope = ωi
M=0
-π φi+1
n=0 n=S
sample index
Figure 3.5: Family of cubic phase interpolation functions with different values of M (after
McAulay and Quatieri (1986)).
is that the final phase is measured modulo 2π, so in fact θ (S) should be expressed as
θ (S) = ζ + γS + αS2 + βS3 = φi+1 + 2πM (3.16)
for some integer M. M defines a family of cubic phase functions with different amounts
of phase unwrapping (see figure 3.5).
The solutions for α and β are parameterized in terms of M as follows:
3 1
α( M) = (φi+1 − φi − ωi S + 2πM) − (ωi+1 − ωi ) (3.17)
S2 S
−2 1
β( M) = 2 (φi+1 − φi − ωi S + 2πM) + 2 (ωi+1 − ωi ) (3.18)
S S
McAulay and Quatieri show that the optional choice of M is that which makes the phase
curve maximally smooth (the maximally smooth choice for figure 3.5 is M = 2). This can
be defined in terms of minimizing the second derivative of θ (n) with respect to sample
index. The result is that the maximally smooth M is the closest integer to x, where

1 S
x= (φi + ωi S − φi+1 ) + (ωi+1 − ωi ) (3.19)
2π 2
The complete phase function is
θ (n) = φi + ωi n + α( M)n2 + β( M)n3 for n = 0 . . . S (3.20)
It is important to note that cubic phase interpolation will generally only work for unmodified
resynthesis. Partials that have undergone time dilation, transposition, or frequency shifting
cannot be synthesized reliably with this method since the resulting phase functions are no
longer “smooth” enough to avoid frequency modulation artifacts. Methods for preserving
phase under modifications remain an interesting open problem. It may prove effective to
resynchronize phase only at certain moments (as suggested by the work in Robel (2003)
and Fitz and Haken (2002)), or slightly modify the breakpoint frequencies to smooth out
the phase function. Phase interpolation in the unmodified case is important for recovering
the residual signal, which is discussed in section 3.6.
3.4 Data Storage
A popular data storage model in analysis-synthesis applications is a sorted list of time
frames. Each frame contains a list of peaks (amplitude, frequency, and phase) that are
connected to successive time frames via either index numbers or pointers.
Rather than represent the analysis data as a sorted list of time frames, SPEAR uses
a list of partials that are represented by breakpoint functions of time versus amplitude,
frequency, and phase. With this storage model the implementation of cut, copy, and paste
operations, of both entire partials or segments, is straightforward. Additionally, partials
can be individually shifted, stretched, or compressed in time without resampling to fixed
frame time points. A further advantage is that the storage model can easily support
multirate analysis data (Levine 1998) or time reassigned breakpoints (Fitz 1999).
3.4.1 Time-span Frames
A common operation, particularly during synthesis, is determining the list of partials (and
breakpoint segments) crossing a particular time point. In the frame based storage model,
a binary search of sorted frames quickly determines the partials that are active at any
particular time.
With a list of breakpoint partials, this query requires an iteration through all the
partials in the data set. For a short sound, this may only require looking at several hundred
elements, but for longer sounds there are likely to be many partials of short duration.
For example, the total number of partials for a one minute sound could be well over
twenty-thousand. During synthesis, continued iteration over thousands of partials is too
inefficient. A solution is to maintain a parallel data structure of time-span frames. For
each frame, a list is kept of all partials that have breakpoints within the frame’s time
span — typically on the order of 0.125 seconds. Figure 3.6 shows an example of ten partials
segmented into five frames where each frame contains a list of its active partials. The
active partial lists maintain a relatively constant size as partials turn on and off. For a
typical sound, this reduces active partial lookups to an iteration over only several hundred
elements.
3
5 7 8
2
9
4
0 6
t0 t1 t2 t3 t4 t5
0, 1, 2, 0, 1, 2, 0, 1, 3, 0, 1, 3, 8 0, 1, 8, 9
3, 4 3, 4, 5 4, 6, 7
Figure 3.6: Data structure for division of partials into time-span frames. Comma delimited
lists for each frame indicate the active partials.
3.5 Transient Sharpening
As observed in section 3.1.1, the choice of window length directly affects both the frequency
and temporal resolution of the analysis. For polyphonic or inharmonic sounds in particular,
we desire a long analysis window so as to be able to resolve closely spaced partials.
However, as the window length increases, the temporal resolution decreases. This is a
particular problem for sounds that contain transients. All signal energy is spread across
the window and assigned to a breakpoint located at the temporal center of the analysis
window rather than at the precise location of the transient. Moreover, since analysis
windows overlap, the energy will be distributed across multiple windows and assigned to
multiple breakpoints. This results in the familiar pre- and post-echo effects of traditional
STFT analysis/synthesis.
Fitz (1999) shows how the method of reassignment can be used to improve the temporal
resolution and reduce pre- and post-echo. The method of reassignment begins by noting
that traditional STFT analysis assigns all energy included in the window to the temporal
center of the window. When there is significant energy in a particular frequency band that
is located near the edge of a window, this energy should instead be assigned to the center
of gravity (time-frequency centroid) of the window. This time-frequency coordinate can be
computed from the derivatives of the phase spectrum. The time-frequency centroid will
be located at the point where phase is changing most slowly (where the partial derivative
of phase with respect to time and phase with respect to frequency is zero).
Auger and Flandrin (1995) and Plante et al. (1998) show how to efficiently compute the
point of reassignment as a ratio of Fourier transforms. The reassigned time for frequency
bin k is given by ( )
Xht (k) Xh (k)
t(k) = t + <e (3.21)
| Xh (k)|2
where t is the original center of the analysis window (in samples). Xh (k ) is the DFT of
the signal using the usual analysis window h(n). Xht (k) is the DFT of the signal using a
special time-weighted analysis window ht(n). ht(n) is defined as the product of h(n) and
a ramp function from − M2−1 to M −1

2 , where M is the window length.
M−1

ht(n) = h(n) n − n = 0, 1, 2, . . . , M − 1 (3.22)
2
The real portion of this ratio represents the time reassignment in samples.
The reassigned frequency for bin k is given by

( )
Xhω (k ) Xh (k )
ω ( k ) = ωk − = m (3.23)
| Xw (k)|2
where ωk is the bin frequency, and Xhω (k ) is the DFT of the signal using a frequency-
weighted analysis window hω (n). hω (n) can be computed by taking the DFT of h(n),
multiplying by a frequency ramp, and then taking the inverse DFT. Note that the time-
frequency center of gravity is computed independently for each frequency bin.2
Pre- and post-echo can be dramatically reduced by removing breakpoints with large
time corrections (Fitz 1999; Fitz and Haken 2002). Specifically, if the time correction is
larger than analysis hop size (and the hop size is less than the window length — which is
always the case in this implementation), a breakpoint can be safely removed since it will
be detected in a neighboring frame. Figures 3.7 and 3.8 show the analysis of a castanet
click that occurs at about 0.868 seconds.
Note that the transient event is well localized, and pre- and post-echo are dramatically
reduced. In order to avoid a complete amplitude discontinuity at the onset of a partial,
a zero amplitude breakpoint is joined to the head of each partial. Informal observations
indicate that although the reassignment method does an excellent job of localizing sharp
transients, there is some loss of energy, likely due to the removal of the breakpoints with
large time corrections. The addition of a zero amplitude breakpoint creates a short fade
into each transient which has the additional benefit of raising the average signal level
around the transient event. Further experimentation should be conducted to determine
whether transient breakpoints would benefit from an additional amplitude boost.

2 Note that equations 3.21 and 3.23 differ from those in Auger and Flandrin (1995), Plante et al. (1998), and
Fitz (1999) in that we add the reassignment term for time reassignment and subtract for frequency reassignment.
The sign depends on whether or not the STFT is defined with a time reversed window. We use the definition in
Figure 3.7: Breakpoints from the analysis of a castanet click using the standard sinusoidal
model.
Figure 3.8: Breakpoints from the analysis of a castanet click using the time reassignment
method. Breakpoints with large time corrections have been removed.
3.6 Noise Modeling
Serra and Smith were the first to develop extensions to sinusoidal modeling for representing
noise. The SMS (Spectral Modeling Synthesis) system offers two different ways to model
non-sinusoidal components. In the residual model, the sound is first analyzed using
sinusoidal modeling. Next, the signal is resynthesized using the oscillator bank method
with cubic phase interpolation as described in section 3.3.4:

N −1
x̂ (n) = ∑ Âi (n)cos(θ̂i (n)) (3.24)
i =0
The residual signal e(n) is then computed by subtracting the resynthesized signal x̂ (n)
from the original x (n).
e(n) = x (n) − x̂ (n) (3.25)
Because phase is preserved in the resynthesis, this subtraction can take place in the time
domain. The residual e(n) will typically consist of noise, attack transients, and pre- and
post-echo. The original signal can of course be perfectly reconstructed by adding the
residual to the resynthesis. However, typically we wish to make some sonic transformations
prior to resynthesis. In this case a more flexible modeling of the residual is required.
The stochastic model represents the residual in the frequency domain. For each frame l
of the analysis, we compute a residual magnitude spectrum from the magnitude spectrum
of the original signal and the magnitude spectrum of the phase accurate resynthesis.
| El (k)| = | Xl (k)| − | X̂l (k)| (3.26)
Xl (k ) is the DFT of the original signal at frame l and X̂l (k ) is the DFT of the sinusoidal
resynthesis. Assuming that e(n) is a stochastic signal, it can be modeled as a set of
magnitude spectra with random phase. Time stretching or transposition of the residual
can be accomplished by interpolating new magnitude spectra | Êl (k )|, setting the phase
equation 2.13 (page 20) in which the window is not time reversed. Auger and Flandrin and Plante et al. define
their STFT with a time reversed window. Thanks to Kelly Fitz for explaining the reason for this difference.
spectra to random values between −π and π, applying the inverse DFT, and then overlap
adding.
Optionally | El (k )| can be computed directly in the frequency domain by subtracting the
frequency domain peaks detected in frame l from | Xl (k )|. | El (k )| may also be approximated
as a spectral envelope. The spectral envelope can be viewed as a filter that shapes a flat
noise spectrum.
Several variants of the deterministic plus stochastic model have been proposed. Pampin
(2004) developed the ATS system which separates the energy in El (k ) into 25 bands; each
band corresponding to a step on the Bark frequency scale. The noise energy in each band
may then be optionally redistributed among the partials.
3.6.1 Bandwidth Enhanced Sinusoids
This approach is similar to the bandwidth enhanced additive sound model described in
Fitz (1999) and implemented in Loris. Bandwidth enhanced synthesis models the signal
as a sum of bandwidth enhanced oscillators — sinusoidal oscillators that are amplitude
modulated by a lowpass filtered noise signal χ(n).

2πF (n)
y(n) = [ A(n) + β(n)χ(n)] cos n (3.27)
fs
Equation 3.27 defines a bandwidth enhanced oscillator with time varying amplitude A(n)
and time varying frequency F (n). The amplitude of the lowpass filtered noise is controlled
by β(n). Multiplication of a sinusoid by noise in the time domain results in the convolution
of the noise spectrum with the sinusoid spectrum. The multiplication shifts the noise
spectrum so that it is centered at frequency F (n).
In Loris, each breakpoint has an additional noise parameter κ that varies between 0
and 1. If Â(n) is the local average partial energy, the the bandwidth enhanced oscillator is
defined by
h√ √ i
2πF (n)

y(n) = Â(n) 1−κ+ 2κχ(n) cos n (3.28)
fs
When κ = 0, y(n) is a pure sinusoid. As κ increases, the amplitude of the narrowband
noise also increases. This model has the advantage that noise is tightly coupled to the
sinusoidal partials. As the sound is edited and manipulated — for example as partials are
transposed — the noise will follow. It is sometimes convenient that the noise amplitude is
controlled by a single parameter κ. However since the noise parameter is coupled to overall
partial amplitude, several square root operations must be invoked, which is somewhat
inefficient for realtime resynthesis.
SPEAR currently does not implement noise analysis, but it will resynthesize noise of
partials imported from Loris SDIF files (see section 4.6 for more information about the
use of SDIF). Loris defines the lowpass filtered noise signal as white noise filtered by a
3rd order Chebychev IIR filter with cutoff of 500 Hz, and SPEAR attempts to match this
behavior. Although the Loris noise representation is compact (based on a single parameter
κ), it lacks generality and is not an ideal choice for cross program data exchange. A more
flexible system would specify noise amplitude and bandwidth independently of sinusoidal
amplitude. Such a data model would better support different noise analysis strategies and
would allow for experimentation with different noise bandwidths. Expanding the number
of noise parameters would require either a new SDIF matrix type or an extension of an
existing type. Expanding the noise handling capabilities of SPEAR is an important area
for future work.

64
4. Editing and Transformation
4.1 Graphical User Interface
One of the motivations behind the development of SPEAR was the desire for a graphical
interface that would allow the user to make a quick visual assessment of analysis quality.
In spite of the best efforts in the choice and implementation of the analysis algorithms,
sonic artifacts are common. Even with a reduced set of available analysis parameters,
some trial and error is usually necessary to determine the optimal settings for a particular
application.
With systems that are command-line oriented (Loris, ATS, SNDAN), the user is left
to either resynthesize and listen carefully for artifacts, or use an additional plotting or
visualization tool. SPEAR offers an intuitive interface that allows the user to quickly zoom
in on analysis artifacts. Problematic areas of the sound can be isolated and auditioned,
and the data can be modified as desired. For example, spurious sinusoidal tracks can be
deleted, attenuated, or joined with other neighboring tracks.
The analysis data is displayed with time on the abscissa and frequency on the ordinate.
As in a grayscale spectrogram, varying levels of shading indicate the relative amplitudes
of partials. Figure 4.1 illustrates the overall user interface. Following the model of time
domain waveform editors, the amount of detail shown for the breakpoint functions varies
according to a user controlled zoom factor. At high zoom levels, every breakpoint is shown
and clickable handles allow individual manipulation in time and frequency (figure 4.2).
At lower zoom levels, the decrease in detail avoids visual clutter and significantly speeds
redraw.
4. Editing and Transformation 65
Figure 4.1: Zoomed-out display showing lasso selection and playback control sliders.
Figure 4.2: Zoomed-in display showing amplitude adjustment dialog.

point/partial selection scroll
selection time region select frequency region
rectangular selection lasso selection
frequency shift transpose selection
time/frequency shift time shift selection
time stretch/compress playback scrub
draw partial
Figure 4.3: The SPEAR tool palette.
A tool palette (figure 4.3) offers different modes for graphical selection and manipula-
tion. Interaction follows a selection and direct manipulation paradigm (Shneiderman 1983).
For example, clicking on a partial selects it and shift-clicking adds partials to the current
selection. Choosing the transpose tool then allows the selected partials to be transposed by
dragging up or down. Additional selection modes allow the user to sweep out areas in
time and frequency or to draw arbitrary time-frequency regions with a “lasso.” Following
the model of graphics editors, arbitrary selections can be made up of a union or difference
of individual contiguous selections.
Editing tools and commands that operate on the current selection include transposition,
frequency shifting, time shifting, time stretching, and amplitude adjustment. Unlimited
undo/redo is supported for all editing operations, and most editing can be performed
while the sound is being synthesized. For further real-time control, sliders allow adjustment
of the overall transposition level, amplitude, and playback speed (figure 4.4). New partials
can be added by drawing with the pencil tool. New breakpoints are added with amplitude
proportional to the dragging speed.
Figure 4.4: The SPEAR playback control sliders.

4.2 Selection Model
During the development of SPEAR, two possible selection models and implementations
were considered. The first may be termed the “region selection“ method. In this model, a
selection is represented as a polygon (concave or convex) in the time-frequency plane. (It
would also be possible to represent the selection in the time-amplitude plane if desired.)
Selections would be defined by sweeping out regions graphically or by setting numeric
parameters. More complex selections could be represented as the union of several regions.
The second selection model is “point selection“ method. In this case, the selection is
represented as a collection of references to the specific breakpoints of interest.
Given the desire to allow both extremely detailed editing and complete flexibility in
the temporal positioning of breakpoints, it was decided that the point selection model was
best. To reduce implementation and interface complexity, this is the only selection model
that has been implemented to date.
There are a few drawbacks to this model. In some cases it may be desirable to define a
transformation (such as an amplitude envelope) that occurs across a certain time span. In
this case, the region selection model is ideal. When the user requests such an operation
on a point selection (for example, an amplitude fade), the bounding region of the point
selection must first be computed. It may also be desirable to define a time-frequency
region (such as a formant) that remains fixed when partials are shifted or transposed. In
this case, the region selection model is required.
Point selections are implemented by associating a selection list with each breakpoint
array. The selection is stored efficiently as a range set data structure (Phillips 1999). Rather
than maintaining the index of each breakpoint, the range set represents contiguous runs
of breakpoints by the start and ending points of the run. More formally, if n breakpoints
starting at index i are selected, the range set represents this selection as the pair (i, i + n),
rather than the list [i, i + 1, i + 2, . . . , i + n − 1]. Discontiguous selections are represented as

an ordered list of pairs: (i0 , i0 + n0 ), (i1 , i1 + n1 ), . . . , (i j , i j + n j ) . Given that contiguous
selections are the most common case, the range set is very efficient in both time and
space requirements. The range set also supports operations to efficiently add and coalesce
ranges, as well as delete and split ranges.
4.2.1 Rule-based Selection
Selections can also be made programatically. For example, the user can choose to select all
partials with a duration less than some specified value. Or, the user can choose to select
partials with an average amplitude less than some desired level. Currently the number of
choices for rule-based selection is quite limited, however one can imagine any number of
selection criteria: minimum or maximum frequency, average frequency, frequency range,
frequency periodicity (vibrato rate), phase stability, amplitude range, maximum amplitude,
etc. Comparison (=, >, <) and logical operations (NOT, AND, OR, XOR) could be used to
develop any number of rule-based selection criteria. The extent of possibilities suggests
that a scripting language (rather than a GUI) would be best way to implement flexible
rule-based selection. It is hoped that future versions of SPEAR will include this capability.
4.3 Editing Features
The SPEAR document model supports cut, copy, and paste of any selected data. When
pasting data from one document to another, care must be taken to make sure data types
match. Breakpoints are required to have time, frequency, and amplitude components. They
may also have phase and noise components, but these are optional (for a discussion of
the noise component, see section 3.6.1, page 62). Analyses performed with SPEAR will
have at least time, frequency, amplitude, and phase. Since analysis data may be imported
from other software and via different file formats (see section 4.6), the document model
must be prepared to support different breakpoint types. When pasting into an existing
document, the pasted data is modified to match the destination document’s breakpoint
type. For example, when pasting data that has only time, frequency, and amplitude into a
document that also has phase, the newly pasted data will have phases set to zero.
In future versions of SPEAR it may prove beneficial to allow greater flexibility of point
types. Additional data types could specify specify stereo panning, 3D spatial location,
noise bandwidth, etc. Partials are currently implemented as multi-dimensional arrays
with contiguous allocation. A flexible point model would be better implemented with
collections of one-dimensional arrays, each array corresponding to one data type.
contiguous multi-dimensional array
collection
of
arrays
Figure 4.5: Different breakpoint storage models.
SPEAR supports unlimited undo/redo for all editing operations. Early in the design
process it was decided that the undo implementation should be simple and robust, yet
flexible enough to support arbitrary edits and transformations. Each undo/redo operation
is modeled as an ordered collection of sub-operations consisting of any of the following:
• selection
• edit partial
• insert partial
• remove partial
These sub-operations are bundled together into a single undo or redo operation. An
undo stack stores the state before the operation (this is the data required to restore the
document to its previous state), while the redo stack stores state after the operation. To
avoid excessive dynamic memory use, the data for edit, insert, and remove operations is
stored on disk. Selections are stored in locally addressable memory. Generally this does
not present a storage problem since the range map data structure is quite compact.
A selection operation consists of the partial indices and breakpoint selection ranges.
An edit operation is defined as any operation on a partial that does not require the
insertion or removal of any other partials (and therefore does not change any partial index
numbers). Insert or remove operations are those that will alter the partial indices in the
data model. Operations that can split partials into multiple segments (such as cut or delete)
are modeled as a sequence of edit, insert, and remove operations. It is crucial that the
atomic sub-operations are ordered properly so that the precise order of partial indices is
always restored.
4.4 Time Expansion and Contraction
SPEAR supports two different modes for time expansion and contraction. The first is
the “independent“ mode. In this mode, each partial is expanded or contracted relative
to the starting time of the partial (or selection), which remains fixed in time. If tk (n)
represents the time of nth breakpoint for partial k, then the independent expansion for
breakpoint n given by the following tk (0) + α[tk (n) − tk (0)] where α is the expansion factor.
In “proportional“ expansion/contraction mode, the new time is relative to two fixed time
boundaries Tmin and Tmax . Breakpoints that fall before Tmin are left unchanged. Breakpoints
falling between Tmin and Tmax are scaled. Breakpoints the occur after Tmax are shifted
earlier (for contraction) or later (for expansion). The following equation summarizes:




 tk (n) for tk (n) ≤ Tmin

t0k (n) = Tmin + α[tk (n) − Tmin ] for tk (n) > Tmin ∧ tk (n) ≤ Tmax



 tk (n) + (α − 1) [ Tmax − Tmin ]

for tk (n) > Tmax
Proportional expansion/contraction scales the time base uniformly and accomplishes the
same thing as traditional time warping. Independent expansion/contraction can produce
interesting effects that “smear“ and overlap the evolution of each individual partial (in the
case of expansion), or emphasize short sinusoidal segments (in the case of contraction).
Figures 4.6–4.8 on page 71 illustrate the different time expansion modes.

Figure 4.6: Original partials prior to time expansion.
Figure 4.7: Partials time expanded in independent mode. Note that the start time of each
partial is the same as in the unexpanded version.
Figure 4.8: Partials time expanded in proportional mode. Note that the start time of each
partial is scaled relative to time zero.
4.5 Frequency Modifications
Currently the SPEAR GUI only supports a few basic frequency-based transformations.
Frequency shifting and transposition can be performed either by dragging the selection
vertically with the appropriate tool or by entering a specific frequency shift (in hertz) or
transposition (in floating point semitones). The interface is shown in figure 4.9
Figure 4.9: Dialog boxes for transposition and frequency shifting.
Frequency “flipping“ inverts all frequencies in the range Fmin to Fmax around the
axis 12 ( Fmin + Fmax ). For frequency f k (n) of partial k and breakpoint n, the new flipped
frequency f k0 (n) is given by
f k0 (n) = Fmax − ( f k (n) − Fmin )
Frequency flipping can be effective when applied to specific frequency bands. For example,
one might wish to try flipping the 3rd through 11th partials of an inharmonic timbre.
Figure 4.10 illustrates the effect. Frequency flipping tends to be less interesting when
applied to an entire instrumental sound, as the strong lower partials invert into high
frequency regions (often above 10 kHz). Strong energy in this area is often perceived
simply as a loud noise band.
4.6 Data Exchange
A central feature of SPEAR is support for data exchange with existing analysis synthesis
packages. For example, one might wish to use SPEAR’s visualization to compare the
analysis results from different software. The Sound Description Interchange Format (SDIF)
(Wright et al. 1999a) is supported both for import and export.

(a) Selected partials of a bell sound prior to frequency flipping.
(b) Selected partials of a bell sound after frequency flipping.
Figure 4.10: Frequency flipping.
SDIF was developed out of a need for a standard file format that could represent
sounds in ways other than as a sequence of time domain samples. SDIF is particularly
well-suited to storing spectral models. An SDIF file consists of a sequence of frames that
are tagged with a 64-bit floating point time value (measured in seconds). Each frame is
also tagged with a 32-bit integer stream ID which allows multiple logical channels per
file. Each frame consists of one or more matrices that contain the data. Various standard
matrix types, which are tagged with a 4-byte code, have been defined for different types of
analysis data and are shown in table 4.1.
For SDIF import, SPEAR supports 1TRC (sinusoidal tracks), 1HRM (harmonic sinu-
soidal tracks), RBEP (reassigned bandwidth enhanced partials — a matrix type used by
Matrix Type Description

1FQ0 Fundamental frequency estimates
1STF Short time Fourier transform data. The time tag
in a 1STF frame is the time of the center of the
window, not the beginning.
1PIC Picked spectral peaks with amplitude, frequency,
phase, and confidence values
1TRC Sinusoidal tracks with track index, amplitude,
frequency, and phase values
1HRM Sinusoidal tracks with track harmonic number,
amplitude, frequency, and phase values
1RES Exponentially decaying sinusoids/resonances
with amplitude, frequency, decay rate, and ini-
tial phase values
1TDS Time domain samples
RBEP Reassigned, bandwidth enhanced sinusoidal
tracks with track index, amplitude, frequency,
phase, time-offset, and noise values
1ENV Sampled spectral envelope
1NVT Name-value table storing text or numerical
header information (value) for arbitrary keys
(name)
EASM Matrix storing all breakpoints of single partial.
Supported by SDIF-Edit (Bresson and Agon 2004).
Table 4.1: Some of the standard SDIF matrix types.
Loris), and 1STF (short-time Fourier transform frames). Data can be exported either as
1TRC frames or RBEP frames. In the case of 1TRC frames, resampling of the breakpoint
functions must be performed to conform to the fixed frames of the 1TRC file type. For
RBEP frames, each point has a time offset value. By collating all breakpoints into frame-
sized chunks and properly setting the time offset, any distribution of breakpoints can be
losslessly represented with an RBEP SDIF file.
Additional formats are supported for importing — the .mq and .an formats from
SNDAN (Beauchamp 1993) and the .ats format from the ATS package (Pampin 2004). The
data exchange options are summarized in table 4.2.
Two custom text file formats have been implemented for both import and export.
par-text-frame-format represents resampled frames, where each line of the file con-
Format Import Export

SDIF 1TRC × ×
SDIF RBEP × ×
SDIF 1HRM ×
SDIF 1STF ×
SNDAN .mq ×
SNDAN .pv ×
ATS .ats ×
Text frames × ×
Text partials × ×
Table 4.2: Supported file formats.
tains the data of a single time frame. Index numbers are used to link sinusoidal tracks.
par-text-partials-format represents exact breakpoint functions, where each line of the
file represents a complete partial. Each line is separated by a newline character. Both for-
mats begin with a preamble. The first line of the preamble is either par-text-frame-format
or par-text-partials-format depending on the type of file. The second line specifies the
data types for each breakpoint. Subsequent lines give data about the number of partials
and/or frames. The final line of the preamble is either frame-data or partials-data de-
pending on the type. Figures 4.11–4.12 show the frame based format and figures 4.13–4.14
show the partials based format.
Although the SPEAR graphical interface is intuitive and fairly feature complete, it
may not offer enough flexibility for more advanced uses. For example, batch analysis and
transformation of many sound files would be quite tedious. The ability to import and
export data gives motivated users the option to implement batch transformations with
other applications (see section 5.3 for details). The motivation for creating a GUI based
analysis/synthesis tool arose from the dearth of options. Existing options such as SNDAN,
ATS, Loris, and IRCAM’s Additive were all primarily command-line based applications.
The ideal environment would offer strong graphical and scripting capabilities, and it is
hoped that future versions of SPEAR will incorporate a scripting engine to facilitate a
wider variety of use cases.

par-text-frame-format
point-type index frequency amplitude
partials-count <J>
frame-count <K>
frame-data
<frame-time0> <N> <index0> <freq0> <amp0> ... <indexN> <freqN> <ampN>
<frame-time1> <N> <index0> <freq0> <amp0> ... <indexN> <freqN> <ampN>
...
<frame-timeK> <N> <index0> <freq0> <amp0> ... <indexN> <freqN> <ampN>
Figure 4.11: The par-text-frame-format specification. Following the frame-data line, each line contains the breakpoints for one
frame. N indicates the number of peaks in each frame. The index values connect peaks from frame to frame. Each line is separated
by a newline character.
par-text-frame-format
point-type index frequency amplitude
partials-count 5
frame-count 7
frame-data
0.000000 5 4 79.209549 0.000271 3 199.714951 0.000551 2 350.347504 0.002547 1 540.650085 0.000262 0 708.789856 0.000477
0.010000 5 4 77.832207 0.000453 3 199.756546 0.000949 2 351.608307 0.013954 1 529.902954 0.000625 0 704.903259 0.004945
0.020000 5 4 75.407600 0.000525 3 201.211563 0.001151 2 352.697296 0.034572 1 533.938293 0.000891 0 701.399414 0.020578
0.030000 5 4 74.918587 0.000409 3 199.946182 0.001182 2 352.384399 0.051492 1 552.429321 0.001080 0 702.354309 0.042789
0.040000 5 4 75.172478 0.000174 3 198.636566 0.001109 2 352.013580 0.059633 1 564.025574 0.000642 0 704.860168 0.063039
0.050000 3 3 199.029831 0.001001 2 353.066437 0.061233 0 706.239319 0.076010
0.060000 3 3 199.642029 0.000907 2 355.612793 0.061363 0 706.990723 0.080715
Figure 4.12: Sample data as par-text-frame-format. There are a total of 5 partials and 7 frames.
76
par-text-partials-format
point-type time frequency amplitude
partials-count <J>
partials-data
<index0> <P> <start-time> <end-time>
<time0> <freq0> <amp0> ... <timeP> <freqP> <ampP>
<index1> <P> <start-time> <end-time>
...
<indexJ> <P> <start-time> <end-time>
Figure 4.13: The par-text-partials-format specification. Following the partials-data line, each pair of lines contains the data
for one partial. The first line of each pair gives overall data for each partial: index number, length (P), start time, and end time. The
second line gives the breakpoints. Each line is separated by a newline character.
par-text-partials-format
point-type time frequency amplitude
partials-count 5
partials-data
0 5 0.000000 0.046440
0.000000 540.650085 0.000262 0.011610 528.172668 0.000683 0.023220 536.151062 0.000970 0.034830 564.025574 0.001158 ...
1 7 0.000000 0.069660
0.000000 350.347504 0.002547 0.011610 351.811279 0.015791 0.023220 353.037323 0.041780 0.034830 351.919281 0.058412 ...
2 7 0.000000 0.069660
0.000000 199.714951 0.000551 0.011610 199.763245 0.001013 0.023220 201.767395 0.001204 0.034830 198.648788 0.001166 ...
3 5 0.000000 0.046440
0.000000 79.209549 0.000271 0.011610 77.610458 0.000483 0.023220 74.562180 0.000542 0.034830 75.172478 0.000314 ...
4 6 0.011610 0.069660
0.011610 704.277527 0.005664 0.023220 700.294800 0.026302 0.034830 703.821472 0.054533 0.046440 706.153931 0.073634 ...
Figure 4.14: Sample data as par-text-partials-format. There are a total of 5 partials.

77
78
5. Compositional Applications
The introductory chapter outlined a number of challenges composers face when working
with the computer: the time/cost problem, the synthesis problem, the control problem,
and the compositional problem. Thus far we have seen that SPEAR most directly addresses
the time/cost problem (by providing fast synthesis) and the control problem (by providing
an agile and highly intuitive user interface). We now turn in earnest to the composition
problem. Musical tools, such as SPEAR, are never neutral agents in the compositional
process. Whether deliberate or not, musical intentions are always latent in the design of
computer music software systems. SPEAR was conceived with particular attention to the
needs of spectral composition.
5.1 Spectral Composition
Spectral composition might be seen as subset of the more general notion of timbral
composition. Many of the notable practitioners of spectral composition are quick to eschew
the idea of a specific “spectral style.” Nevertheless, developments over the past thirty
years do point to consistencies which allow us to formulate a coherent conception of
spectral music. The wariness of labels stems in part from a mistrust of compositional
orthodoxies, serialism in particular, from which many spectral composers were trying
to escape. A central tenet of spectral composition in fact centers around the rejection of
discrete categories in favor of continuums. This distinguishes the spectral approach from
other notions of timbral composition which are more oriented toward categorical and
hierarchical notions: Fred Lerdahl’s Timbral Hierarchies (1987), Pierre Schaefer’s Solfège de
l’objet sonore, or James Dashow’s Dyad System (1999).

5. Compositional Applications 79
Spectral composition can be understood from several perspectives. From an aesthetic
point of view, the notion of continuity and process is central. Tristan Murail speaks in
particular about the organization of continuous frequency space:
“Frequency space is continuous and acoustical reality only has to define its
own temperaments. If we push this reasoning to an extreme, the combination
of pure frequencies could be used to explain all past categories of musical
discourse and all future ones. Harmony, melody, counterpoint, orchestration,
etc., become outdated and are included in larger concepts.” (Murail 1984)
Gérard Grisey holds a complementary view that locates notions of continuity and process
in the temporal domain.
“For me, spectral music has a temporal origin. It was necessary at a particular
moment in our history to give form to the exploration of an extremely dilated
time and to allow the finest degree of control for the transition from one sound
to the next. . . . From its beginnings, this music has been characterized by
the hypnotic power of slowness and by a virtual obsession with continuity,
thresholds, transience and dynamic forms.” (Grisey 2000)
Joanthan Harvey’s view of spectralism stresses the potentially spiritual aspect of the
approach: “History seems grand, for once; spectralism is a moment of fundamental shift
after which thinking about music can never be quite the same again. . . . spectralism
in its simplest form as color-thinking, is a spiritual breakthrough.” Later he comments
specifically:
“In several works, to take a simple example, violins provide upper harmonics
to a louder, lower fundamental and at a given moment they cease to fuse, begin
to vibrate, begin to move with independent intervals and then again return to
their previous state. The images of union and individuation are powerful ones
which have both psychological and mystical implications. ‘The Many and the
One.’” (Harvey 1999)
It is notable that all three composers make a particular point of spectralism’s place
in history. Composers of the postwar generation seemed particularly keen to define
spectralism as a coherent alternative to what they perceived as the cognitively opaque
results of serialism. Spectral composition nevertheless gravitates toward a formalist
approach that privileges individual compositional control. Technology has tended to

function as a means of gaining precision and power. It remains an interesting open area
for exploration to see how spectral approaches might be applied in improvisational or
indeterminate compositional situations.
From a technical point of view, spectral music engages a number of specific concerns.
Pitch is organized in a continuous fashion and pitch structures are often reckoned directly
in cycles per second. In instrumental writing, approximations to the nearest quarter-
tone (or in some cases eighth tone) are used. This is in marked contrast to a precise
just intonation pitch specification.1 The harmonic series serves as an important point of
departure, and often functions as a stable pole (if not the stable pole) in the harmonic
discourse. Harmony often proceeds by a differential approach which unfolds processually.
The notion of “instrumental additive synthesis” often features prominently in spectral
composition. A combination of instrumental sounds with the appropriate frequencies and
intensities can fuse into a single timbral entity. In contrast to some approaches to timbral
composition, instrumental color tends to be exploited not for its differentiating abilities
but for its potential to participate in spectral fusion. Thus individual instrumental timbres
are subsumed into larger musical structures.2 In some spectral works, harmonic models
are derived from characteristic spectra of instrumental sounds. The goal of instrumental
additive synthesis is never to recreate, but rather to reveal latent musical potential is
pre-exisiting sonic material. The archetypal gesture of instrumental additive synthesis is
the arpeggiated spectra which gradually transforms from a series of individual tones to a
fused complex.
As with pitch, the approach to rhythm derives from a temporal sensibility that is
continuous rather than discrete. Rhythmic processes often involve controlled long range
accelerations and decelerations. Additive and subtractive rhythms or rhythmic interpola-
tions (as in the music of Philippe Hurel) are another common strategy. Overlaid rhythmic
1 The approximation approach tends to be found in the music of the French spectralists and their followers.
American music that might be identified as spectral or proto-spectral has tended to follow a just intonation
approach (as in the compositions of James Tenney or Ben Johnston).
2 This technique is a hallmark of spectral music. While instrumental timbre played an important role in
some proto-spectral music (that of Varèse in particular) timbre tended to be more about making elements
more distinct (Varèse’s planes of sound) rather than fusing them together.
pulsations with various rational relationships are also encountered (particularly in the
music of Grisey). In many cases rhythms are approximated, either via quantization or
proportional notation.
A complete summary of all technical aspects found in spectral music is beyond the
scope of this document (see Fineberg (2000) for more details). For our discussion, the
following list of some of the central concerns of spectral music will suffice:
• continuous frequency space
• continuous temporal space
• timbre-harmony continuum
• the use of approximations
• global approach to form
• smooth transitions contrasted with ruptures
• process preferred over development
• appeal to perceptual and psychoacoustic aspects of sound
The last item points to a particularly anti-formalist aspect of spectral music. For all
its technical machinations, spectral music is very much “for the listener.” However, this
tendency to appeal to cognitive “universals” raises interesting concerns. For example,
to what degree might the expressive or communicative ability of the spectral approach
be limited by this focus? In particular, does the pre-occupation with the purely acoustic
preclude the possibility of addressing culture-specific aspects of the musical enterprise?
These open questions point to interesting possible directions for future compositional
work.
Technology has always played an important role in the development of spectral music.
Models derived from electronic processes such as frequency shifting, ring modulation,
frequency modulation, or tape delay can be found in any number of spectral pieces.
The desire for mastery of the sonic continuum led naturally to the use of the digital
computer.3 Digital sound synthesis is complementary to instrumental synthesis and

3 Ironically, one gains reproducible control of the continuum in the discrete world of the computer.
spectral works that combine electronic and instrumental sounds can achieve remarkable
coherence. Computer-assisted composition environments, which will be discussed in
section 5.3, are indispensable for realizing sonic processes. SPEAR is a particularly useful
tool for spectral composition. The high performance GUI makes it possible to visualize
and manipulate complex sounds. Its detailed analyses can be used in the derivation of
harmonic and/or rhythmic models, or the data can be harnessed directly to drive various
forms of digital synthesis.
5.2 Sonic Transformations
A sinusoidal analysis may be viewed as a reservoir of sonic data with almost limitless
musical potential. Possible temporal transformations include replication, time delay, time
dilatation, and time reversal. Amplitude envelopes may be applied (for example, to fade
partials in and out, or to emphasize the attack portion of a sound). Partial frequencies may
be altered by any desired transposition or frequency shifting envelope. The analysis data
can be used to drive other compositional or synthesis processes (for example, granular
synthesis). The following subsections detail a number of possible sonic transformations.
5.2.1 Timbre Hybridization
Properties of different spectral analyses can be combined to create new hybrids. This
process of timbre hybridization is sometimes referred to as cross synthesis. There are many
possible cross synthesis implementations, some of which are found in software such as
AudioSculpt, SoundHack, Shapee, and in commercial plugins such as Prosoniq Morph.
Many of these techniques focus on different ways of combining STFT frames or applying
properties of one STFT frame to another. Sinusoidal modeling offers the possibility of very
fine-grained control over the cross synthesis process.
Dynamic morphing between two collections of partials A and B can be achieved by
matching common partials and interpolating their frequencies and amplitudes. Such a
procedure has been implemented as in Common LISP as follows:

1. Temporally align the two sounds by means of timebase envelopes that map output
times to time points in the original sound.
2. Define an interpolation envelope that maps output time to relative fraction of A or
B that should be present in the output — a value of 0 indicates the sound should
consist of all A, 1 indicates all B, intermediate values should be a blend of A and B.
3. Compute a succession of blended spectral frames to be synthesized. This proceeds
as follows:
• Advance an output time variable t by a predetermined hop.
• For each value of t, determine the active partials in A and B and compute two
frames At and Bt with the interpolated frequencies and amplitudes of the active
partials at time t.
• For each value of t, determine the current blend factor x from the interpolation
envelope.
• Determine a set of matching frequency/amplitude points in At and Bt . Matching
is controlled by a frequency tolerance. Interpolate the amplitude and frequency
of each of the matching points according to the blend factor x. Unmatched
points can be discarded or have their amplitude adjusted according to x (thereby
creating a cross fade of the unmatched partials).
4. Synthesize the blended frame and window with a tapered window function. Overlap
add into the output.
This morphing algorithm is somewhat crude, but is reasonably effective for certain kinds
of sounds. It could be improved by computing a new set of blended partials, rather than
overlap adding phase-incoherent frames in a granular fashion. As it is, the algorithm only
works well if sounds contain similar frequency content. Harmonic sounds need to share
the same fundamental frequency. A more sophisticated algorithm, such as that used in
Loris, would tag the harmonics of A and B relative to a fundamental frequency track.
Like harmonics can then be interpolated and smoothly glide to new frequencies. For the
specifics of the Loris morphing algorithms see Fitz et al. (2003).

A more sophisticated algorithm should also be able to interpolate between more than
two sources. It would be musically useful to create long sequences of seamless morphs
between a series of sounds as with IRCAM’s Diphone Studio. Despite its limitations, this
basic algorithm shows promise and has been used to create some effective morphs between
harmonic and inharmonic timbres.
Another approach to cross synthesis is to keep the frequency content fixed and ma-
nipulate amplitudes to affect a change in the spectral envelope (see section 2.3 for a
brief discussion of spectral envelopes). A hybrid can be created by imposing the spectral
envelope of one sound on another. Although SPEAR does not currently include spectral
envelope capabilities, a number of Common LISP functions have been developed for
manipulating spectral envelopes and applying them to SPEAR analyses (see table 5.3). The
functions are designed to work with spectral envelope data generated by SuperVP, a pow-
erful command line analysis/synthesis tool developed at IRCAM. Successful experiments
have been conducted with the application of vocal spectral envelopes to tam-tam sounds
(and visa-versa).
5.2.2 Spectral Tuning
Spectral tuning adjusts the frequency content of a sound to match the frequencies of a
predetermined spectrum or harmonic field. The target frequencies might be derived from
any number of sources: octave repeating scales or chords, synthetic spectra, frequencies
derived from other spectral analyses, distortion and warping of existing spectra, etc.
Spectral tuning using the phase vocoder has been explored extensively by a number of
composers and researchers including Eric Lyon and Christopher Penrose with FFTease
(Lyon 2004), Paul Koonce with PVC, and Trevor Wishart with Composers Desktop Project
(CDP) (Wishart 2000).
Spectral tuning with a sinusoidal model allows very precise control of the evolution of
each partial. In a typical application, the tuning is static: the local amplitude and pitch
fluctuations of each partial are maintained and the partial is transposed so that its average
frequency matches the target. Local pitch fluctuations can be attenuated or emphasized
as desired. Dynamic tuning transposes a partial according to a variable function. If the
rate of frequency change is not too rapid, the retuning can still maintain local frequency
perturbations which are crucial to the character of most natural sounds. A simple example
of dynamic tuning would be the application of vibrato or glissando.
Parameters can control the strength of retuning. For example if a partial is separated by
a distance of 3 semitones from its target, a retuning strength of 70% would transpose the
target by 2.153 semitones rather than the full 3 semitones. Or a parameter can specify that
partials are retuned only if they are in close enough proximity to the target. Particularly
interesting results have been achieved with progressively retuning inharmonic timbres so
that a succession of sounds eventually converges on a harmonic spectrum. Spectral tuning
can also be used for cross synthesis where the target frequencies are derived from analysis
of a different sound.
5.3 Software Composition Environments
Although a great deal of useful sonic manipulation and transformation can be accom-
plished directly with the SPEAR GUI, much of the software’s compositional utility is
revealed when it is used in conjunction with other tools. Computer-assisted composition
and performance environments such as OpenMusic, Common Music, Max/MSP, RTCmix,
and SuperCollider (to name a few), can be used to import, manipulate, export, and/or
synthesize SPEAR analysis data.
SPEAR’s support for text based file formats (as well as standard SDIF formats) makes
cross program data exchange relatively straightforward. The text formats are particularly
useful because they are human readable and easy to use with interactive interpreted
programming languages such as Java, Python, Scheme, and Common LISP. Import and
export of the SDIF formats is more difficult and time consuming to implement for casual
programmers. The author’s own experience has shown that even though SDIF is robust and
well-defined, it is not necessarily widely understood among computer music practitioners.

Even with supporting SDIF libraries such as those provided by CNMAT or IRCAM, it may
prove more expedient (particularly in one-off applications) to use the text data formats.
5.3.1 OpenMusic
OpenMusic is a computer-assisted composition environment developed at IRCAM using
the Common LISP language. It is a visual programming environment based on patching of
pre-defined or user-defined objects. The patch paradigm provides full access to Common
LISP programming, including iteration and recursion. Nodes in the object graph are
evaluated for their results and possible side-effects. A user request for evaluation will
trigger evaluation of all parent nodes (figure 5.1). OpenMusic includes graphical editors for
MIDI data, breakpoint functions, and common practice notation (figure 5.2). OpenMusic
was developed as a successor to PatchWork a similar LISP based environment for visual
programming.
Figure 5.1: OpenMusic patch that evaluates the arithmetic expression 10 + (2 × 10) .
The environment may be extended with user defined libraries. New OpenMusic func-
tion are defined as CLOS (Common LISP Object System) methods using the om:defmethod!
macro. SPEAR import and export has been integrated with OMTristan, a custom OM
library created by Tristan Murail. OMTristan, which has evolved over decades, includes a
wide variety of compositional algorithms for the generation of chords and spectra, tempo-
ral computations, MIDI communication, and general data manipulation. The spdata object
in OMTristan is specifically designed for efficient storage and manipulation of frame based
spectral data. The methods spear-read and spear-write convert between the SPEAR
par-text-frame-format file type and OM spdata objects.

Figure 5.2: OpenMusic notation editor displaying a chord sequence.
Figure 5.3 shows an OM patch that reads SPEAR data, transforms the data, displays
the transformed data in common practice notation, and writes out the transformed data
to a new text file. The patch functions as follows: The spear-read object (located at the
top) takes the name of a text file as input and outputs an spdata object which is stored
as an embedded object (this avoids having to re-read the data from disk on subsequent
evaluations). The spdata output is split into three streams: the frequencies, the amplitudes,
and the partial index numbers. Each stream is a LISP list of two elements. The first element
is a list of times for each frame and the second element is a list of lists where each sublist is
the frame data (either frequencies, amplitudes, or indices). The time base is stretched (the
“dilation” parameter) and the frequencies distorted in the dist-frq sub-patch. The data is
reassembled into frames and converted to a chord sequence in the visu-chseq sub-patch.
The times, indices, frequencies, and amplitudes are also routed to the spear-write object
which outputs a new text file with the transformed data. The data can then be opened in
SPEAR for resynthesis or further transformation.
5.3.2 Max/MSP
Max/MSP is a widely used graphical patching environment for realtime media processing.
SPEAR SDIF data can be imported into Max/MSP using the CNMAT SDIF objects (Wright
et al. 1999b). The SDIF-buffer stores the SDIF data in a named buffer which can be
accessed using the SDIF-tuples object. SDIF-tuples outputs matrix data at the specified
Figure 5.3: OpenMusic patch that reads and writes SPEAR data.
time (with optional interpolation) as a Max list. CNMAT’s sinusoids~ object implements
an oscillator bank (with optional bandwidth enhancement as described in section 3.6.1)
that is driven by a list of frequency amplitude pairs. The threefates object can be used to
manage the sinusoidal track indices output from the SDIF buffer and format the data into
frequency amplitude pairs suitable for the sinusoids~ object.
0., 10. 10000 7.1, 2.1 5000 loadbang loadbang
read-stream-number moanin.add.sdif 0 read-stream-number kyorei.add.sdif 0

line 0. 10.
SDIF-buffer ss-toot sdif-buffer ss-boot
2.1
shakuhachi sax
tff
tuples time $1 columns 1 2 3 tuples time $1 columns 1 2 3 0.
SDIF-tuples ss-toot SDIF-tuples ss-boot
threefates threefates
zl slice 100 zl slice 100 0.
tbl prepend sinusoids frequency-scale $1
f 1. sin-transform
list-interpolate
sinusoids~
List-interpolate with sines
While this fast and dirty interpolation yields some
results, it would be more musical if we paid
attention to the actual meaning of the SDIF data.
Trying to interpolate between timbre, contour, and
amplitude envelope all it once is a crazy idea, if
you think about it.
•!start audio
startwindow stop
dac~
Figure 5.4: Max/MSP patch from the CNMAT Spectral Synthesis Tutorials that interpolates
between two sinusoidal models to achieve a dynamically controllable cross synthesis.
The limitation of 256 elements with the standard Max list objects does put some
constraints on the complexity of the spectral data. Michael Zbyszynski’s “Spectral Synthesis
Tutorials” suggest some solutions (Zbyszynski et al. 2007). Possibilities include the use of
custom C or Java objects, the use of Javascript, the use of Jitter matrix processing objects,
or the use of IRCAM’s FTM extensions.

The use of sinusoidal models in the Max/MSP/Jitter environment offers numerous
options for interactive transformation and resynthesis. The ability to manipulate spectral
data directly in Max using Java and Javascript eases the burden of managing large spectral
data sets. Future work could lead to the development of “partial based” storage models
rather than the frame based approach implicit in the SDIF 1TRC format. As we shall
see in the next section, the partials based format (collections of breakpoint functions) is
particularly well suited for certain types of musical operations.
5.3.3 Common Music
Common Music (CM) is an algorithmic composition environment developed by Heinrich
Taube. Like OpenMusic, CM is implemented in Common LISP. It is built around the notion
of generalized input/output streams which can include MIDI files, MIDI ports, Open
Sound Control streams, CSound scores, CLM notelists, and CMN scores. Compositional
facilities include the scheduled processes, a full featured pattern iteration system, and a
flexible graphical interface called Plotter. Plotter facilitates display and editing of breakpoint
functions, histograms, scatter plots, and piano roll style MIDI data (Taube 2005). CM can
be tightly integrated with CLM (Common Lisp Music), a Music-N style acoustic compiler.
Current versions of CM include a spectral composition toolkit than can read SPEAR
data from par-text-frame-format and par-text-partials-format files. The data is stored
as nested LISP lists. For manipulating large sounds, it can be more efficient to store spectral
data in LISP vectors (contiguous arrays). A LISP toolkit has been developed for reading,
writing, and manipulating par-text-partials-format data. A complete sound is stored as
a LISP list of sp-partial data structures. Each sp-partial represents a single breakpoint
envelope with time, frequency, and amplitude arrays. Functions exist to transpose, time
warp, and amplitude scale partials. The adjustments can be controlled dynamically with
envelopes. The function for transforming single partials are summarized in table 5.1, page
91. Many of the transformation functions take an optional copy parameter which, when
(copy-partial partial)
Return a newly allocated copy of partial.
(offset-partial partial offset [copy])
Shift the time of partial by offset seconds.
(transpose-partial partial interval-ratio [copy])
Transpose the frequencies of partial by interval-ratio.
(transpose-partial-env partial interval-ratio strength-env [copy])
Transpose the frequencies of partial by a ratio ranging from 1 to interval-ratio. strength-env

is an envelope controlling the percentage of interval-ratio applied.
(time-scale-partial partial timescale [copy])
Apply a fixed time scaling ratio timescale to partial.
(amplitude-scale-partial partial ampscale [copy])
Apply a fixed amplitude scaling ratio ampscale to partial.
(reverse-partial partial [copy])
Reverse the temporal evolution of partial.
(amp-env-partial partial env [copy])
Apply an amplitude scaling envelope env to partial.
(sp-partial-avg-freq partial)
Compute the average frequency of partial.
(sp-partial-avg-amp partial [points])
Compute the average amplitude of partial. If points is non-nil, the average is computed based on
the first points breakpoints. If points is nil (the default) all breakpoints are used to compute the
average.
(sp-partial-interp partial time)
Interpolates the frequency and amplitude of partial at time seconds. The frequency and amplitude
are returned as multiple values (in that order). Uses a binary search for efficiency.
Table 5.1: Transformation functions applied to individual partials.
non-nil, applies the transformation to a new copy of partial, leaving the original unchanged.
The default value of copy is nil, meaning the transformation is applied destructively.
A retune-partials function can transpose a collection of partials to a new set of
target frequencies. The retune-partials function assumes partials with relatively constant
frequency (the retuning is based on the average frequency of the original partial). Functions
(retune-partials partials target {keyword value}*)
Transpose each of the partials of partials so that its average frequency matches a frequency
contained in the list target. Because the target frequency set may cover a vastly different frequency
range from the original partials, the average partial frequencies are first rescaled to lie within the
minimum and maximum frequencies of target. Each rescaled frequency is then matched to the
closest frequency in target. The matching is controlled by the following keyword arguments:
:low-limit Only tune partials above this frequncy. Default is 0 Hz.
:high-limit Only tune partials below this frequcny. Default is 3000 Hz.
:delay-curve Envelope to apply a frequency dependent delay to each partial. De-
fault is nil.
:amp-curve Envelope to apply frequency dependent amplitude scaling to each
partial. Default is nil.
:flatten Scale the ambitus of each partial. Values < 1.0 reduce the amount of
frequency variation of each partial. When 0.0, all frequencies are set
to the average frequency.
:strength Adjust the amount of frequency remapping. When 1.0, the partials
are retuned to exactly match the target frequencies. When 0.0, the
partials remain at their original frequencies.
:strength-curve Adjust the amount of frequency remapping over time. Partials can be
made to glide between original and target frequencies.
:index-scramble Scramble the targets such that the original partials will remap to
less proximate target frequencies. When 0, there is no scrambling
(the default). When 1, remap to the neighbor of the closest target
frequency. When n, remap to the nth neighbor (either above or
below).
(sp-partials-apply-vibrato partials {keyword value}*)
Apply slow frequency modulation (vibrato) to partials (a list of partials). This function requires
CLM. The following keyword arguments are supported:
:vibfreq The frequency of the vibrato in hertz. Default is 6.5.
:vibinterval The width of the vibrato in semitones. Default is 0.5.
:start-time Starting time for the onset of vibrato (in seconds). Default is 0.
:end-time Ending time for the vibrato (in seconds) or nil if vibrato continues to
the end. Default is nil.
:copy If non-nil, return new copies of the partials leaving the originals
unchanged. Default is true.
Table 5.2: Frequency adjustment functions applied to collections of partials.
for frequency transformation are summarized in table 5.2, page 92. Spectral envelope
functions are show in table 5.3, page 93.
5.4 Compositional Examples
A straightforward application of SPEAR is the derivation of harmonic models for instru-
mental composition. Pitches can be transcribed manually from the GUI or by importing
(sp-partials-flatten-spectral-env partials spectral-env [amp-scaler])
Flatten the spectral envelope of the sound represented by partials by dividing the partial ampli-
tudes by the spectral envelope spectral-env. The spectral envelope can be generated by reading
an SVP text formatted spectral envelope analysis. Apply an optional amplitude scaler amp-scaler
(default value 0.025) following the division.
(sp-partials-apply-spectral-env partials old-spectral-env new-spectral-env {keyword

value}*)
Apply a new spectral envelope new-spectral-env to the sound represented by partials.

old-spectral-envelope is the current spectral envelope of the sound. The following keyword
arguments are supported:
:transpose Transposition ratio.
:time-warp Time remapping envelope.
:freq-warp Frequency remapping envelope.
:freq-func Generalized frequency transformation. freq-func is a function of
three arguments: frequency, partial-index, and time. The function
should return the new desired frequency value.
:filter-func Generalized amplitude transformation. filter-func is a function
of four arguments: amplitude, frequency, partial-index, and time.
The function should return the new desired amplitude value.
:copy If non-nil, return new copies of the partials leaving the originals
unchanged. Default is true.
Table 5.3: Spectral envelope functions applied to collections of partials.
into other environments. Rather than relying on individual spectral snapshots, the evolu-
tion of a collection of partials can be incorporated into the composition model.
In the author’s work Subterrain (2008) for B[ clarinet, string trio, and electronics,
clarinet multiphonics were analyzed. The SPEAR GUI was indispensable for extracting
only the most stable and/or musically interesting partials from these complex sounds.
The data was imported into Common Music and amplitudes were averaged across entire
partials, which gave a very good sense of the overall spectral color of the multiphonic.
Further manipulation of the multiphonic spectra, including frequency shifting and ring
modulation, were used to create harmonies and pitch reservoirs.
Spectral retuning using SPEAR and OpenMusic was used to create a number of sounds
in Tristan Murail’s Pour adoucir le cours du temps (2005) for 18 instruments and synthesized
sound. In this piece, the partials of tam-tam sounds were retuned to the frequencies of
artificially generated spectra. A cow-bell analysis also undergoes similar retuning. Of the
work, Murail writes,

“The ‘noises’ (breath, grainy sounds, metallic resonances) of the piece have
been domesticated, tuned to instrumental harmonies through the use of a
specific technique for altering the internal components of the sounds. This is
how the sound of a gong can become a harmony, or the virtual cow-bells can
be constantly varied and altered by their musical context. The synthesized
sounds are triggered from an onstage MIDI keyboard — they are mixed and
spatialized in real-time by the program Holophon developed by the GMEM.”
(Murail 2005)
Spectral tuning was used to generate several of the sounds used in the author’s work
Tear of the Clouds (2007) (see appendix A for the complete score). The low vocal sounds
(at m. 14, reh. A and m. 203) were created by retuning and resynthesizing samples of a
choral ensemble. Although these particular source samples exhibited a high standard of
performance and excellent recording quality, they were not ideally suited for resynthesis.
Because they consisted of a sum of many ensemble voices (all in unison pitch), and because
they were recorded in a reverberant environment, partial tracking resulted in a large
number of short low amplitude partials in addition to the fundamental and harmonics. For
retuning purposes, a cleaner representation was desired. IRCAM’s Additive program was
used to perform fundamental frequency tracking and to analyze a fixed set of harmonic
partials (up to 22 kHz). The SDIF 1TRC file from Additive was imported into SPEAR and
then exported as a text file. This was then imported into LISP where the retuning was
performed.
At m. 14, the retuning target consists of selected partials of a compressed harmonic
spectrum built on G1 (49 Hz). The tuning was only applied to the original partials ranging
from 49 Hz to 1850 Hz. Because the original sound had a fundamental of 82 Hz, retuning
and compressing the entire spectrum would have resulted in a considerable loss of high
frequency presence. Since most of the harmonic “color” of a sound is found below 2000
Hz, the harmonics above 1850 Hz were retained in their original tuning (another option
would be to replicate partials to fill in the missing upper harmonics). The retuned data was
then imported into SPEAR and synthesized. Further post processing, including granular
time stretching, was then applied. The granulation produced a sound of the required
duration and helped re-impart some of the reverberant quality of the original sound that
was lost in the analysis.
Further applications of spectral tuning in Tear of the Clouds include the retuning of
tubular bell, piano, and tam-tam sounds. The penultimate electronic sound of the piece
is built from a retuning of a piano sound. The target spectrum consists of a distorted
harmonic series built on D]1 (39 Hz), in which the even partials are stretched and the
odd partials are unchanged. For this particular sound, the retuning is dynamic. The
attack maintains the original tuning and the partials glide to the target spectrum over a
period of 2.75 seconds. Rather than sliding to the nearest target frequency, the retuning
was configured so that the partials make wider frequency excursions (either up or down),
crossing each other as they proceed.
Figure 5.5: Dynamic retuning with crossing partials.
5.5 Performance Models
Although not extensively explored in the current work, some words are in order about the
potential applications in performance and improvisation. The SPEAR interface itself could
be viewed as a crude performance interface where one can scrub through the sound using
the mouse (Garton 2007). One can easily imagine a richer instrument that would give
control of elements such as playback speed, transposition, frequency shifting, harmonic
parity (balance of even and odd partials), etc. These potentials would ideally be realized
in a flexible environment such as Max/MSP, rather than in the more confining space of
the SPEAR user interface. A variety of gestural and tactile inputs could be used to drive
resynthesis parameters.
97
6. Conclusions
The preceding chapters have given a thorough account of the technical aspects of the
design and development of SPEAR while also touching on some possible compositional
applications. I do not wish to claim that SPEAR represents a revolutionary new computer
music tool. It has been preceded by more than fifteen years of research and development
in the field of spectral analysis and modeling. However, I hope it is clear that SPEAR is an
important evolutionary convergence of a number of mature technologies.
Although SPEAR has experienced reasonable popularity among a wide variety of
technologically oriented music practitioners, the software is still very much a work-in-
progress. Although the interface works well, it is still limited in many ways. A multi-
demensional display should be developed that allows one to view and edit all aspects
of a partial’s evolution: amplitude, noise bandwidth, phase, stereo (or multi-channel)
panning, etc. The user should be able to draw breakpoint envelopes and curves to control
the transformation of any parameter. Rule based selection (see section 4.2.1) should be
implemented. Moreover, an interface that allows one to specify persistent selections (like
audio regions in a waveform editor) would be incredibly useful. A user interface for
sound morphing and hybridization would also be welcome. Since it is impossible to
envision all of the possible desired editing and transformation possibilities, SPEAR needs
a built-in scripting language and command-line tools that allow users to develop their
own processes.
SPEAR must also keep pace with developments in spectral analysis and synthesis. To
build more compact and malleable spectral models, SPEAR needs robust fundamental
frequency tracking. The option should be available to perform analyses with fixed numbers
6. Conclusions 98
of harmonics (as with IRCAM’s Additive and Diphone programs). Models with fixed
numbers of partials are much easier to manipulate in real-time environments such as
Max/MSP or SuperCollider.
Some form of noise modeling, as discussed in section 3.6, should be implemented. The
transient sharpening method described in section 3.5 is successful to a point, but to handle
wide-band polyphonic audio and percussive attacks, a separate transient detection and
representation method is needed. The multi-resolution sinusoidal modeling techniques
described in Levine (1998) should also be implemented. Although multi-resolution model-
ing significantly increases the implementation complexity of partial tracking (particularly
at the band boundaries), it has the potential to considerably reduce pre- and post-echo
artifacts. The combination of time-reassignment, multi-resolution analysis, noise modeling,
and improved partial tracking would represent a formidable analysis engine.
In order to tie together these more sophisticated analysis techniques, a more advanced
user interface is required. Each component of the analysis would be assigned to its own
layer:
• fundamental frequency layer
• sinusoids layer
• noise layer
• transient layer
• spectral envelope layer
Each of these layers could be (optionally) coupled to the others. For example, when
layers are coupled together, adjustments to the fundamental frequency track would
correspondingly adjust the frequencies in the sinusoids layer. As the partials are transposed,
the amplitudes could be adjusted to match the spectral envelope layer. Peak detection and
tracking could be applied to the spectral envelope data so that formant trajectories and
widths could be adjusted in a manner analogous to individual partials. Given sufficient
time and effort, all of these ideas are feasible.

6. Conclusions 99
Returning to some of the issues raised in the opening chapter, we must consider what
a tool like SPEAR tells us about timbre and musical organization. For all its abilities,
SPEAR perhaps raises more questions than it answers. Although obvious, it’s worth
stating the overall conclusion: additive synthesis is not a sufficient model for timbre.
Although additive models are malleable, they tend to be fragile — minor adjustments to a
few partials can destroy timbral fusion and coherence. Higher level models are needed.
Spectral envelopes are a good first step. We should also look for models of the micro-level
evolution of partials. The amplitude and frequency fluctuations can be correlated between
partials and with parameters of the overall evolution of the sound. Finally, we must
recognize that the time varying aspects of spectra are often far more important than
the instantaneous relationships. We need timbral models than can capture micro-level
rhythm, grainy textures, and impulses. The TAPESTREA software, which uses a sinusoids
+ stochastic background + transients model, suggests a particularly promising approach to
high level timbre modeling (Misra et al. 2006).
From a compositional perspective there is clearly more to learn about timbre and
tools such as SPEAR can help with these compositional approaches. But it is not enough
for compositional work to merely reflect or reinforce habitual modes of understanding
timbre. Experimental composition can reveal new ways to listen to timbre, previously
unheard subtleties, surprising relationships, and unanticipated modes of organization.
With the acknowledgment that speculations on the future are often more about setting
an agenda than trying to make a neutral prediction, we might ask what role computer
music might play in this process and in turn what will be the nature of tools to come? Or
more precisely, what do we want from our musical tools of the future? Successful software
tends to embody at least some of the following attributes:
1. ease of use and configuration
2. high performance
3. large feature set
4. reliability
6. Conclusions 100
5. backward compatibility
6. initial novelty
7. openness and extensibility
At present, commercial software tends to excel at items 1–3, while failing miserably in
terms of openness and user extensibility. Extensibility has typically been the strength
of software with an experimental and open source pedigree. Commercial software is
frustrating because it tends to be expensive and creatively limiting. Free software is
frustrating because it tends to be difficult to configure and maintain. Of what benefit is
extensibility if one is left with no time to exploit it? Future software efforts must bridge
this gap. We must avoid the “sealed black boxes” of the commercial world and the arcane
configuration, instability, and continual tabula rasa upheavals of the open source world.
Negative proscriptions aside, what potentials are there for music software of the
future? An environment that melds a powerful acoustic compiler and general purpose
programming language with a flexible and high-performance GUI seems ideal. Such an
environment should include flexible time-based displays that can handle common practice
notation, breakpoint functions, audio data, and spectral models. It must not be a neutral
environment — it must make assumptions. It must assume a user who wishes to work
with common practice diatonic music and quantized meter and rhythm; it must assume
a user who wishes to work with millions of sonic granules on a proportional time scale;
it must assume a user who wishes to precisely notate irrational rhythms and microtonal
inflections; it must assume a user who wishes to capture and manipulate performance data
in real time. Technologically, this is all within the realm of possibility today. It remains to
be seen when, if, and/or how this may come about. Tools of this sort are not envisioned to
make composition facile. They should exist to engage and inspire the human imagination
and creative spirit.

101
References
Abe, Mototsugu, and Julius O. Smith. “Design Criteria for Simple Sinusoidal Parameter
Estimation based on Quadratic Interpolation of FFT Magnitude Peaks.” In Proceedings of
the 117th Audio Engineering Society Convention. AES, 2004.
Amatriain, Xavier, Maarten de Boer, Enrique Robledo, and David Garcia. “CLAM: An
OO Framework for Developing Audio and Music Applications.” In Proceedings of
17th Annual ACM Conference on Object-Oriented Programming, Systems, Languages and
Applications. Seattle, WA, 2002. http://mtg.upf.edu/publicacions.php.
Auger, F., and P. Flandrin. “Improving the Readability of Time-Frequency and Time-Scale
Representations by the Reassignment Method.” IEEE Transactions on Signal Processing 43:
(1995) 1068–1089.
Beauchamp, James W. “Unix Workstation Software for Analysis, Graphics, Modification,

and Synthesis of Musical Sounds.” In Proceedings of the 94th Audio Engineering Society
Convention. AES, 1993. Preprint no. 3479.
Bonada, Jordi, and Xavier Serra. “Synthesis of the Singing Voice by Performance Sampling
and Spectral Models.” IEEE Signal Processing Magazine 24, 2: (2007) 67–79.
Bresson, Jean, and Carlos Agon. “SDIF Sound Description Data Representation and
Manipulation in Computer Assisted Composition.” In Proceedings of the International
Computer Music Conference. Miami, FL: ICMC, 2004, 520–527.
Cadoz, Claude, Annie Luciani, and Jean Loup Florens. “CORDIS-ANIMA: A Modeling and
Simulation System for Sound and Image Synthesis: The General Formalism.” Computer
Music Journal 17, 1: (1993) 19–29.
Cormen, Thomas H., Charles E. Leiserson, and Ronald L. Rivest. Introduction to Algorithms.
Cambridge, MA: MIT Press, 1990.
Dashow, James. “The Dyad System (Parts Two and Three).” Perspectives of New Music 37,
2: (1999) 189–230. http://www.jamesdashow.net/download/DyadSystemPt2-3.pdf.
Depalle, Philippe, Guillermo García, and Xavier Rodet. “Tracking of Partials for Additive
Sound Synthesis Using Hidden Markov Models.” In IEEE International Conference on
Acoustics, Speech and Signal Processing. Minneapolis, MN, 1993, 225–228.
References 102
Fineberg, Joshua. “Guide to the basic concepts and techniques of spectral music.” Contem-
porary Music Review 19, 2: (2000) 81–113.
Fitz, Kelly. The Reassigned Bandwidth-Enhanced Model of Additive Synthesis. Ph.D. thesis,
University of Illinois Urbana-Champaign, Urbana, IL, 1999.
Fitz, Kelly, Lippold Haken, and Brian Holloway. “Lemur—A Tool for Timbre Manipulation.”
In Proceedings of the International Computer Music Conference. Banff, Canada, 1995, 158–161.
Fitz, Kelly, Lippold Haken, Susanne Lefvert, Corbin Champion, and Mike O’Donnell.
“Cell-utes and Flutter-Tongued Cats: Sound Morphing Using Loris and the Reassigned
Bandwidth-Enhanced Model.” The Computer Music Journal 27, 4: (2003) 44–65.
Fitz, Kelly, and Lippopld Haken. “On the Use of Time-Frequency Reassignment in Additive
Sound Modeling.” Journal of the Audio Engineering Society 50, 11: (2002) 879–892.
Garton, Brad. “Multi-language Max/MSP.”, 2007, Accessed August 27, 2008. http:
//www.cycling74.com/story/2007/3/26/145938/572.
Grey, John M. “Multidimensional perceptual scaling of musical timbres.” Journal of the

Acoustical Society of America 61, 5: (1977) 1270–1277.
Grisey, Gérard. “Did You Say Spectral?” Contemporary Music Review 19, 3: (2000) 1–39.
Harris, Fredric J. “On the Use of Windows for Harmonic Analysis with the Discrete Fourier
Transform.” Proceedings of the IEEE 66, 1: (1978) 51–83.
Harvey, Jonathan. In Quest of Spirit: Thoughts on Music. Berkeley, CA: University of

California Press, 1999.
Lagrange, Mathieu, Sylvain Marchand, Martin Raspaud, and Jean-Bernard Rault. “En-
hanced Partial Tracking Using Linear Prediction.” In Proceedings of the 6th International
Conference on Digital Audio Effects. London, UK: DAFx-03, 2003, 402–405.
Lagrange, Mathieu, Sylvain Marchand, and Jean-Bernard Rault. “Partial Tracking based
on Future Trajectories Exploration.” In Proceedings of the 116th Audio Engineering Society
Convention. AES, 2004.
. “Enhancing the Tracking of Partials for the Modeling of Polyphonic Sounds.” IEEE
Transactions on Audio, Speech, and Signal Processing 15, 5: (2007) 1625–1634.
Laroche, Jean. “Synthesis of Sinusoids via Non-Overlapping Inverse Fourier Transform.”

IEEE Transactions on Speech and Audio Processing 8, 4: (2000) 471–477.
References 103
Laroche, Jean, and Mark Dolson. “Phase-vocoder: About this phasiness business.” In
IEEE ASSP Workshop on Applications of Signal Processing to Audio and Acoustics. IEEE, 1997,
19–22.
. “New Phase-Vocoder Techniques for Pitch-Shifting, Harmonizing and Other Exotic

Effects.” In IEEE ASSP Workshop on Applications of Signal Processing to Audio and Acoustics.
IEEE, 1999, 91–94.
Lerdahl, Fred. “Timbral Hierarchies.” Contemporary Music Review 1, 2: (1987) 135–160.
Levine, Scott N. Audio Representations for Data Compression and Compressed Domain Processing.
Ph.D. thesis, Stanford University, Stanford, CA, 1998.
Lindemann, Eric. “Music Synthesis with Reconstructive Phrase Modeling.” IEEE Signal
Processing Magazine 24, 2: (2007) 80–91.
Loy, Gareth. “Composing with Computers—a Survery of Some Compositional Formalisms

and Music Programming Languages.” In Current Directions in Computer Music Research,
edited by Max V. Matthews, and John R. Pierce, Cambridge, MA: The MIT Press, 1989,
chapter 21, 291–396.
Lyon, Eric. “Spectral Tuning.” In Proceedings of the International Computer Music Conference.
Miami, FL: ICMC, 2004, 375–377.
Maher, Robert C. An Approach for the Separation of Voices in Composite Musical Signals. Ph.D.
thesis, University of Illinois, Urbana, IL, 1989.
Matthews, Max V. “The Digital Computer as a Musical Instrument.” Science 142, 3592:
(1963) 553–557.
McAulay, Robert J., and Thomas F. Quatieri. “Speech Analysis/Synthesis Based on A

Sinusoidal Representation.” IEEE Transactions on Acoustics, Speech, and Signal Processing
34, 4: (1986) 744–754.
Misra, Ananya, Perry R. Cook, and Ge Wang. “Musical Tapestry: Re-composing Natural
Sounds.” In Proceedings of the International Computer Music Conference. ICMC, 2006.
Murail, Tristan. “Spectra and Pixies.” Contemporary Music Review 1, 1: (1984) 157–170.
. Pour adoucir le cours du temps. Editions Henry Lemoine, 2005.
Nuttall, Albert H. “Some Windows with Very Good Sidelobe Behavior.” IEEE Transactions
on Acoustics, Speech, and Signal Processing 29, 1: (1981) 84–91.
References 104
Pampin, Juan. “ATS: A System for Sound Analysis Transformation and Synthesis Based
on a Sinusoidal plus Crtitical-Band Noise Model and Psychoacoustics.” In Proceedings of
the International Computer Music Conference. Miami, FL: ICMC, 2004, 402–405.
Phillips, Andrew T. “A Containter for a Set of Ranges.” Dr. Dobb’s Journal 17, 6: (1999)
75–80. http://www.ddj.com/cpp/184403660?pgno=1.
Plante, F., G. Meyer, and W. A. Ainsworth. “Improvement of Speech Spectrogram Accuracy

by the Method of Reassignment.” IEEE Transactions on Speech and Audio Processing 6, 3:
(1998) 282–286.
Puckette, Miller. “Phase-locked vocoder.” In IEEE ASSP Workshop on Applications of Signal

Processing to Audio and Acoustics. IEEE, 1995, 222–225.
. Theory and Techniques of Electronic Music. World Scientific Press, 2007. http:
//crca.ucsd.edu/~msp/techniques/v0.11/book.pdf. Online edition draft, December
30, 2006.
Roads, Curtis. Microsound. Cambridge, MA: The MIT Press, 2001.
Robel, Axel. “A New Approach to Transient Preservation in the Phase Vocoder.” In

Proceedings of the 6th International Conference on Digital Audio Effects. London, UK: DAFx-
03, 2003.
Robel, Axel, and Xavier Rodet. “Efficient Spectral Envelope Estimation and its Application
to Pitch Shifting and Envelope Preservation.” In Proceedings of the 8th International
Conference on Digital Audio Effects. DAFx-05, 2005.
Rodet, Xavier, and Philippe Depalle. “Spectral Envelopes and Inverse FFT Synthesis.” In
Proceedings of the 93rd Audio Engineering Society Convention. AES, 1992. Preprint no. 3393.
Rodet, Xavier, and Adrien Lefèvre. “The Diphone program: New features, new synthesis
methods, and experience of musical use.” In Proceedings of the International Computer
Music Conference. Thessaloniki, Grece: ICMC, 1997, 418–421.
Sandowsky, George. “My 2nd Computer was a UNIVAC I: Some Reflections on a Career
in Progress.” Connect: Information Technology at NYU http://www.nyu.edu/its/pubs/
connect/archives/index01.html.
Schoenberg, Arnold. Theory of Harmony (Harmonielehre). Translated by Roy E. Carter.

Berkeley, CA: University of California Press, 1983. First published 1922.
Schwarz, Diemo. Spectral Envelopes in Sound Analysis and Synthesis. Master’s thesis, Institut
für Informatik, Stuttgart / IRCAM, 1998.
References 105
. “Current Research in Concatenative Sound Synthesis.” In Proceedings of the

International Computer Music Conference. Barcelona, Spain: ICMC, 2005.
Serra, Xavier. A System for Sound Analysis/Transformation/Synthesis based on a Deterministic

plus Stochastic Decomposition. Ph.D. thesis, Stanford University, Stanford, CA, 1989.
Shneiderman, Ben. “Direct Manipulation: A Step Beyond Programming Languages.” IEEE

Computer 16, 8: (1983) 57–69.
Smith, Julius O. “Viewpoints on the History of Digital Synthesis.”, 1992, Accessed July 5
2007. http://ccrma.stanford.edu/~jos/kna/kna.pdf.
Smith, Julius O. “Virtual Acoustic Musical Instruments: Review and Update.” Journal of
New Music Research 33, 3: (2004) 283–304.
Smith, Julius O., and Xavier Serra. “PARSHL: An Analysis/Synthesis Program for Non-
Harmonic Sounds Based on a Sinusoidal Representation.” In Proceedings of the In-
ternational Computer Music Conference. Champaign-Urbana, IL: ICMC, 1987, 290–297.
http://ccrma.stanford.edu/~jos/parshl/parshl.pdf.
Stilson, Timothy. Efficiently-Variable Non-Oversampled Algorithms in Virtual-Analog Music

Synthesis. Ph.D. thesis, Stanford University, Stanford, CA, 2006. http://ccrma.stanford.
edu/~stilti/papers/TimStilsonPhDThesis2006.pdf.
Taube, Heinrich. “λGTK: A Portable Graphics Layer for Common Music.” In Proceedings
of the International Computer Music Conference. Barcelona, Spain: ICMC, 2005.
Terhardt, Ernst, Gerhard Stoll, and Manfred Seewann. “Pitch of complex signals according
to virtual-pitch theory: Tests, examples, and predictions.” Journal of the Acoustical Society
of America 71, 3: (1982) 671–678.
Välimäki, Vesa, and Antti Huovilainen. “Oscillator and Filter Algorithms for Virtual
Analog Synthesis.” Computer Music Journal 20, 2: (2006) 19–31.
Varèse, Edgard, and Chou Wen-chung. “The Liberation of Sound.” Perspectives of New
Music 5, 1: (1966) 11–19.
Wishart, Trevor. “Computer Sound Transformation: A personal perspective from the U.K.”,
2000, Accessed August 26, 2008. http://www.composersdesktop.com/trnsform.htm.
Wright, Matthew, Amar Chaudhary, Adrian Freed, Sami Khoury, and David Wessel. “Au-
dio Applications of the Sound Description Interchange Format Standard.” In Proceedings
of the 107th Audio Engineering Society Convention. AES, 1999a. Preprint no. 5032.
References 106
Wright, Matthew, Richard Dudas, Sami Khoury, Raymond Wang, and David Zicarelli.
“Supporting the Sound Description Interchange Format in the Max/MSP Environment.”
In Proceedings of the International Computer Music Conference. ICMC, 1999b.
Zbyszynski, Michael, Matthew Wright, and Edmund Campion. “Design and Implemen-
tation of CNMAT’s Pedagogical Software.” In Proceedings of the International Computer
Music Conference. ICMC, 2007.
Zeitlin, Vadim. “The wxWindows Cross-Platform Framework.” Dr. Dobb’s Journal 24, 3:
(2001) 106–112.
107
Appendix A. Tear of the Clouds score
Tear of the Clouds is a composition for 13 instruments and electronic sounds. It was first
performed May 6, 2008 by the Manhattan Sinfonietta, Jeffrey Milarsky, conductor and
music director. In performance, the ensemble includes a MIDI keyboard that functions in a
dual role: as a traditional keyboard that controls a variety of sampled and quasi-synthetic
pitch sounds, and as a trigger for pre-synthesized sonic complexes. The tuning of the
keyboard is dynamic and adjusts to the prevailing non-tempered harmony. Pre-synthesized
sonic events were realized using a combination of SPEAR, Common Music, and CLM.
Realtime synthesis and sample playback was done using SuperCollider.
A.1 Program Note
As the highest lake in the Adirondacks, Lake Tear of the Clouds is the Hudson River’s
source. The piece was inspired in large part by my experiences and observations living
along this waterway — the vast wide expanse at Tappan Zee, the churning gray whitecaps
of a stormy day, the vistas from the New Jersey Palisades, and an imaginary slow motion
dive from the top of the cliffs into the murky depths. Some of the sonic materials used in
both the ensemble and electronics are derived from recordings of water sounds made at
various points along the river. Other sound spectra are derived from sources such as bells,
tam-tams, and voices. In many instances these inner structure of these sounds are retuned
via a process of analysis and resynthesis.

A. Tear of the Clouds score 108
Tear of the Clouds

Duration — 10'54"
Instrumentation
Flute doubling Piccolo

Oboe
B b Clarinet doubling B b Bass Clarinet
Bassoon
Horn
Trumpet
Trombone (with F trigger)
Percussion
crotales (upper octave, written C5–C6)
vibraphone (with motor)
marimba (4 1/3 octave, A3–C7) sandpaper
congas blocks sus. cym. temple blocks
temple blocks (set of 5)
sandpaper blocks
(one block mounted, the other free and
playable with one hand) bass dr. bongos tam-tam Chinese
Chinese cymbal cym.
medium suspended cymbal
2 bongos
2 conga drums, high and low (or toms)
large tam-tam
bass drum
Computer / electronic sounds (see explanation below)

Keyboard
88-key MIDI keyboard with sustain pedal, connected to computer
Violin
Viola
Cello
Contrbass
All instruments sound the written pitch with the following standard exceptions:
Piccolo — 1 octave higher

Crotales — 2 octaves higher
Contrabass — 1 octave lower
Accidentals Strings
B quarter-tone flat exaggerated bow pressure creating a low,

Bb three-quarter-tones flat noisy, “crushed” sound
µ quarter-tone sharp resume normal bow pressure
Â three-quarter-tones sharp Note: strings should play with ordinary vibrato unless otherwise indicated
Whenever possible, woodwinds should execute

quarter tones using alternate fingerings.
Trumpet and horn should use alternate
fingerings based on the 7th and 11th partials.
Electronics
There are two types of electronic sounds indicated in the score, those than are triggered by the MIDI keyboard and those
that are triggered by the computer.
The MIDI keyboard functions much like a traditional keyboard playing a variety of sampled and quasi-synthetic pitched
sounds. A timbre description is indicated in quotation marks (e.g. “bowed pluck”). The computer operator takes care of
switching timbres at the appropriate time. The actual pitches produced, which are often microtonally inflected, are
indicated in the staves labeled “Synthesis.”
The computer operator also triggers various sonic events — chords, sustained tones, noise complexes, etc. These are also
generally indicated in the “Synthesis” staves.
Rev 3/16/08
C SCORE
Tear of the Clouds
Michael KLINGBEIL
q = 66 Bw w œ
>
& ∑ Ó Œ ‰ mœ ˙. ˙. Ó
J
ø
Flute
3 4 Í 3 F 4 π F 3 2
bw w œ
∑ Ó
&4 4 µ˙> ˙ 4œ œ œ œ ˙ œ 4 4
ø 4
Oboe
Í F π π F
>
∑ ∑ Œ ‰ œJ œ œ ˙ œ œ‰ Ó ∑ ∑
take Bass Clarinet
& J
ø
B b Clarinet
Í F
Bassoon
? ∑ ∑ ∑ ∑ ∑ ∑ ∑
Horn & ∑ ∑ ∑ ∑ ∑ ∑ ∑
3 4 3 4 3 2
B>˙ . œ‰
con sord.
&4 ∑ ∑ ∑ ∑ J ∑ ∑
4 4 4
ø 4 4
C Trumpet
f
œ. ˙ œ
con sord.
Trombone
? ∑ ∑ Ó ⋲J ∑ ∑ ∑
π F
vibraphone
˙ ˙
˙
arco
Percussion & ˙ Œ Ó Œ ∑ Ó ∑ ∑
F f
° (lock pedal down)
3 c1 . 4 3 4 mµ >œœ c1
3 2
& ⋲B b œœœ .. µ µ Bœœœ
15 c6 c7
m B œœ .. mœ B œœ
& 4 ⋲ µ œœ .. mœ Â œœ Bœœ
J 4 µœœ 4 Œ b œœ 4 4 4
p p
Synthesis
f
? Œ
œ µœ
Computer & ∑ ∑ ∑ ∑ ∑ ∑ ∑
b m œœœœ
“bells”
∑ ∑ ∑ ∑ Œ Ó ∑ ∑
“big pluck”
&
15
œ
& œ mœ . Ó ∑ ∑ ∑ ∑ ∑ ∑
Keyboard
f F
? ∑ ∑ ∑ ∑ ∑ ∑ ∑
bw w œ
o
q = 66
o o B>œ œ . ˙ œ
& ˙. w œ Œ Ó
ø ø
Violin 5
3p 4 F 3 f p 4 π 3 2
> Bw w œ
B ∑ ∑ & 4 Œ ÂœJ ˙ ˙ œ Ó
4 4 4 4
ø 4
Viola
3
Í p
≤ ≥ ≤ w w œ
B n>œ µ>œ . ˙ w ˙. ˙ œ
Ó
ø
Cello
f p F p
>o
sul G (sounding
ȯ . o o œo o
?
gliss. harm.
∑ Ó &˙ œ œ r‰ . Œ Ó ∑ ∑
as written)
œ
Contrabass
Í F f
© Copyright 2008 Michael Klingbeil

3/16/08
2 Tear of the Clouds
q = 84 œœ
mw w mw w mœ œ œ µ œ œ Bœ
œ µœ µ œ . œ ‰ ‰ Âœ ˙
⋲ œ µœ
8
& ∑
Fl.
J
2 4 p 2 6 5 ƒ p p 3
œ œ
8
w w w w œ œ Âœ µ œ œ µ œ Â œ mœ B œ œ . ⋲ ‰ nœ ˙
&4 ∑ œ
4 4 4 J 4
Ob.
p F ƒ p
5 6
Cl. & ∑ ∑ ∑ ∑ ∑
B˙ . ˙ œ œ ‰. n˘œ mœ œ œ œ
? B Bb œ J
8
Bsn. ∑ Œ R ‰ Ó ‰
π P p p f p F
3
8 con sord.
& ∑ ∑ ∑ ∑ Ó ‰ 3
œ œ œ œ
Hn.
2 4 2 5 3
p F
8
∑ ∑ ∑ ∑ Ó ‰ bœ œ œ œ
(con sord.)
Tpt. &4 4 4 4 4
p3 F
?
8
Tbn. ∑ ∑ ∑ ∑ ∑
medium Chinese cym.

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ lowj conga
sus.cym.
~~~~~~
⋲ œj . œj ‰ Ó
low bongo
œ̊
8
∑ ÷ ‰ œ ˙ ‰ œ̆ ‰
yarn
& J
w œ œ J
fl
Perc
π π P F p
4 µœœ
c2
2 2 5 3
B œœœ
8
& B Bœœ
15
eeeeeeeeeeeeeeeeeeeeeeeee µ B œœœ
4 Bµ œœœ
&4 m
Synth 4 Bb œœ 4 4
?
Comp. & ∑ ∑ ∑ ∑ ∑
& ∑ ∑ ∑ ∑ œœ ∑
15
œ
œ œ mœ
& ∑ ∑ ∑ ∑ œœ œ Ó.
p
Kbd.
f
? ∑ ∑ ∑ ∑ ∑
q = 84
Âw µw Âw µw Âœ µœ µœ
œ µœ µœ µœ
⋲ œ Bœ m œ Bœ m œ œ œ œj n Oœ Ȯ
8
& ∑
Vn.
œ J
2 4 p 2 F 6 5 f 7 p 3
µw µw µ œ Âœ µœ œ œ œ ‰ . BBOœ Oœ
pizz. arco
&4 ∑ Âw Âw j ‰ ‰ Âœ œ œ œ
Vla.
4 4 œ 4 R 4
p F f P3 f p
mO O
B & ⋲ œJ . Bœ œ Bœ œ B œ œ Bœ œ Bœ œ Bœ œ Bœ œ . Bœ . ‰ ? m œœ œœ œœ œœ ‰ m œJ œ
3 7
∑ ∑ ‰
ø
Vc.
F p P3 f p
.
& ∑ ∑ Œ œ œ œ æ ?˙ œ
ø mœ
Cb.
3/16/08
Tear of the Clouds 3
A jet
q = 60
j
whistle
˙ bœ .
13 flz.
& ⋲ Œ Ó ∑ Ó Œ ‰ µœ ˙ œ œ ˙.
æ æJ mœ
ø
Fl.
3 F 4 ƒ 2 4 3 F 4
˙ œ œ œ œ
13
∑ ∑ ∑ ∑
&4 4Œ
4 4 4
ø 4
Ob.
F
œ œ µœ
Bass Clarinet
? mœ nœ m œ œj ‰ Ó
13
& ∑ ∑ ∑ ∑ Œ
ø
m˙
Cl.
fl
ƒ
6
B ˙ œ ? µœ mœ nœ œ
m œ b œ œj ‰ Ó
13
J ‰ ∑ ∑ ∑ ∑
ø
Bsn.
fl
ƒ
6
13 senza sord.
& j ‰ ∑ ∑ Œ
m˙ . ˙.
ø 3œ F ø
œ œ œ
3˙ ø
œ
Hn.
4 2 4 4
13 senza sord.
&4˙ œ ‰ ∑ ∑ ∑
4Œ œ ˙
J
ø ø̇
Tpt.
4 4 4 4
F
?
13 senza sord.
Tbn. ∑ j ‰ Œ ∑ ∑ ∑
Sf >˙ œ œ ˙
p p
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
j
13
÷ ∑ ‰ Œ Ó ∑ ∑
bass drum
Perc
œ . ˙.
fl π̇ F
f
13
3 4 2 4 3 4
Bµ µ œœœ
c3 / c0
&
15
&4 Bœœœœ
4 4 4 4 4
ƒ
Synth
? mÂm œœœ
voice
œ
13
Comp. & ∑ ∑ ∑ ∑ ∑ ∑
∑ ∑ ∑ ∑ ∑ ∑
“big pluck”
&
15
& ∑ ∑ ∑ ∑ ∑ ∑
ƒ
Kbd.
? ∑ Œ ∑ ∑ ∑ ∑
˙.
A °
13
Oœ Oœ Oœ Oœ q = 60
Vn. & J J J ‰ Œ Ó ∑ ∑ ∑ ∑
3 f 4 2 4 3 4
O O m Oœ Oœ
⋲ µ œj . œ
&4œ œ J J J ‰ Œ Ó ∑
4Ó 4œ œ œ œ ˙.
ø
Vla.
4 4 4
f F
Oœ Oœ j
? B Bœœ œœ ‰ Œ Ó ∑ ∑ Ó ‰ Âœ ˙ .
J J J
ø
Vc.
? ≥
Cb. œ œ œ œ ˙ ˙ w ˙. ˙.
f ƒ F f
3/16/08
>
B˙ ˙ w œ bœ ˙ . ˙ œ nœ œ µœ
J J
19
& œ Œ
ø ø
Fl.
4 F p f
3 3
4 2
19
œ µ˙ œ œ ˙ œ ˙. œ œ mœ . œ œ. µœ ˙ œ œ. œ m>œ
&4 J ‰ 4 ‰
ø ø 4
ø ø
Ob.
F F f
3
? œ ˙ œ. mœ ˙ . œ >œ
19
Œ Â˙ . œ
œ . ˙
ø ø
Bs. Cl.
Ḟ f
?
19
bw ˙ w œ bœ ˙ . ˙. œ œ
Bsn.
> ƒ p
π
j j j
19
& ‰ Œ 3
j j
3
œ
Bb œ ˙ œ œ ˙ œ œ.
ø ø ø
Bœ œ œ
3
œ ˙. œ œ Bb œ œ œ. >
Hn.
4 F 2 4 F f
19
œ ˙ œ œ ˙ œ. œ ˙ Bœ . >œ
&4 œ ‰ Bœ ˙
4˙ 4 œ ‰
ø J
ø ø ø
J J
Tpt.
F F f
? r‰.
19
‰ j
m>˙
ø
bw ˙ ˙. œ µœ ˙ œ ˙
Tbn.
> ƒ
f
19
÷ Œ Ó ∑ ∑ ∑ ∑
tam-tam
Perc
œ
f
19
4 c1
2 4 c2 c3 c4 c5 c6
&
15
&4 4 4
Synth
? ‰ j Œ Œ ‰. r Ó j Œ j
bœ Bœ µœ mœ
œ œ
19
Comp. & ∑ ∑ ∑ ∑ ∑
& ∑ ∑ ∑ ∑ ∑
15
& ∑ ∑ ∑ ∑ ∑
f
Kbd.
? Œ Ó ∑ Œ Ó ‰ j Œ Œ ‰. r Ó j Œ j
bœ œ mœ mœ
> > > œ œ
b˙ ˙ œ m˙ . œ œ ˙ œ. mœ ˙ œ. Bœ œ >œ
19
˙ ˙ œ ˙. œ
& Ó
ø
Vn.
4 2F 4 π P f
&4 œ œ
mœ Âœ
˙
˙
œ œ œ œ µ Âœœ ˙˙ œœ .. B µ œœ ˙˙ œ.
œ.
µœ œ .
Bœ œ .
µœ
Bœ
œ.
œ.
>
µœ
Vla.
4œ œ 4 œ œ J J
π F π F f
? Âw
w ˙ w ‰ j
µœ ˙ . ˙ œ œ œ
Vc.
f f ƒ f P
? j
w ˙ Bw ˙. œ. œ ˙ œ. œ
> > >
Cb.
ƒ >
3/16/08
œ mœ œ œ >˙ œ µœ œ
q = 76
µœ œ œ
B>œ ˙ œ µœ œ œ œ œ. J bœ m œ Âœ Âœ j
24
& œ œ.
ø >
Fl.
3 p f 3 4 f 3
œ µœ œ œ œ. mœ œ µœ œ
⋲ b œ b œ Â œ bœ n œ µ œ œ Œ m œj µ œ m œ
7
.
24
œ œ bœ œ ˙ 4 ‰ R Ó
3
&
>
ø 8 4 œ >
Ob.
F f 4:3
? œ œ œ œ µ>œ œ œ œ ‰ Œ >
Œ µ œJ œ œ œ µ œ œ bœ
24
J Œ ∑ ∑ Œ
ø
Bs. Cl.
ƒ
3
3
? µ˙ œ. m˚œ
24
œ ˙ œ ∑ Œ Ó
œ bœ . >
ø
Bsn.
f f ƒ
Bb˚œ œ .
24
Œ Œ
3
& œ Bbœ œ . œ J œ œ
œ Bb œ œ œ 3 œ 3 œ. 4 ̇f >
Hn.
> p ḟ
b˚˙ . >œ
& œ bœ œ œ̊ œ œ
24
Œ Œ
œ >œ ˙ 4 mœ œ œ œ 8 œ.
ø J
Tpt.
4
F f
Â˚œ
3
µ˙ œ œ. œ
? ˚
. Œ Bœ œ œ
24
⋲ Âœ . œ‰ Ó Œ
œ œ
ø J
Tbn.
f p f f 3
> >x xxxx >x xxxx⋲ . ~~~~~~~~~

sus. cym.
x xxxx ® ‰ Œ
temple blocks
‰.
24
÷ Œ ∑ ∑ ‰ Œ Œ Œ Œ xxx x x ‰
6 3
œ
>
Perc
f ƒ
7
f p
24
3 3 4
r
c7 c8 c9 c10 c11 c12 c13 c14 c15
&
15
œ Bœœ
. B œœ œ
&Œ ‰ µ œ 4 b œœ Œ œ
Synth µ Â œœœ 4 8
bœœ µœ
?‰ j ⋲ j r µœ Bœ œ Œ mœ
j ‰.
Âœ
µ œ Âœ œ œ
R
24
œ 3
Comp. & ∑ ∑ ∑ ∑ ∑
& ∑ ∑ ∑ ∑ ∑
15
‰. b œœ œœ ggg b œœœœ Ó.
&Œ œ œ Œ ∑ ∑ ggg œ ∑
R ggg
ggg ƒ
Kbd.
ggg bœœ mœ œ
?‰ j gg œ œ mœ œ œ
m œ œ Âœ œ œ
Œ ∑ ∑ Œ Œ Œ j ‰. R
* œ
° *° *° 3
*°
œ m˙ œ. œ œ
Bœ œ m µ œœ œ̊ ˙
q = 76
˙ œ. B>˙ œ. œ œ œ µœ µœ œ . œ œ
œ œ
24
& Œ
ø œ J
Vn.
ƒ π 3 3F 4 ƒ 3
˚
5 r
5:3 œ
mœ œ
B œ µœ œ mœ Âœ ˙ . ⋲ ⋲ BBœœ .. œœ œœ ..
& 4 Ó. mœ œ j œ mœ µœ
mœ œ œ
4 œ Œ 3
ø
Vla.
8 µœ œ
ƒ π f
>˙ . œ
? µœ œ. œ œ œ. Œ Œ Bœ œ Bœ œ
j‰ &
œ µœ œ œ œ J œ
Vc.
f p f ƒ
3
3
o œo ? Œ
(sounding
as written)
? Ó Œ &œ Œ
sul G
j j
3
œ œ œ œ œ œ œ œ œ œ
ø œ œ
Cb.
Ï ƒ
3/16/08
B q = 69
b˙ ˙ œ
œ œ. bœ nœ mœ œ
29
Âœ œ m œ
Fl. & Œ ∑
Ï
3
2 3 4
œ œ m œ b˙ ˙ œ
œ. œ µœ œ œ
29
& œ Œ ∑
Ob.
4 4 4
Ï
˙ œ
7
? µœ nœ œœœ
œ œ B˘œ m˘œ œ̆ m˘œ ˘ mœ œ mœ µ œ bœ n œ bœ b œ n œ µ œ œ œ bœ bœ n œ µœ œ
29 7
mœ
7 7
œ. œ bœ bœ bœ n œ b œ b œ n œ
œœœ œ bœ
3
œ bœ œ
Bs. Cl.
> >
9
Ï 5 π
5
? mœ œ mœ
29
∑ Œ bœ j ‰
fl fl fl fl flœ >˙ ˙ œ
ø
Bsn.
f Ï
r
& œ ‰.
29
‰ Bb ˙ ˙ œ Œ ∑
3
œ
>
Hn.
ƒ 2 3 4
bœ . n˙ ˙ œ
& ⋲ œJ . œ
29
bœ Œ ∑
Tpt.
4 4 4
ƒ
>œ
3
˙
? Œ
29
‰ ∑
bœ œ ˙.
Tbn.
f p >
3
f p
~~~~~~~~~~~~ j >œChinese cym.
œæ œæ œæ
29
hi conga
÷ œ Œ ∑ Ó
3 low conga
œ œœ
J >
Perc
f S π
œ
c18 c1
µ mµ œœœœ
29 c16 c17
2 3 4
&
15
mœ Œ
& Œ ‰
3
œ t Œ bœ
œ 4 4 4
j p
Synth
? ⋲ µœ . t Œ œ ‰. r
bœ ‰. R Bœ œ œ
29
Comp. & ∑ ∑ ∑ ∑
mœ œ Œ
7
& Ó œœ ∑ ∑ ∑
15 “bowed pluck”
mœ œ
bœ mœ œ
& ⋲ j m œ œ bœ ∑ ∑ ∑
œ.
p
Kbd.
5
f
? ∑ ∑ t Œ ‰. œ ‰. r
R œ œ œ
bœ
* B q = 69 °
√
œ bœ µ˙ µ˙ œ.
œ. œ mœ µœ mœ æ J
29
& ⋲ m ⋲ Œ ∑
Vn.
J 6
ƒ 2 3 4
bœ . œ µ˙ µ˙ œ.
pizz.
˝
Vla. & ⋲ J æ 4 J B⋲ œ
4 ∑
4
Ï ƒ
˙ œ. ?⋲ ˝
pizz.
& ⋲ µ œj . j
æ̇ ∑
3
œ Âœ J mœ
Vc.
ƒ ƒ
> > > > >
? ∑ ∑ œ œ mœ µœ mœ µœ œ
3
> Â>œ >œ œ ˙

Cb.
f 5 p
3/16/08
33
Fl. & ∑ ∑ ∑
4
33
Ob. &4 ∑ ∑ ∑
œ œ œ µœ œ mœ bœ nœ
7
? œ œ œ mœ
7
b œ œ b œ b œ b œ bœ b œ b œ œ µ œ œ b œ bœ n œ µ œ b œ b œ . b œ b œ bœ n œ b œ bœ n œ µ œ œ œ b œ bœ œ œ
33 6 5 3
bœ œ bœ b œ b œ nœ bœ bœ nœ
7:6 3
Bs. Cl.
bœ
6
œ.
5
? œ œ.. œ œ œ œ
33 5
Ó
3
bœ µœ bœ µœ œ
5
J mœ . mœ mœ œ bœ .
5
Bsn.
mœ œ
π
33
Hn. & ∑ ∑ ∑
4
33
Tpt. &4 ∑ ∑ ∑
?
33
Œ ∑ ∑
˙.
ø
Tbn.
33
B.D. beater
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
tam-tam l.v.
÷ ∑ Œ .
ø̇ w
Perc
p
33
4
&
15
t
4
mœ œ . .
Synth
µœ œ bœ . Â œ œ µ œ bœ œ
t µœ œ µ œ µ œ mœ œ µ œ b œ œ µ œ µ œ b œ µœ µ œ m œ Œ
œ Bœ œ µœ œ j
3
Âœ µœ µœ Bœ
3 3
7 6 6 bœ œ
33
Comp. & ∑ ∑ ∑
& ∑ ∑ ∑
15
& ∑ ∑ ∑
Kbd.
œ bœ . mœ œ . . nœ œ mœ
t œ
mœ bœ œ m œ œ mœ œ œ m œ b œ œ m œ m œ b œ mœ œ m œ Œ
œ œ œ mœ œ m œ œ b œ œj
3
œ mœ 3 3
*°
7 6 6
33
Vn. & ∑ ∑ ∑
4
arco
B Ó Œ ⋲ j Œ ∑
ø ø
4 µœ . œ œ ˙
Vla.
F
Ÿ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
sul pont.
? j
arco
‰ j j ‰ Œ ‰ œ (µœ ) ˙
bœ ˙ . ˙.
ø œ
ø
Vc.
F π
? Œ ‰ mœ œ œ œ œ œ Œ Ó
J
ø ø ø
Cb.
˙.
F
3/16/08
36
Fl. & ∑ ∑ ∑
36
Ob. & ∑ ∑ ∑
œ Bœ œ œ
3
? µœ œ µœ œ µœ œ b œ œ œ b œ bœ b œ b œ n œ µ œ œ œ œ
36 5 5
µœ
7
bœ œ b œ bœ b œ n œ b œ n œ bœ
3
bœ bœ
3
bœ
Bs. Cl.
p
7 poco cresc…
œµœ œ µœ œœ œ œ œ mœ œ
5 7 3
? œœ œœ œ mœ µœ
36 5
œ b œ m œ mœ bœ n œ m œ nœ bœ mœ œ b œ mœ n œ m œ b œ b œ m œ œ m œ b œ n œ œ n œ m œ b œ n œ œ bœ bœ œ
6
Bsn.
mœ mœ œ m œ œ bœ œ m œ
p 6:5
7
5
36
Hn. & ∑ ∑ ∑
36
Tpt. & ∑ ∑ ∑
?
36
Tbn. ∑ ∑ ∑
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
˙. œ. œ̆ œ œ b œ œ bœ
medium marimba
?mallets œ œ ˙
36
Perc ÷ Ó Œ ‰ J bœ
F p F 6
ø
36
&
15
t bw w
µ œ œ œ Bœ .
Synth
œ
t ⋲ œ mœ bœ µœ ? ‰.. rK œ œ œ Bœ œ µ œ œ œ Bœ
3 5
‰ œ µœ µœ bœ µœ µ œ µ œ œ bœ µ œ µ œ
7
Bœ µœ mœ nœ œ bœ
µ µœ œ . mœ µœ œ œ œ
5
5
3 5
5
36
Comp. & ∑ ∑ ∑
& ∑ ∑ ∑
15
& ∑ ∑ ∑
Kbd.
œ mœ œ œ œ .
t ⋲ œ mœ bœ mœ ? ‰.. rK œ œ œ œ œ mœ œ œ mœ m œ œ bœ œ œ œ
3 5
‰ œ mœ bœ mœ mœ œ nœ œ bœ mœ
7
œ m mœ œ . mœ œ œ œ œ
5
*°
5
3 5
5
36
Vn. & ∑ ∑ ∑
Vla. B ∑ ∑ ∑
~~~~~~~~ Ÿ~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Ÿ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

ord. sul pont.
œ µœ µœ m ˙ .(œ )
?
œ
5
Bœ œ . Œ µ ˙ (œ ) œ bœ œ œ µœ
œ mœ
µœ . bœ bœ
Vc.
P p 3
F
œ. œ µœ µœ
? Ó ⋲ J œ bœ µ œ
µœ œ œ mœ mœ nœ j œ bœ m œ µœ
5
œ œ mœ œ .
7
5
Bœ œ œ. Bœ . œ œ.
Cb.
P P
3
3/16/08
39
Fl. & ∑ ∑ ∑
39
Ob. & ∑ ∑ ∑
bœ œ œ œ
? œ œ œ œ œ Bœ œ œ œ œ bœ œ bœ œ bœ . œ. bœ
œ bœ bœ µœ œ œ µœ nœ œ bœ œ
39
Bs. Cl. œ œ bœ
3
P 6 f p subito
œ
? œ œ nœ mœ œ nœ œ œ œ œœœ m œ b œ m œ m œ m œ bœ œ Â œ
39 5 6
µœ œ n œ œ bœ m œ n œ m œ œ bœ m œ mœ n œ m œ œ b œ m œ m œ
6
œ œ bœ mœ nœ œ bœ
6
œ
5
œ œ. bœ mœ œ œ œ mœ
Bsn.
f p subito
6
3 cresc… 6
6
39
Hn. & ∑ ∑ ∑
39
Tpt. & ∑ ∑ ∑
?
39
Tbn. ∑ ∑ ∑
œ bœ #œ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
? œ #œ œ bœ œ œ œ œ œ.
œ bœ ‰. r
39
‰ J ‰ J bœ Œ œœ ˙˙ ..
ø
Perc
5 3
p π f p
39 c2 / 12
&
15
t w w
w
π
Synth
? ‰. œ œ
R
œ µœ . œ. œ œ t ? mœ
5
bœ µœ bœ
µœ œ .
39
ṗ
Comp. & ∑ ∑ ∑
& ∑ ∑ ∑
15
& ∑ ∑ ∑
Kbd.
? ‰. œ œ
R
œ mœ . œ. œ œ t ? mœ
5
bœ œ bœ
mœ œ ˙.
*° *°
39
Vn. & ∑ ∑ ∑
n œ µ œ µ œ µ œ bœ œ Ÿ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
3
œ bœ ⋲
sul pont.
Bœ œ
5
B ‰. r œ œ Bœ . œ Bœ œ
µœ œ œ Âœ œ Bœ µ w (b œ )
ø
Vla.
3
Í 5
5 f
~~~~~~~~~~~~~~~~~~
ord.
? ˙
sul pont.
Ó ∑ Œ Â˙ .
ø ø
Vc.
œ
? bœ µœ Âœ œ œ
⋲ µ œj .
pizz.
J ‰ Œ ∑ Œ Œ
œ
Cb.
p F
3/16/08
42
Âœ ˙. œ œ ˙ ˙
& ∑ ‰ J
ø
Fl.
3 p 2 3
b˙ . œ ˙ œ œ
42
& ∑ Œ
ø 4 4 4
Ob.
P
bœ bœ Bœ œ . œ
bœ bœbœ
? bœ bœ bœ bœ nœ bœ .
6
œ œ œ b œ œ b œ œ œ b œ bœ bœ œ œ mœ œ œ bœ bœ œ mœ œ bœ bœ b œ
42 3
bœ œ.
5 5 5
Bs. Cl. œ bœ œ bœ .
3 3
œ µœ .
3
œ
7
œ bœ œ mœ œ Âœ bœ b œ œ bœ bœ
? bœ bœ nœ œ mœ
42
mœ mœ mœ bœ œ œ œ Œ bœ œ
3
‰
œ œ
Bsn.
P
3
5
42
& ∑ Ó
ø
Bb ˙ 3 œ œ œ œ 2 ˙
Hn.
3
42
p
& ∑ ∑ ∑ ∑
Tpt.
4 4 4
?
42 con sord.
∑ ∑ Â˙ . ˙
ø
Tbn.
p
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
? ‰. r
42
∑ ∑ ÷
tam-tam
Perc
œ> ˙ ˙
p π P
42
3 2 3
&
15
t
Synth w w 4 ˙. 4 ˙ 4
? Bœ bœ bœ Bœ
b œ Bœ t µ œ œ œ Bbœ
5 6
œ bœ ˙ n œ bœ m œ
œœ˙
F
5
3
42
Comp. & ∑ ∑ ∑ ∑
& ∑ ∑ ∑ ∑
15
& ∑ ∑ ∑ ∑
Kbd.
? œ bœ bœ œ t m œ œ œ bœ
5 6
œ bœ ˙ n œ bœ m œ bœ œ Ó
œœ˙
F *
5
3
42
w œ ˙ œ œ
& ∑ Bw œ ˙ œ
ø ø
Vn.
3 p 2 π 3
µ˙ ˙ œ œ ˙. ˙
ord.
B Ó µ˙ ˙ œ œ 4 ˙. &4
ø ø
Vla.
4
p π
? ∑ ∑ ∑
˙.
ø
Vc.
œ
F
r mœ . bœ .
? Œ œ bœ ‰ µœ Œ ‰. bœ Œ
R Œ œ µœ ⋲ J Ó ∑
Cb.
J 3 J
5
f
3/16/08
œ µœ
œ.
46
& Ó ∑ ∑ Œ ‰ œ œ
ø J
Fl.
3 4 p f
˙
46
Œ
&4 m˙ . ˙. 4Œ m œ- Â œ ˙
ø >
Ob.
Í F P
œ ˙
? œ bœ œœœ take B b Clarinet
46
œœœ œ bœ Œ ∑ ∑
ø
Bs. Cl.
œ
ƒ
F
6
mœ . œ œ mœ œ œ.
? œ Âœ œ B mœ œ œ œ J ?
46
Œ œ mœ mœ œ œ œ œ œ mœ mœ Â œ ⋲ Ó
5 5
mœ . bœ œ.
Bsn.
p 3 6 π
? ‰. r
46
& Œ ∑ ∑ Ó Œ œ
3 ˙
ø
Hn.
4 Í
46
& 4 m˙ j ‰ ∑ ∑
ø œ œ ˙
øœ
Tpt.
4
p
? ‰. µœ
46
˙ Œ Ó ‰ j Œ R
ø ø
bœ ˙ œ ˙
ø
Tbn.
p Í
46 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
÷ ˙. ∑ ∑ ∑
ø
Perc
46
3 c3 / 20 4
&
15
t
4 4
p
Synth
t ⋲ j Ó. ‰. r
µœ . bœ
46
Comp. & ∑ ∑ ∑ ∑
∑ ∑ ∑ ∑
“organ pluck”
&
15
& ∑ ∑ ∑ ∑
p
Kbd.
t ⋲ j
mœ . ˙ ˙. ˙. ˙. œ. bœ
µ˙ ˙.
œ
µ µ Ȯ .. Ȯ
46
Vn. & µ BȮ

3 p 4
&4
œ B m m Ȯ Ȯ Oœ µ µ Ȯ .. Ȯ
bb Ȯ
Vla.
4
p
Ȯ
? ÂÂ Ȯ .. Ȯ
ord.
Œ
B˙ ˙.
Vc.
p
µœ . o
sul D (sounding o o
? ⋲ J r
sul A
Ó ∑ Œ ‰.
as written)
Cb. & ˙. ˙ œ
F p
3/16/08
& æw æ̇. æ̇ ⋲ œr ‰
50
Ó Ó
tongue ram
Fl.
p 3 4 3 >O 4
ƒ
Bœ
50
Ob. &w 4œ µ˙ 4 ˙ œ œ 4 ˙ 4
f F f
œ
œbœ œ bœÂœ
B b Clarinet
6
? ‰.
50
∑ ∑ Ó œœ
6:5
& œ µœ µœ œ µœ œ œ
µœ mœ œ
Bs. Cl.
p 5
p
mœ
œ œ œ œ œ µœ
? ⋲ m œ Â œ µœ m œ Â œ µ œ œ Â œ µ œ œ œ œ œ µ œ mœ œ œ . œ œ œ mœ œ
6
‰ œ µ œ œ œ µ œ œ œ Âœ R ‰ . ‰ mœ œ œ µœ
6
R ‰.
50 3
mœ Œ
Bsn.
œœ œ
p p p
5 3
6 6 6
? ˙. mw ⋲ œ. ˙
50
Œ
Hn.
œ ˙ J
3 4 Í 3 Í 4
50
∑ ∑ ∑
& 4 ⋲ œj .
ø
Tpt.
4 4 ˙ 4
Í
? ˙.
50
bœ ˙. j
Tbn. ∑ ⋲ Âœ . ˙
Í
50
Perc ÷ ∑ ∑ ∑ ∑
50 c4 / 33
3 4 c5 / 40
3 c6 / 44
4
&
15
t
Synth 4 4 4 4
t µœ ⋲ œJ .
50
Comp. & ∑ ∑ ∑ ∑
& ∑ ∑ ∑ ∑
15
& ∑ ∑ ∑ ∑
Kbd.
t
w ˙. mw œ œ. ˙
j >
ÂÂ Ȯ Oœ Ȯ µ µ Oœ Ȯ .. Oœ ‰ ⋲ BBOœ .. Oœ Oœ
50
Vn. & Oœ J
F p 3 4 F π 3 f 4
>
B Oœ Ȯ Oœ Ȯ Oœ j
Oœ ‰ m Oœ .. Oœ Oœ
Vla.
4 4 µBȮ .. 4⋲ J 4
F p F π f
>.
? Oœ ‰ bb Ȯ Oœ Ȯ .. Âw ⋲ bb Oœ . Ȯ
Vc.
J J
F p Í f p
o j
? ⋲ m œJ . œ Œ
pizz.
∑ ∑ ‰.
3
Cb. &w R œ
ƒ F
3/16/08
œ ˙
µœ mœ J Âœ œ . œ
54 3
Fl. & Œ m˙ . J nœ œ J œ œ
4 p 2F 3 5 f 4 2
. œ œ ˙ œ Bœ
œ Âœ µœ œ œ. œ
7
&4 œ mœ J µ œ œ Bœ
54
Ob.
4 8 4 mœ nœ œ m˙ . 4
ƒ
5 5
bœ œ b œ Â œ b œ œ . n œ bœ Ÿ~~~~~~~~~~~~~~ 3
r ‰.
54
∑ ∑ Œ ‰
7
& œBb œ j
bœ bœ Âœ
µœ œ œ œ
Cl.
œ b œ(nœ ) œ
f p f
5
œ µœ œ mœ œ œ œ œ
œ µœ œ œ œ
6
? œ œ œ Bœ µ œµ œ œ J ˙ œ
54 3
Ó ‰ ‰ J J
Bsn.
œ.
F f p F
3 3
3
? œ
54
Œ Ó ∑ ∑ ∑
ø
Hn.
4 2 5 4 2
54
∑ ∑ ∑
&4 4 8 4 Œ bœ . ˙ 4
>œ >
Tpt.
f
?
54
˙ Ó ∑ ∑ ∑
ø
Tbn.
54
Perc ÷ ∑ ∑ ∑ ∑
54
4 c7 / 48 2 5 4 c8 / 52
2
&
15
t b µ œœ
& 4 µ Bœœœœ Œ Ó
Synth 4 4 8 bœ 4
t ?œ mœ
Œ
54
Comp. & ∑ ∑ ∑ ∑
& ∑ ∑ ∑ ∑
15
& ∑ ∑ ∑ ∑
Kbd.
t œ ?˙ . ˙ œ œ. mw
Oœ ÂÂ Ȯ .. Oœ Oœ m O Oœ ..
>
Oœ n Oœ ⋲ m wO
54
& mœ
ø
Vn.
4P 2 f 5 4 ƒ 2
>
Oœ Oœ b Oœ Oœ .. Oœ µ µ Oœ
BBȮ .. Oœ sul pont.
B ⋲ ‰.
4 Œ Œ r
4
ø 4 8 mœ œ 4
Vla.
P f Í
? Oœ O. O O Âµ Oœ Oœ .. Oœ
sul D
‰ Œ µ˙ .
ø J
Vc.
F F
O. O O K ˙o .
arco (sounding
arco sul D sul G as written)
? ⋲ œj . œ
pizz.
Cb.
J ‰ ∑ &
p p f F
3/16/08
b>œ µ>œ
q = 80
58
˙
& ∑ µœ œ œ œ œ œ
œ œ œ
> >
Fl.
2 5 3 P 3 2 3
œ œ mœ œ µœ œ bœ µœ mœ mœ m œ mœ œ œ ˙ œ œ µœ
mœ œ œ µ œj j J
58
Ob. &4œ 3 8 8 œ œ. 4 4 8
f 5
7
ƒ
œ œ bœ œ œ
mœ µœ µœ mœ µœ Œ.
58
Cl. & œ b œ bœ 6
∑ ∑ ∑
6
p
?
58
∑ ∑ ∑ ∑
Bsn.
˙
p
?
58
Hn. ∑ ∑ ∑ ∑ ∑
2 5 3 3 2 3
œ Bœ
œ mœ œ
58
j œ œ
Tpt. &4
˙ 8 mœ œ. 8 œ 4 œ 4 8
F P f
3
?
58
Tbn. ∑ ∑ ∑ ∑ ∑
58
Perc ÷ ∑ ∑ ∑ ∑ ∑
2 5 3 mœœœ
c9 / 60
3 j 2
c10 / 68
3
bœœ
58
& m Bµ œœœ ‰
µ µB œœœœ
15
m µ µ œœœœ Œ. m œœ ‰ B mœœ
Synth
&4 8 Â œœ 8 4 Ó 4 8
? Bœ . ‰ bœj
58
Comp. & ∑ ∑ ∑ ∑ ∑
& ∑ ∑ ∑ ∑ ∑
15
& ∑ ∑ ∑ ∑ ∑
Kbd.
? ˙ œ œ. œ. ˙ œ bœ ˙
Âœ Bœ .
µœ œ mœ
q = 80
Ȯ
œ µœ µœ mœ
58
& ‰ œ œ. ∑ ∑
J
Vn.
fÍ
3
2 5 f 5
3 3 2 3
œ œ.
œ µœ Âœ Œ .
ord.
B bœ ∑ ∑
3
Vla.
4 œ. œ œ œ 8 &8 4 4 8
cresc… f fÍ
gliss. harm.
? Œ ? mO . ∑ ∑
sul D
Vc.
˙ & O.
f f p
ȯ o
gliss. harm.
? ‰ bœj
pizz.
Œ ∑ Ó ∑
sul G
&
œ.
Cb.
f f
3/16/08
œ̆ œ. œ
œ bœ Bœ œ µ>œ >œ b œ µ œ
œ. µœ mœ œ œ nœ
63
Fl. & ‰ Œ ∑
f f
3
3 2 4 7 3 2
3
3
œ. œ mœ mœ œ µœ œ ˙ Ÿ~~~~~~~~~~~~~~
˙ (œ)
63
Ó ∑
Ob. &8 4 4 8 4Œ 4
p
3
µ œ Â œ œ œ bœ œ Ÿ~~~~~~~~~~~~~
‰ œœ œ
63
& ∑ Œ œ bœ œ mœ ∑ Œ b ˙ ( b œ)
Cl.
œ Bœ ˙
ƒ p f
P
7
3
˙ ˙
?
63
Bsn. ∑ ∑ ∑ ∑
f p
? ˙ ˙
63
Hn. ∑ ∑ & ∑ Œ Bb˙

3 2 4 f p 7 3 f 2
œ. œ œœ Bb œ . œ n˘œ ˚ œ̆
mœ
63
mœ . b˙
&8 j œ bœ œ œ 4 œ.
5
4 . 4 œ œ 8 4
>œ œ œ
Tpt.
F 3 ƒ 5
? ⋲ µœ . j
63
mœ . œ Ó
3
Tbn.
J mœ µ˙ ˙ œ. ˙ œ
f P f Chinese cym.
>œ j >œ >œ >œ ˙~~~~~~~~~~~~~~~

Chinese cym.
>
with yarn
œæ œæ
63 medium bongos congas
stick end of mallet
÷ ∑ ‰ Ó ‰ ‰ Ó Œ
æ
œ
yarn
Perc
æ œ
J
p
Í f p F
3
3
œ /. 72
œœœ ...
3 bc11 2 j 4 7 œ 3 2
63 c12 / 76 c13 / 80 c14 / 84 c15 / 88
& Bµ œœ .. m Bœœœ µ µ Bœœœ

15
œœ m œ
3
Bµ œœ .. bœ m µb bœœœœ n n µ œœœ
µ bœ
mœ n œœ µ œœ 4 Bµ œœœœ œœ bmbµ œœœœœ
&8 4 Â œœ 8 4Œ 4
Synth
mœ j mœ œ
? œ œt
œ. œ
Bœ œ œ mœ mœ
63
∑ ∑ ∑ ∑ ∑
3
Comp. &
& ∑ ∑ ∑ ∑ ∑
15
& ∑ ∑ ∑ ∑ ∑
f
Kbd.
? t œ
3
j
œ. œ œ mw œ. ˙ ˙.
œ. Ÿ~~~~~~~~~~~~~~
œ bœ mœ nœ µœ œ ˙ µœ . œ
Bœ ˙ m ˙ (Bœ )
63
Vn. & Ó Œ Œ
f ƒ 7 fÍ Í
3
3 2 4 3 2
Bœ . mœ bœ œ nœ œ œ ˙ ˙æ̇
3
Ó ∑
Vla. &8 4 4 8 4Œ 4
f ƒ Í
µ µOœ .. mœ nœ œ œ œ µœ ˙
? Ó ∑ Œ
µ˙
Vc.
f f Í
3
j ˙ ˙
?
arco
bœ ∑ ∑
3
Cb.
µœ . œ bœ
p F f p
3/16/08
Ÿ~~~~~~~~~ >œ . ˙
K œ mœ œ mœ
m œ (œ) >œ >œ
> > > >œ
q = 76
r B>œ ⋲ mœ œ µœ µœ . µœ . mœ œ
& ⋲ . µœ
68
Fl. ∑
2 p ƒ 4 3ƒ 4 f5
5
9
~~~~~~~~~~~~~~~~~~~~~~~ m˘œ œ µœ ˙
&4˙ . œ Âœ µ œ
68
Ob.
4 R‰ Œ Œ ⋲ mœ
4 Œ
4 ∑
8
ƒ ƒ p
œ µœ
7
b œ bœ Bœ Bœ ˙.
6:5
œœ Bœ m œ
& ⋲œ
68
µœ
3
œ ∑
Âœ œ mœ
Cl.
œœ œ œ. œ mœ f
>
6
fÍ
6
œ œ. m œm œ œ
6
? œ œ œ mœ
68 5
∑ mœ ˙. Œ Ó
mœ œ œ bœ nœ œ œ
Bsn.
f 6
j 3 œ B>œ ˙.
68
& j j œ œ œ Œ Bb˙ œ Bœ
3
3
2œ œ
Hn.
œ 4 œ œ ƒ 3 Ï 4 p 9
>œ œ ˙. œ œ >œ . ˙
& 4 œ. m>œ
68
∑
Tpt.
4 4 4 8
F ƒ p Ï
? œ œ Bœ œ œ Bœ œ bœ
68
∑ J œ ˙ Œ µœ µœ . ˙
J
ø
Tbn.
P f f p
3 3
68 ~~~~~~~~~~~~~~~~~~~~~~~
˙ P
÷ ∑ ∑ Œ Ó
tam-tam
Perc
œ
f F
c19 / 104 c20 / 108
œ j 9
bœœœ µ œœ
68
2 4 c16œœ/ 92 c17 / 96 c18 / 100
3 4 c21 / 112 c22 / 116
& Bœ
µ m œœœ mœ
15
mœ
B n Bœœœ œ
3
Bbµ œœœ n œœ
Âm µ œŒœœ bB n œœœœ
B
3
Bœ ‰ B b µ œœœ ‰ b Ân œœœœœ b œœ Œ Œ ‰ m Âœœ

Synth
&4 4 œ 4 4 µœœ 8
µ œœ µœ œ
t ? ‰ œ ‰ Bn œœ mœ œ ˙. œ bœ
œ
œ 3
3
68
Comp. & ∑ ∑ ∑ ∑
∑ ∑ ∑ ∑
“big pluck”
&
15
& ∑ ∑ ∑ m œœ Œ Œ ‰ m œœ
Kbd. J
F
t Œ ? j
3
mœ œ ˙. ˙. œ bœ
j œ œ
3
œ œ œ œ
° *°
√ b>œ
Bb
œ.
>œ
C
~~~~~~~~~~~~~~~~~~~~~~~~ Ÿ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ q = 76
˙ b w (µ œ ) bœ . œ. œ
œ œ
68
Vn. & Ó œ bœ
2 ƒ 4 Í ƒ 3 F 4 f 5
9
>œ œ w≤
(with tpt.)
Bœ œ m œ ∑
Vla. &4
æ æ æ 4
Í ƒ
4 4 w
> π
8
ƒ f
œ̋ Ÿ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>
µ w (œ)
? œ mœ B œ œ.
pizz.
‰ œ ‰ Ó
3
µ ˙æ
Vc. Bœ
ƒ 3 Ï F f π
?
pizz.
œ ˝j
∑ ‰ ‰ ‰ ∑ ∑
3
Cb. œ bœ œ
f 3
Ï
3/16/08
>œ b œ >œ .
m>œ n œ œ
Â œ œ µœ œ . m œ
C
bœ µœ œ ˙.
‰ œr œ µœ µœ œ . œ. œ bœœ œ µœ j œ.œ
72
Œ ‰
7
œ œ
3
&
>œ œ
Fl.
5
P p
5
9 4 ƒ F
7 3
5
72
&8 ∑ ∑ ∑ ∑
Ob.
4
72
Cl. & ∑ ∑ ∑ ∑
?
72
Bsn. ∑ ∑ ∑ ∑
bœ œ . j
72
Hn. & ˙ œ œ Œ Bb ˙ ˙. œ. ⋲ ∑
9 4 π p ƒ
Bb>œ Bb œ bœ
5
mœ ˙ mœ bœ mœ
72
&8 ∑ Ó Œ ‰ Œ
Tpt.
4 œ m˙ .
f
µœ ˙ fÍ
? ˙ œ œ œ. œ ˙.
72
Tbn. Œ Œ Bœ œ Œ
π f P f π
œ #œ œ
7
œœ
bœ nœ œ Œ
medium vibraphone
œbœ bœ ‰ J
72
÷ ∑ Œ Ó ∑ œ bœ Œ
mallets (motor off)
&
#œ
Perc
f 7
F
° *° *°
7215
9 4
&
/ 120 / 124 / 128
j
µ œœ Bœ
b m µ œœœ bœœ ‰ m m œœœœ
Synth
&8 4 Œ µœ
Ó
µ Â µ œœœ
Œ Œ
œ
? ˙ œ Bµµ œœœ m œœ ‰ µ œœ
j
œ. œ. Bœ
mœ Bœ
72
Comp. & ∑ ∑ ∑ ∑
& ∑ ∑ ∑ ∑
15
& ∑ ∑ ∑ ∑
Kbd.
b˙ .
? ˙ œ t ˙. ∑ ? œ
œ. œ. mœ n˙ . ˙. œ.
*° *° *°
œ
œ œµ œ . œœ
⋲ >œ
C
œ œ œ µ œ µ œ µ œ Bœ m˙ . œœ œ œ
R œ mœ bœ
72 6
& bœ . Bœ µ œ Œ œ
Bœ ˙ ˙.
Vn.
f 4 f
5
p p
7 5
9 5
œ œ œ µœ >O O Ȯ bœ œ
&8 Œ ‰ J J Bœ m œ b œ B Bœ µ œ Bœ œ œ œ Œ ⋲ µœ bœ
œ
4 Â˙ . œ œ
Vla.
f 3
p f 6
Bœ œ œ. œ ? µ œ µ œ Bœ œ
7:5
B œ
5
Œ ‰ J µœ œ Bœ ∑
œ œ Âœ œ j Bœ mœ µœ Âœ µœ µœ mœ Bœ œ nœ
3
œ mœ
Vc.
f 5
ƒ
j Bœ˝
pizz.
? ∑ Ó ‰ Âœ œ œ j ⋲ Œ Ó
Cb.
µ˙ œ.
f ƒ
3/16/08
œ Bœ nœ œ œ mœ œ œ. œ œ bœ µœ œ . œ
œ bœ œ
œ
µœ nœ œ nœ œ Âœ bœ œ µœ m œ bœ œ mœ
J Bœ n œ œ œ bœ
76
& Œ ‰ œœ
mœ œ Bœ .
>
Fl.
f 5p f
6
5 6 3 3 4
œ b œ µ œ m œ µœ .
76
Ob. & Ó Œ œ 4 œ ˙ 4 œ Âœ . œ Âœ . œ œ 4
p f
5
5
œ œ bœ Bœ Bœ
œ œ µ˙ œ bœ œ œ œb œ J
76 5
& ‰ mœ œ bœ
5
mœ œ . ˙ œ mœ œ Âœ mœ œ œ . œm œ
Cl.
>
f
7
?
76
Bsn. ∑ ∑ ∑
76
Hn. & ∑ ∑ ∑
3 5 4
76
& ∑ ∑ ∑
Tpt.
4 4 4
?
76
Tbn. ∑ ∑ ∑
j œ #œ œ nœ
#œ j œ #œ œ œ J œ œ
œ n œ nœ
76 3
& Ó bœ ‰ bœ ‰
5
œ œ.
œ œœb œ œ
œ #œ bœœ .. œ bœ œ .
#œ
Perc
œ
f p f 5
*° pedal phrases…
7615
3 mœœ / 144
5 4
& œœ j
/ 132 / 136 / 140
µœ µ µB œœœ
3
œ
& Ó m µ m œœœœœ Bb œ
BB œœœœ œ
b bn œœœ
µ œ Ó
4 B œœ 4 4
>œ
Synth
? Ó Bœ m n œœ Âœ
œ J
3
76
Comp. & ∑ ∑ ∑
& ∑ œ mœ ∑
15
b œ bœ
& ∑ œ Ó ∑
Kbd.
œœ
œ f
7
? ˙ œ mœ J ∑ ∑
3
* °
µ˙ .
⋲ mœ Â œ œ œ œ œ j mœ œ mœ bœ œ œ µ œ n œ µœ µ œ
76 7 3 5:3
& j œ œ œ œ œ
3 3
Vn.
œ µœ œ 3
5:3
œ
3 5 4
œ œ Âœ . bœ
B œ Bœ 4 œ & œ œ µœ
Vla.
m˙ mœ 4 œ µœ µœ œ 4
5:6
? ˙ Œ ‰. r
5
œ œ. µœ œ µœ µœ œ
b˙ œ bœ œ œ R
œ
Vc.
f F
3
5
Cb.
? ∑ ∑ ∑
3/16/08
œ œœ œ
µœ mœ µœ œ mœ µœ mœ œ
µœ . œ œ mœ
œb œ bœ . µ œ œ bœ µœ µœ µœ µœ
bœ œ bœ b œ b œ b œ œ œ m œ
œ . œ bœ mœ
mœ œ mœ nœ œ nœ bœ
œ mœ œ ‰.
79
Fl. & 5
œ 7:5
5
4 5
7:5 3 6 6 6
œ mœ œ bœ Âœ mœ mœ œ œ mœ bœ œ
7:6
œ œ µ œ m œ µ œ bœ mœ n œ m œJ
œ. mœ œ œ mœ mœ œ n œ mœ œ mœ
3
j
R ⋲ œ œ Bœj . œ œm œ œ mœ nœ
79
Ob. &4 ⋲ œ mœ .
5
mœ œ œ bœ mœ œ nœ mœ nœ œ. bœ Bœ Bœ mœ œ œ bœ œ mœ . œ œ mœ nœ ⋲ œ bœ . m œ . n œ .
œ œ œ œ bœ mœ œ bœ œ œ œ œ .mœ .œ .œ .œ .
79
Cl. & mœ 6:5 6:5

7 7
r m œ bœ .
6:5
mœ . nœ œ œ œ
? B bœ œ nœ œ œ œ bœ œ œ j
79
∑ Œ œ. œ bœ bœ œ mœ œ.
mœ œ
Bsn.
f 7:6
7:6
79
Hn. & ∑ ∑ ∑
4
79
Tpt. &4 ∑ ∑ ∑
?
79
Tbn. ∑ ∑ ∑
œ #œ œ bœ bœ œ n œ œ œ #œ œ œ #œ
œ bœ œ nœ œ bœ b œ œ œ b œ n œ n œ
œ œ
nœ #œ bœ œ # œ bœ
79
œ ⋲ œ œ œ
Perc & bœ 5
œ bœ œ 5
nœ #œ 5
79
4
&
15
&4
Synth
œ
?
79
Comp. & ∑ ∑ ∑
& ∑ ∑ ∑
15
& ∑ ∑ ∑
Kbd.
? ∑ ∑ ∑
*
bœ œ œ. .
œ œ œ mœ œ œ œ œ nœ µœ
œ
J
œ mœ œ œ œ œ m œ œ œ œ œ nœ µœ œ bœ œ œ
œ œœ r
79
& œ mœ œ œ œ œ
5
Vn. 6:5 œ œ. œ mœ œ 5
5
4 5 5
5
‰ bœ bœ œ œ µœ œ œ œ bœ œ mœ œ nœ nœ œ
&4 µœ œ J R ⋲ ‰ j j œ Âœ œ
œ J µœ œ œ œ . œ nœ
5:3
R
œ µœ œ . œ µœ
Vla.
j r
4œ :5œ
3
j r
5:3
4 œ :5 œ
5:3 5:3 5
œ nœ œ B mœ œ . bœ œ bœ nœ œ œ mœ nœ
bœ nœ µœ œ œ mœ n œ
5:3 9:8
? œ µœ µœ
5:3
œ œ µœ µœ œ œ œ mœ œ œ µœ œ œ
Vc.
3
œ œ µœ œ 9:8
Âœ .
9:8
Cb.
? ∑ ∑ ∑
3/16/08
œ œ mœ œ m œ ⋲ œ œ bœ nœ b œ n œ bœ . œ. œ œ mœ œ m œ œ œ b œ n œ bœ œ bœ . œ œœ
œ mœ œ mœ bœ nœ bœ . œ.
mœ œ mœ nœ œ nœ mœ œ mœ nœ œ nœ œ mœ . œ œ
œ
82
Fl. & 6
mœ 6
6 3
7 7
6 6
3 5:3 5:3
4 7 7
œ œ œ mœ œ mœ œ œ œ. œ mœ œ œ œ. œ œ œ
82
œ œ mœ mœ mœ mœ nœ mœ œ œ mœ mœ mœ nœ mœ œ œ mœ mœ nœ œ mœ mœ œ œ
Ob. & 4 5 5 4 5 5
5 5 5
œ. œœ œ œ . nœ œ.
6:5 6:5
œ œ nœ bœ bœ œ bœ
82
œbœ œ œ œ œ b œ m œ nœ
6
œ œbœ bœ œ œ œ œœ
bœ b œ œ nœ œ œ ⋲
œ œ œ
bœ b œ
Cl. & 7:5
œ mœ œ . œ mœ œ œ 6 œ
6 6
7:5 3 6
6 6
bœ œ bœ nœ œ œJ bœ b œ n œ œJ bœ
6
œ.
B œ b œ œ mœ J œ b œ œ b œ bœ œ œ n œ bœ n œ
6:5
bœ bœ œ mœ œ nœ mœ œ œ b œ œ b œ bœ R mœ œ œ œ
82
Bsn. œ
7:6 6:5
6:5 6:5 5
6:5 5
82
Hn. & ∑ ∑ ∑
3 4
82
& ∑ ∑ ∑
Tpt.
4 4
?
82
Tbn. ∑ ∑ ∑
œ.
bœ nœ œ œ bœ bœ n œ œ bœ n œ œ bœ œ
bœ nœ œ bœ bœ n œ œ bœ œ œ
œœ
#œ #œ . bœ #œ #œ #œ
82
Perc & R bœ œ œ œ bœ œ œ bœ œ œ bœ
6 5 5
5
6:5 6:5 6:5 5 5
82 / 148
3 4
&
15
& 4 4
Synth
? Bœ
82
Comp. & ∑ ∑ ∑
& ∑ ∑ ∑
15
& ∑ ∑ ∑
Kbd.
? Bœ Œ Ó ∑ ∑
mœ nœ œ mœ nœ µœ mœ µœ œ mœ œ nœ bœ µœ µœ œ mœ
bœ œ µ œ Âœ . n œ œ
bœ œ œ n œ m œ œ œ . bœ œ µœ
œ mœ µœ nœ bœ œ œ
œ µœ œ µœ œœ J œ œ
82
œ µ œ µ
Vn. & œ 5
6 6
5
5 5
3 7:5
7:5 4 6 6
bœ œ m œ œ mœ œ mœ œ mœ nœ œ mœ . mœ nœ œ œ µœ bœ nœ mœ œ
& œ µœ œ œ
nœ œ µœ œ µœ œ œ bœ µœ œ µœ œ œ bœ œ µœ R
Vla.
4 µœ 4 7:5
7:5 7:5 7:5
5
œ. bœ mœ n œ œ bœ mœ nœ . œ . bœ œ nœ . bœ œ œ n œ m œ nœ
B mœ œ bœ nœ œ nœ
œ µœ R J œ bœ n œ œ œ
œ µœ bœ nœ œ œ µœ mœ bœ n œ
Vc.
6:5 Âœ œ µ œ œ œ 5 5
5
5 5 5 5
6:5 5
6:5
Cb.
? ∑ ∑ ∑
3/16/08
œ œ mœ œ mœ bœ nœ bœ œ œ. œ œ n œ œ m œ œ b œ n œ bœ . œ. œ œ mœ œ mœ bœ nœ bœ œ. œ œ mœ œ mœ nœ bœ œ
œ œ œ œ
85
Fl. & 7 7
3
œ œ œ. œ œ. œ œ.
85
mœ mœ œ mœ mœ œ œ mœ mœ mœ mœ œ œ mœ mœ mœ œ œ mœ mœ œ . mœ œ œ m œ mœ œ mœ
Ob. & 5 4 7 7
5 6 7 7
œ œ. œ œ œ bœ œ œ. nœ bœ bœ œ œ œ œ nœ bœ œ œ nœ bœ
œ œ œ bœ b œ nœ œ
œ œ œ œ œ œ œ
œ œ
85
Cl. & 6 5:3

6 3 5:3 5:3
7 7
œ bœ nœ bœ nœ bœ bœ bœ
B bœ œ bœ bœ œ œ b œ n œ bœ b œ œ b œ n œ bœ bœ œ bœ œ bœ nœ bœ bœ œ bœ
85
œ bœ bœ nœ bœ bœ
3
Bsn.
5 7:5 7:5 7:5 7:5
-̇ œ- œ- -̇
85
Hn. & ∑
3
F
3 3
85
& ∑ ∑
Tpt.
4
?
85
Tbn. ∑ ∑
85
bœ œ œ
#œ bœ œ #œ œ bœ œ bœ
œ
R #œ œ bœ
œ
#œ bœ #>œ
Perc & œ œ bœ œ œ œ bœ œ bœ œ œ bœ
5
3
5 5 5 7:5 6
85
3
&
15
& 4
Synth
85
Comp. & ∑ ∑
& ∑ ∑
15
& ∑ ∑
Kbd.
? ∑ ∑
œ. nœ µœ nœ œ mœ nœ µœ œ mœ nœ µœ mœ nœ bœ µœ œ bœ µœ µœ
bœ œ bœ œ bœ œ œ µœ œ œ µœ œ œ µœ bœ œ œ µœ
œ œ
85
œ µ œ µ
Vn. &
7
6 6 7
3 7 7
mœ œ œ µœ mœ œ mœ œ œ µœ œ œ œ µœ mœ œ œ œ µœ œ œ µ œ œ µ œ œ œ µ œ œ µ œ œ µ œ œ µ œ œ
Vla. & µœ œ µœ œ µœ œ 4 µœ
3
7:5 7:5 6 6 6 6
œ bœ œ nœ bœ œ nœ bœ œ bœ œ bœ œ
B bœ nœ œ œ bœ nœ œ œ b œ n œ bœ œ œ bœ bœ œ n œ b œ b œ œ œ n œ bœ b œ œ œ n œ
Vc.
3 6 6 6 6 3 5:3 5:3
Cb.
? ∑ ∑
3/16/08
œ œ bœ œ œ œ œ
87 œ mœ œ mœ œ mœ œ mœ œ œ mœ œ mœ œ œ mœ œ mœ œ œ œ mœ œ œ œ mœ œ œ mœ œ œ mœ œ mœ œ mœ œ mœ œ mœ ®
mœ mœ mœ mœ mœ mœ mœ mœ
Fl. &
cresc… 4
œ
mœ mœ mœ œ œ mœ mœ mœ œ œ mœ mœ mœ œ œ mœ mœ œ œ mœ
m œ œ m œ m œ œ m œ m œ m œ mœ m œ m œ mœ m œ m œ m œ m œ m œ m œ m œ m œ
& œ œ œ œ œ œ œ œ
87
Ob.
4 7
7 7 7 7 7 7
cresc…
œ œ œ œ œ bœ n œ œ œ œ b œ œ œ œ bœ œ œ œ bœ bœ bœ bœ bœ bœ bœ bœ bœ bœ
œ œ œ œ œ œ œ œ œ œ œ œ œ œ œ œ œ œ œ œ œ œ œ œ œ ⋲
87
Cl. &
cresc…
7 7 7 7
B œ bœ œ bœ œ bœ œ b œ œ œ bœ b œ œ n œ b œ bœ œ n œ b œ b œ œ b œ bœ œ b œ bœ œ bœ b œ œ b œ b œ œ b œ b œ
nœ bœ bœ bœ nœ bœ
87
Bsn. œ bœ bœ nœ bœ bœ
cresc… 6 6 7
- œ- œ- œ- œ- -̇ œ-
87
Hn. & œ
F cresc…
4 5
f
87
& ∑ ∑
Tpt.
4
?
87
Tbn. ∑ ∑
b œ #>œ bœ > bœ > œ bœ b>œ œ b œ b>œ œ b œ b>œ œ b œ > œ b œ > œ b œ >œ bœ œ bœ œ bœ

bœ œ œ # œ bœ œ œ # œ
87
Perc & œ œ bœ œ œ œ œ œ œ
3 8:5 8:5 8:5 8:5 8:5
cresc…
87
4
&
15
& 4
Synth
87
Comp. & ∑ ∑
& ∑ ∑
15
& ∑ ∑
Kbd.
? ∑ ∑
µœ µœ µœ mœ µœ mœ µœ mœ µœ mœ µœ mœ µœ mœ µœ mœ µœ mœ µœ mœ µœ mœ µœ mœ µœ mœ µœ mœ µ œ mœ
bœ œ
œ µ œ œ œ µœ œ µ œ œ µœ œ µœ œ µœ œ µœ µœ µœ µœ µœ µœ µœ µœ µœ µœ
87
Vn. &
4
œ œ œ œ œ œ œ œ œ œ œ
cresc…
µœ œ µœ µœ ⋲ µœ µœ µœ µœ µœ µœ µœ µœ µœ µœ µœ µœ
Vla. & µœ œ µœ œ µœ œ µœ µœ µœ 4 µœ µœ µœ µœ µœ µœ µœ µœ µœ µœ ®
6
cresc… 7 œ œ œ œ œ
B b œ b œ œ n œ b œ b œ œ n œ b œ bœ œ b œ b œ œ b œ b œ œ bœ œ œ b œ œ œ œ
œ
œ
œ
œ
œ
œ
œ
œ
œ
œ
œ
œ
œ
œ
œ œ
œ ?
bœ bœ bœ bœ bœ bœ bœ bœ bœ
æ
Vc.
7
cresc…
Cb.
? ∑ ∑
3/16/08
q = 96 ˘ ˘ ˘ ˘ ˘ m˙ .
w mw ˙. m˙ . œ̆ œ̆ ˘ œ̆ m œ œ̆ œ̆ ˘ œ̆ m œ œ̆ œ̆ ˘ œ̆ m œ œ̆ œ̆ ˘ œ̆ m œ œ̆ œ̆ ˘ œ̆ m œ
Âœ Âœ Âœ Âœ Âœ
89
Fl. &
ƒ f 3 Ï
5 5 5 5 5
3 5 5
œ̊ ˙.
89
mw mw m˙ . m˙ . m˚œ n˚œ µ˚œ m˚œ
Ob. & 4 4 4 4
ƒ f Ï
bw bw b˙ . b˙ . ˘ ˘ ˘ ˘ ˘ ˘ ˘ ˘ ˘ m˙ .
89
B˘œ œ̆ œ̆ œ̆ m œ Bœ œ̆ œ̆ œ̆ m œ B œ œ̆ œ̆ œ̆ m œ Bœ œ̆ œ̆ œ̆ m œ Bœ œ̆ œ̆ œ̆ m œ
Cl. &
ƒ f 5 5 5 5 5 Ï
B w œ bœ œ . n œ œ m œ œ b œ. œ µ œ m œ œ bœ. œ m œ œ b œ. œ m œ œ b œ. œ. œ.
> . . > . . > . .
œ > . .
œ > ˙.
89
Bsn. Ó mœ
ƒ Ï
f
5 5 5 5
5
˚ Bb ˙ .
& ‰ Bb œJ ˙ . ˙.
89
bœ œ bœ mœ
  œ
5 
Hn.
fÍ 3 ƒ  3 Ï 5
f
Bb ˙ .
& w ˙. mœ mœ nœ œ mœ mœ nœ œ mœ mœ nœ œ mœ mœ n œ œ m œ mœ nœ œ
89
Tpt.
4 4 œ œ œ œ œ 4 4
fÍ ƒ f Ï
w ˙. B˙ .
5 5 5 5 5
?
89
B˚œ œ̊ n˚œ b˚œ µ˚œ
Tbn.
fÍ ƒ f Ï
b wæ ~~~~~~~~~~~~~~~~~~~~~~~
sus. cym.
˙.
89
& b www ∑ ∑ ÷
æ
Perc
p
ƒ
89
3 5 3 5
&
15 / 152 / 156 / 160 / 164
µœ
Ó µ µ µ mœœœœœ µÂ œœœ n µ m µBœœœœœœ
& 4 4 œ œ œœ 4 4
Synth
? œ
noise wave
µœ œ µœ µœ
mœ
89
Comp. & ∑ ∑ ∑ ∑
∑ ∑ ∑ ∑
“big pluck”
&
15
& ∑ ∑ ∑ ∑
f
Kbd.
? œ >˙ .
∑ ∑ mœ œ mœ mœ
q = 96
. ˙.
µ m ww µ m ˙˙ . µ˘œ b˘œ µ˘œ œ̆ µ˘œ b˘œ µ˘œ œ̆ µ˘œ b˘œ µ˘œ œ̆ µ˘œ b˘œ µ˘œ œ̆ µ˘œ b˘œ µ˘œ œ̆ ˙.
æ æ œ̆ œ̆ œ̆ œ̆ œ̆ æ
89
Vn. &
ƒ 3 5 f 5 5 5 5 5
3 Ï 5
w ˙. œ̊
µw µ˙ . œ œ̊ œ̊ B ˙æ.
Vla. & æ 4 æ B
4 œ̊ Œ & 4 ˙. 4
ƒ f Ï
?
Bw
æ B˙æ. & Œ bœ œ œ ˙æ.
œ m œœ  b˙ .
Vc.
f  mœ Ï
j j j
 j
 j
œ œ œ œ œ m˙ .
?
æ̇. bœ- b œ- b œ- b œ- b œ-
Cb.
w
f ƒ f Ï
3/16/08
w
take Piccolo
93
Fl. & Œ ∑ ∑ ∑ ∑
5 2 4 3 5 4
˙. œ µœ œ mœ >œ ˙. œ >œ . œ. œ
œ œ mœ ⋲ mœ œ mœ
93
Ob. &4 4 >œ 4 4 ˙ œ 8 4

f Ï ƒ
˙ œ œ Bb œ œ œ m>œ ˙. œ m>œ . œ. œ
6 3
bœ
œ bœ œ ⋲ œœ
93
& œ bœ œ
b˙ œ
Cl.
f >
Ï >œ ƒ
3
˙. œ
6
˙. œ œ œ ⋲ œ œ œ. œ. œ
B ? œ œ mœ
93
œ mœ ˙ œ
> >
Bsn.
f 5
Ï
5 ƒ
> > ˙. >
w œ ‰ ? œ & Bb œ mœ œ œ ⋲ ? œ & Bbœ . œ. œ ?
93
& J ˙
fl fl fl
Hn.
5 2 ƒ 4 3Ï ƒ 5 4
˙. œ œ Bb>œ >œ >œ >œ >œ œ Bb>œ ˙. œ
R ⋲ bœ . .
93
œ̆ œ̆
Tpt. &4 4 bœ 4 4 b˙ œ > 8œ œ
4
> f ƒ
3 3
Ï
w œ >œ ˙. œ >œ . œ. œ
? J
93
‰
œ œ
>
Tbn.
ƒ Ï̇ ƒ
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ >œ ‰ >œ >œ >œ >œ æ̇ œ.
hi bongo
93 Chinese cym.
÷ ‰ Ó Œ ⋲ ∑
low conga
w œ
J œ J
> >
Perc
f Ï f f F
6
93
5 2 4 3 5 4
&
15 / 168
&4 4 4 4 8 4
Synth
? œ ˙. Œ œ. œ. œ
bœ b˙ œ
93
Comp. & ∑ ∑ ∑ ∑ ∑
∑ ∑ ∑ ∑ ∑
“deep synth”
&
15
& ∑ ∑ ∑ ∑ ∑
Kbd.
? w Œ œ ˙. Œ œ. œ. œ
bœ b˙ œ
˙. œ. œ œ bœ œ
˙. œ. >œ ˙ œ œ œ mœ >œ . œ. œ
bœ B œ œ m˚œ m˚˙
æ æ Bœ œ œ
93
Vn. &
f 2Ï f ƒ
6
5 4 3 5 4
3
B>œ ˙ œ œ B>œ . œ. œ
˙æ̇. œ . B µœ œ œ œ œ œ⋲ œ œ œ
5
&4 . œ 4 œ œ ˙ œ œ. œ. œ
Vla.
œ. 4 4 œ 8 4
f  ̇
Ï f
5
ƒ
& æw bœ nœ ˙ œ œ ⋲ œ µœ œ
5
b˙ œ nœ . œ. œ
µœ œ
5
w œ nœ œ  b>œ ˙ œ œ  b>œ . œ. œ
Vc.
f Ï f ƒ
? w µœ œ œ œ m>œ ˙. œ m>œ . œ. œ
bœ œ œ œ b˙ œ
 
Cb.
f Ï f ƒ
3/16/08
98
Fl. & ∑ ∑ ∑ ∑
4 5 9 4
œ̊5 j œ. œ œ Bœ
œ œ œ mœ J
98
œ œ
Ob. &4
>
œw œ ˙. 8 8 mœ- m œ. >œ >œ œ µ œ 4
F ƒ Ï P F ƒ
m˚œ œj
6
mœ . œ œ œ Bœ
˙. bœ œ m œ J
98
& œ bœ
> œ bw œ- m œ. b>œ n>œ œ µ œ
Cl.
F ƒ Ï P F ƒ
œ œ mœ mœ
6
œ j œ. œ >
>œ œ œ m œ m œ œ
? . J
98
œ mœ œ mœ m
> w  œ ˙. œ-
Bsn.
F ƒ Ï P F ƒ
6
? æw Bb˚œ ?œ ˙. µœ . œ
98
flz.
& & j
flz.
Hn.
4 f F ƒ
æ
5Ï æ 9 mœp-
bœ .
œ. >
œ œ œ. œ 4
F
æ . œ. œ œ œ. œ
98
bœ
8 bæœ œ
flz.
&4 J
flz.
bw  bœ ˙. æ 8 mœ- œ. > 4
Tpt.
f F ƒ ƒ p F
œ̊ œ. œ œ. œ
? æw æ æ bœ. µ>œ . œ œ J
98 flz.
flz.
Tbn.
œ ˙. œ-
f F ƒ Ï p F
æ̇. x. j
temple blocks
Œ.
98
÷ Œ œ Œ x
œ- ‰ Ó
low bongo
Perc
fl
œ ˙ æ æ
f ƒ f f (senza dim.) p
98
4 5 9 c1 / 12 4
&
15
&4 8 8 4
Synth
? œ œ. œ Œ.
bw bœ ˙. ˙.
98
˙.
Comp. & ∑ ∑ ∑ ∑
& ∑ ∑ ∑ ∑
15
& ∑ ∑ ∑ ∑
Kbd.
? œ œ. œ Œ.
bw bœ ˙. ˙.
˙.
œ̊ œj œ. œ œ µœ œ
œ- œ µ œ œ m œ œ
bœ œ
mw mœ ˙. æ æ œ m œ- J
98
Vn. & µ œ-
9P F ƒ
6
4 5 4
B˚œ Bœ . œ Âœ
œ œj œ. œ µ œ Âœ œ œ
B mœ œ & µ œ µ œ- µ œ- œ œ m œ
Vla.
4 w œ ˙. 8 æ æ 8 b œ-
P 6 ƒ
J 4
œæ. œæ œ œ mœ œ j B
6
& µœ œ bw œ j bœ ˙. bœ µ œ- µ œ- œ Â œ µ œ nœ
b œ µœ bœ . œ œ-
Vc.
 P F ƒ
? µ˚œ j µœ œ µ˙
Cb. œ œ bw œ bœ ˙. J Œ Œ.
p f p
3/16/08
q = 108
œ œ mœ œ mœ œ œ r
6
j
Piccolo
œ ‰. mœ œ œ mœ œ
102
Œ ‰ œ œ œœ ‰ Œ ∑
6
Picc. & œ mœ
4 p f p 2 4
˚
6
œ Bœ œ œ œ œ œ œ ˚ ˚ œ̊ m œ ⋲
œ œ œ œ >œ œ Bœ mœ µœ œ
œ œ œ œ̊ b œ µ œ
102
&4 œ œ mœ œ œ B Œ ‰ ‰ œ œ
4‰
3
> 4
Ob.
f f ƒ
6
ƒ
6
P
6 3
5
Bœ œ > œ ˚ ˚ Â˚œ µ˚œ œ̊

5 6
œ œ œ n œ µ œ œJ
6 6
n
6
œ œ B
Bœ b œ b œ œ ‰ ‰ œ œ œ µ œ œJ
102
& bœ œ œ nœ œ ‰ bœ ⋲
œ mœ
3
Bœ b
œ mœ œ œ bœ œ
Cl.
>
P f ƒ f
3
ƒ
bœ œ œ œ œ̊ m˚œ ⋲
? m œ m>œ œ œ bœ œ bœ nœ nœ nœ œ̊ m˚œ œ̊
102
bœ nœ mœ Œ Ó ‰
œ
Bsn.
P f ƒ
5
ƒ
5
6
j j
102
& j ‰ ‰ Bœ œ œ Bœ ‰ Œ Ó ∑
4 œp fl
Hn.
p F S 2 4
102
œ j ˘ ˚ ˚ m˚œ œ̊ Bb˚œ ⋲
‰ ‰ bœ ‰ Œ Ó
Tpt. &4 J bœ œ œ
J 4‰ mœ bœ 4
p p F S ƒ
œ
? J œ̆
102
‰ µ˙ . ‰ Œ Ó ∑
Tbn.
J
p p S
hi bongo
œ̆
102
Perc ÷ ∑ J ‰ Œ Ó ∑
S
102
4 2 4
&
15 c3 / 20 / 28
j b œ.
B œœ m n œœ B œœ ..
3
nœ
3
&4 Ó
µ˙
m Bµ ˙˙˙˙˙ Bµb bœœœœ ‰ µ µ m n œœœœ Œ BBµ µ œœœœœœ Œ ‰ µ œœœ ...
Synth n œ 4 Â œ. 4
bœ mœ
? Ó m˙
Âœœ ‰ œ Œ µ œœœ Œ ‰ Âm œœ ..
3 J
3
102
Comp. & ∑ ∑ ∑
∑
“big pluck”
&
15
j bœœœ
œ m m œœœœ
3
nœ
3
& ∑ bœœœ ‰ m œœ Œ n œœœœ

m Œ ‰ J Œ
b œœ
b m œ
J
Kbd.
F œœ
? ∑ ∑ ‰ mœ Œ
J
œ Bb œ œ µœ µœ q = 108
´ µ œ´
mœ œ µœ œ Âœ Bœ œ µ œ >œ Âœ Bœ µ œ œ œ ˝ ´ œ́ bœ
œ œ µœ ‰
æ
pizz. arco
Âœ œ́
102
‰ ‰ r Œ ‰ ⋲
æ æ æ
& µœ œ
3
Vn.
œ œ
f
6
4 F
2 ƒ 4
6
5
œ@ œ@ @ @
˝j
6
Bœ m œ ˝ bœ œ
œ mœ Bœ ‰ ‰
6
‰ œr œ œ
pizz.
µœ œ
pizz. arco
B mœ œ µœ ‰
4‰ œ Œ
æ æ æ
&4 µœ
3
µœ œ mœ œ
3
mœ œ 4
Vla.
f 3 5
> 3
µ œ bœ Bœ œ mœ µœ Bœ mœ œ œ ˝j
B œ mæœ æ œ µœ mœ œ
@ @ @œ œ@ œ œ bœ œ ? œ œ
pizz.
Vc.
Âœ ‰ œ ‰ Âœ Œ
f
5 3
6 6 3
˙ œ œ ˝ ˝ ˝j
? µ œ Âœ œ ‰ bœr Œ
pizz.
Œ œœ ‰ ‰ ‰ µœ Œ
3 3
Cb.
mœ
f 5
3/16/08
mœ µ œ bœ œJ œ œ mœ œ
œ œ œ mœ œ œ mœ œ mœ
⋲ œ m œ œ m œ ⋲ b œ n œ œ n œ ⋲ œ bœ œ œ ⋲ . œ œ œ œ bœ œ œ m œ n œ œ bœ n œ œ n œ m œ
105
Picc. & ‰ ∑
f ƒ 4
6
4 6 7 2 5 5
bœ n œ n œ m œ
5 7
µœ œ mœ œ
⋲ mœ mœ œ mœ ⋲ œ mœ œ
bœ
⋲ œ bœ œ J µœ œ m œ œ œ
mœ µœ n œ b œ œ œ m œ
œ µœ œ b œ b œ n œ
µœ œ mœ Âœ Âœ m œ œ m œ
105
‰
Ob. &4 4 Œ 4 4
Ï
7
f f
5 6
6 3 5
7
Bœ œ B œ œR µœ œ bœ nœ œ µœ œ
⋲ bœ bœ ⋲ nœ bœ œ œ œ µœ mœ nœ
µœ mœ R ‰. œ
105
Œ ⋲ j œ
5
µœ œ m
3
& œ µœ œ œ œ
µœ œ œ
m n
µœ
Cl.
œ Ï
7
f f
6
5 3
S
? œ mœ œ œ mœ
j >œ m œ œ b œ œ œ m
m ⋲ mœ œ n
105
⋲ œ Œ ‰ nœ Œ œ œ œ œ œ œ
 œ œ. > > >
Bsn.
S f ƒ Ï
7
5
5
^j ^
‰ œj
105
Hn. & Œ ⋲ mœ . Œ ∑ Bb w œ
4 S S 2 5 p ƒ 4
105
Bbœ^
Ó Œ ‰ J ∑
Tpt. &4 4 4 Œ . œ 4
S ṗ ƒ
^ µ œ^
? ⋲ µ œj .
105
Œ Œ ‰ J ∑
w œ
Tbn.
S S p f
œ
nœ œ #œ œ œ bœ #œ nœ œ bœ ⋲ ⋲ nœ œ œ
105
÷ ∑ ‰
vibraphone
j
5
& #œ œ
3
bœ œ #œ bœ #œ
Perc
p
j
/ 36 (senza pedale)
µ bœœœ µ œœœ
/ 34
4 2 5 4
⋲ µ Bœœœ...
105
& Bœœœ Ó
15 / 30 / 44
œ
B œ ..
mmœœœ ⋲ m b Bœœœ ... Œ
Bœ
µµ œœœ n Âœœœ µ œœ
& 4 Bµ œœœ 4 mœœ
J µ œ Bbœ
Synth 4 4
? m œœ
105
Comp. & ∑ ∑ ∑
j
& œ ⋲ œœœ... Œ m œœœ œœ ∑ ∑
15
m œm œœ œ.
⋲ b œœ .. Œ œ ∑ ∑
& œ mœ
Kbd. J mœ m œœ
ƒ
? ∑ ∑ ∑
µ œ µ œ µ œ Bœ œ µ œ
œ œ œœ
Âœ œ œ œ m œ œ œ œ n œ œ µ œ µ œ ⋲ Bœ œ œ m œ n œ m œ µ œ
⋲ œ œ œ œ
µ
µœ œ œ
m
mœ œ µœ
105
& Œ ⋲ µœ mœ
7
⋲
7
œ œ œ mœ œ
Vn.
7
6
ƒ
5 6
4 2 5 5 5 6
4
œ œ mœ œ œ
œ mœ œ œ µœ œ œ œ bœ
r
œ mœ œ ⋲
5
⋲ Âœ µ œ
arco 8:5 œ 6
B ⋲ œ Âœ nœ ‰ r œ œ & œ œ œ œ µ œ œ œ m œ
œ œ œ 4 µœ œ µœ
m µ m
mœ mœ œ
Vla.
4 4 7 4
5
5 ƒ
œ œ > mœ œ µœ
œ
? µ œ Âœ Â œ œ œ
Âœ œ
œ œ mœ
mœ
œ œ bœ mœ œ mœ œ
arco
⋲ µœ ⋲ J ‰ œ& œ mœ µœ œ
5
œ . .
6
mœ mœ œ œ œ œ n œ œ
œ
Vc.
5 b 5 6
ƒ
5 7 5
.
? ⋲ µ œJ œ
pizz. arco
Cb. Œ Œ mœ Ó ww œœ
ƒ
3/16/08
>
m˙ µ w˚ b>œ œ m œ m œ œ m œ
q = 66
mœ œ mœ mœ œ œ nœ mœ œ œ bœ
mœ mœ mœ bœ nœ R ‰.
108 6
Picc. & Œ ‰ Ó ∑
f 2ƒ f ƒ
5
4 4
3 7
œ œ œ. ˚˙ m w˚ œ ˙ w
mœ œ œ œ nœ mœ mœ mœ œ >œ
œ œ
œ mœ mœ Âœ µœ µœ œ
⋲ m œ Bœ œ m œ
108
m m
Ob. &4 7 4 4
Ï̇ Ï ƒ
6 5
˚ m w˚
6
ƒ bœ œ œ . Bb œ ˙ w
5
œœ m œ œ mœ >œ
mœ
⋲ m œ Bœ œ µ œ n œ œ œ m œ m œ œ µ œ mœ œ Bœ œ
108 5
Cl. & œ 7
ƒ 6 6
Ï Ï
5
ƒ
˚˙
œ œœB m˚w >œ œ œ ˙ w
? œ mœ œ œ mœ œ mœ
108
Ó ⋲
5
Bsn.
m œ
œ fl fl
ƒ fl fl Ï ƒ F
6
5
108
> ˚ ˙. w
& Ó ‰ œ œ ˙ bw Œ
J
Hn.
4 Í 2ƒ 4Ï p ƒ
108 >œ œ ˙ Bw˚
&4 Ó ‰ J Œ
4 4 b˙ . w
Tpt.
Í ƒ̇ Ï p
µœ œ ƒ
mœ œ œ
? ⋲ r µœ œ µœ
108 gliss. 3
Ó ∑ Ó
œ œ
Tbn.
f ƒ b˙ w
6
VII I p Ï
>˙ œ
œ #œ
6 brass
œ #œ #œ
5
mallets
œ #œ #œ
108
‰ œ ‰ Œ Ó ∑ ∑
3
& œ #œ œ nœ #œ J œ œ
Perc
f ƒ
° * °
œ
5
œ 2
7
œ Bœ œ œ œ Bœ œ Â œ m œ
108
4 4
Ó
15
&
/ 46 / 48 / 50 + / 52 / 54 / 56
Bœ œ Â œ œ µ Âœœ
6
&4 Ó ‰ œ œ œ mœ b œ œ œ B œ m œ m µ œœ Œ B n n œœœ
Synth 4 4
? b µœœ
5
Ó Œ
mœ œ mœ œ œ œ ˙ bw œ
108
Comp. & ∑ ∑ ∑ ∑ ∑
“bells”
œ
7
œ œœ œ œ œ œ œ mœ
5
& Ó ∑ ∑ Ó ∑
15
œ œ œ
bœ œ œ œ mœ mœ
6
& Ó ‰ œ œ œ mœ ∑ ∑ Ó ∑
f
Kbd.
5
? ‰ j Ó Ó ∑ Ó. ∑
mœ mœ
œ
—
œ ˙ mw œ Âœ µœ µœ œ œ ˙
q = 66
w
⋲ œ µ œ m œ Âœ œ œÂ œ
œœ µœ mœ œ æ ˙ w
sul E
œ
æ
108
Vn. &
2Ï 4 fÍ ƒ F
6 5
4
6
mœ œ mœ —
œ B Bœ m œ œ B˙ mœ œ mœ mœ œ ˙ w
æ µw
sul C
Vla. &4 ⋲ µ œ œæ œæ 4 ˙ 4 w & œ

Ï fÍ F
ƒ
5
5
Â—
æ
sul G
? œ œ Bœ œ œ Âœ . œ µœ ˙ w
5
& ⋲ œ Bœ m œ œ ∑ ∑ & œ œ bœ œ
œ
æ æ
Vc.
ƒ F
—
5
? Œ mœ œ
sul G
æ æ Ó
≤ ≥ sul pont.
Cb. mœ
æ æ æ b˙ ˙ mœ ˙ œ w
f ƒ p
3/16/08
D bw w w
113
Picc. & ∑ ∑ ∑
Ï 3 2
113 œ
& Œ Ó ∑ ∑ ∑ ∑ ∑
Ob.
4 4
113 µœ
Cl. & Œ Ó ∑ ∑ ∑ ∑ ∑
B
113
œ
Bsn. Œ Ó ∑ ∑ ∑ ∑ ∑
œ
113
Hn. & Œ Ó ∑ ∑ ∑ ∑ ∑
3 2
113
& Œ Ó ∑ ∑ ∑ ∑ ∑
Tpt.
œ 4 4
?
113
Tbn. Œ Ó ∑ ∑ ∑ ∑ ∑
œ
b b œœ
crotales
113
Perc & Œ Ó ∑ ∑ ∑ ∑ ∑
ƒ
113
bµ b Bb œœœœœœ b˙ 3 2
& Œ Ó Ó
15
Synth
&
ø Ï
4 4
?
113
Comp. & ∑ ∑ ∑ ∑ ∑ ∑
H m œœœœœœ
& Œ Ó ∑ ∑ ∑ ∑ ∑
15
& ƒ ∑ ∑ ∑ ∑ ∑ ∑
Kbd.
? ∑ ∑ ∑ ∑ ∑ ∑
wO wO
D
113
& ∑ ∑ ∑ ∑
ø
Vn.
f 3 2
µ
µw O wO
& ∑ ∑ ∑ ∑
ø 4 4
Vla.
f
œ Œ Ó ∑ ∑ ∑ ∑ ∑
Vc. &
?
ord.
Cb.
œ Œ Ó ∑ ∑ ∑ ∑ ∑
3/16/08
bœ . ˙ . ˙. œ . œŸ~~~~~~~~~~
( m œ)œ œ bœ
b>œ ˙Ÿ~~~~~~~~~~~~~
. (b œ ) œ̆ ⋲ œ ˙. >œ œ mœ mœ
‰. R mœ
119
Picc. & Œ œ˙
2 F 4 5
3 4 ƒ F 6
p
119
&4 ∑ ∑ ∑ ∑ ∑ ∑
Ob.
4 4 4
119
Cl. & ∑ ∑ ∑ ∑ ∑ ∑
B
119
Bsn. ∑ ∑ ∑ ∑ ∑ ∑
119
Hn. & ∑ ∑ ∑ ∑ ∑ ∑
2 4 3 4
119
&4 ∑ ∑ ∑ ∑ ∑ ∑
Tpt.
4 4 4
?
119
Tbn. ∑ ∑ ∑ ∑ ∑ ∑
b b œœ ..
soft vibraphone
b œœ
crotales mallets motor on (slow)
œ
Ó. ⋲ J Ó.
119
Perc & ∑ ∑ ∑ œ
#œ Œ ‰ Œ
p f œ 3
F
°
119
2 4 3 4
&
15
µœ
Bm œœœ BBb œœœœ µ m œœœœ
& 4 µ œœœ Œ
4 BBœœ Œ Ó
4 4 Bb œœ
Synth
p p
? m Bœœ Ó
œ
119
Comp. & ∑ ∑ ∑ ∑ ∑ ∑
∑ ∑ ∑ ∑ ∑
“bowed pluck”
&
15
œœœ b œœœ bœ
œœ œœ
& m œœœ Œ Œ Ó ∑ ∑ ∑ Œ Ó
œ
p p
Kbd.
? ∑ ∑ ∑ ∑ ∑ m œœ Ó
œ
Bœ ˙ . ˙.
arco
mœ µ œœ
pizz.
‰ J BBOœ
119 pizz. arco
& J ‰ Œ ∑ ∑ J ‰ Œ Œ
ø ø
Vn.
2 p 4 3π 4 P
j j
pizz. arco pizz. arco
& 4 Bœ ‰ Œ
4 ‰ µOœ Ȯ .. 4 Ȯ .. ∑ ∑ bœ ‰ Œ Œ m n Oœ
ø 4 J
ø
Vla.
p p P
Bœ µ µ Oœ
pizz.
‰ m Oj Ȯ . ? œ ‰ Œ
arco
& ∑ Ȯ .. ∑ ∑ J Œ
mœ .
ø ø
Vc.
p p
o
arco
pizz.
? Œ &œ
sul G
∑ ∑ ∑ ∑ ∑ Œ œ
ø
Cb.
3/16/08
˙ œ.
m>œ m œ œ m œ bœ . ˙ œ >œ ˙Ÿ~~~~~~~~~~~~~~~~~
. (œ) œ ˙. œ ⋲ œ̆ J
œ œ bœ œ
œ mœ
‰.
125
bœ . bœ R
5
& œ. œ œœ œ
5:3
Picc.
œ R
fl
F f p 3
f p
7
125
& ∑ ∑ ∑ ∑ ∑
Ob.
4
125
Cl. & ∑ ∑ ∑ ∑ ∑
B
125
Bsn. ∑ ∑ ∑ ∑ ∑
125
Hn. & ∑ ∑ ∑ ∑ ∑
3
125
& ∑ ∑ ∑ ∑ ∑
Tpt.
4
?
125
Tbn. ∑ ∑ ∑ ∑ ∑
#œ . œ œ bœ
bœ œ œ œ b œ
œ. œ œ bœ œ
œ bœ
7
œ
5
⋲ #œ .
125
&Œ ‰ Œ bœ ‰ Œ
3 3
œ œ
Perc
J J œ #œ œ.
p
5
3
*° *° *°
j b µBœœœœ
125
3
& µµ µ œœœœœ Œ
15
µ m mµ œœœœœœ µ œœ b bœœœœœ
µ œœœ
&Œ ‰ b µ œœ
µ
œ Œ m µ œœ
mœ
œ π b œœ 4
p
Synth
Bœ p
?Ó ⋲ j Œ Œ Œ ‰ j Œ ⋲ Bœj . Ó
µœ . bœ
p
125
Comp. & ∑ ∑ ∑ ∑ ∑
&Œ ‰ j Ó ∑ ∑ m œœœœ Œ Ó ∑
15
m m m œœœœœ m m œœ b bœœœœœ mœ
&Œ ‰ bœ Ó œ Œ Ó ∑ Œ Ó Œ m œœ Ó
œ p
J F
Kbd.
?Ó œ bœ
⋲ j Œ Œ Œ ‰ j ∑ ∑ Œ œ. Ó
mœ . bœ
O mœ ‰ Œ µœ
125 pizz.
&w ∑ ∑ œ Ó Œ J ‰ Ó
J
Vn.
p p p 3
µœ ‰ Œ
pizz.
Vla. & wO ∑ ∑ J Ó ∑
4
p p
O œ
pizz.
Vc.
?w ∑ ∑ J ‰ Œ Ó ∑
p p
o
Cb. &w ∑ ∑ ∑ ∑
p
3/16/08
>
Ÿ~~~~~~~~~~~~~~~~ œ œ œ m œ œ mœ œ . nœ bœ bœ . œ œ œ m>œ œ
œœ œ mœ œ œ nœ œ
˙ (mœ ) œ. œ bœ œ J bœ m œ œ m œ œ
130
Picc. & m˙ œ œ.. 5

3
3F
7
f 2 p F4 2
f
7 7
3
130
&4 ∑ ∑ ∑ ∑ ∑
Ob.
4 4 4
130
Cl. & ∑ ∑ ∑ ∑ ∑
B
130
Bsn. ∑ ∑ ∑ ∑ ∑
130
Hn. & ∑ ∑ ∑ ∑ ∑
3 2 4 2
130
&4 ∑ ∑ ∑ ∑ ∑
Tpt.
4 4 4
?
130
Tbn. ∑ ∑ ∑ ∑ ∑
bœ œœ ˙
#œ #œ œ #œ n˙ œ
130 7
œ œ bœ ‰ Œ j
3
& œ œ œ bœ œ œ œ
3
œ œ œ œ œ œ
Perc
œ
3
F p
*°
3 bœ 2 4 r 2
Bµ œœœ
130 c11 / 60
& Ó
15
œ Bœ bBœœœ
œ Œ Bmm œœœœ µ µ b µ œœœœœ ‰ . µ BBœœœ
&4
p µœ 4 4 œ 4
p
Synth
? b B œœœœ π Bœ
œ Œ
mœ
µœ
130
Comp. & ∑ ∑ ∑ ∑ ∑
& ∑ Ó mœ ∑ ∑ ∑
15
œ bb b œœœœ
& ∑ œ Œ m œœœœ ∑ Œ m œœ Œ ∑
Kbd. œ
? b œœœ œ
∑ Œ Œ ∑ Œ Ó ∑
Â˙ Â˙ ˙
µœ
130
& ∑ Ó mœ ‰ ∑
J
Vn.
3 π 2 4 π p 2
µ˙ ˙ ˙
&4 ∑ j ‰ Œ Bœ ‰ ∑
J 4 4 4
Vla.
œ π π p
p
? µœ bb Ȯ
∑ œ ‰ Ó ∑ BBȮ Ȯ
J
Vc.
p π p
? j mœ
pizz.
Cb. & ∑ µœ ‰ Ó ∑ ∑ Œ
p P
3/16/08
135
mœ œ œ œ
& œ m œ Bœ œ ˙. Œ ∑ ∑
ø
Picc.
w w
4 3
p
135
& ∑ ∑ ∑ ∑ ∑ ∑
Ob.
4 4
135
Cl. & ∑ ∑ ∑ ∑ ∑ ∑
B
135
Bsn. ∑ ∑ ∑ ∑ ∑ ∑
135
Hn. & ∑ ∑ ∑ ∑ ∑ ∑
4 3
135
& ∑ ∑ ∑ ∑ ∑ ∑
Tpt.
4 4
?
135
Tbn. ∑ ∑ ∑ ∑ ∑ ∑
#˙
œ
135
& ‰ œ bœ œ ˙ Œ œ Œ ‰ ‰ ∑ ∑
5
œ
bœ bœ ˙
Perc
œ 3
7
135 c12 / 66
4 c13 / 72 c15 / 84 c16 / 90
3
&
15
bµbœœœ .. Bœ µ œœœ
& ⋲ œ.
. B œœœ
bœ . 4Œ œ
Ó
4
J π
Synth
? Ó µœ
135
œ
Comp. & ∑ ∑ ∑ ∑ ∑ ∑
j
& ⋲ mœ . Œ ∑ ∑ ∑ ∑ ∑
15
b bœœ . œœ
. œœ
& ⋲ œ. Œ Œ Ó ∑ ∑ ∑ ∑
Kbd. J œ
? ∑ ∑ Ó mœ ∑ ∑ ∑
œ
µœ ˙.
sul A
œ j sul D
‰ m OO OO .. OO
135
& Œ ∑ ∑
ø ø ø
Vn.
4 p 3
µœ ˙.
& œ Œ Ó Œ µ µOœ Ȯ .. Oœ Ȯ .. Œ ∑
4
ø ø ø 4
Vla.
p
Ȯ Ȯ Ȯ Ȯ
? Oœ Oœ Ȯ .. Œ Ó Ó ∑
ø ø ø
Vc.
? ∑ ∑ Ó Œ ∑ ∑ ∑
œ
Cb.
3/16/08
q = 90 q = 60
>œ ˙.
E
˙ m˙ ˙. µœ ˙. œ . bœ ˙ . œ. œ
mœ œ mœ
œ bœ œ mœ œ
141
Picc. & Ó
3 Ï
4f π
141
&4 ∑ ∑ ∑ ∑ ∑ ∑
Ob.
4
141
Cl. & ∑ ∑ ∑ ∑ ∑ ∑
B
141
Bsn. ∑ ∑ ∑ ∑ ∑ ∑
141
Hn. & ∑ ∑ ∑ ∑ ∑ ∑
3 4
141
&4 ∑ ∑ ∑ ∑ ∑ ∑
Tpt.
4
?
141
Tbn. ∑ ∑ ∑ ∑ ∑ ∑
hard mallets
œ >œ
œ# œ œ
141
Perc & Ó œ Œ Ó ∑ ∑ ∑ ∑
f
*°
4˙ B˙ m˙
® m œ µœ µ œ µœ m œ µœ œ ˙
c1
141
3 µ˙
Ó œ bœ œ µœ bœ Ó
5
& œ µœ µœ
15
3
π mœ mœ µœ bœ
&4 Ó
3
4
f
Synth
141
Comp. & ∑ ∑ ∑ ∑ ∑ ∑
˙ b˙ ˙ ˙ m˙
"slow synth”
∑ œ bœ œ mœ bœ Ó ∑
5
& œ mœ mœ
15
3
π mœ mœ nœ bœ
& ∑ ∑ ∑ ∑ Ó ∑
3
Kbd.
? ∑ ∑ ∑ ∑ ∑ ∑
°
q = 90 E q = 60
Oœ Ȯ .. Ȯ .. Oœ m Oœ Ȯ .. Oœ µ B Oœ Oœ Ȯ Oœ m O O
mœœ
141
& ∑ ∑ ‰ J J
ø
Vn.
J
p
3
3 4
3
&4 ∑ ∑ ∑ ∑ ∑ ∑ B
Vla.
4
Vc.
? ∑ ∑ ∑ ∑ ∑ ∑
Cb.
? ∑ ∑ ∑ ∑ ∑ ∑
3/16/08
˙. œ µœ . ˙ œ. œ œ œ œ bœ œ. œ œ ˙ µœ œ . . Âœ œ
J œ . nœ œ œ
147
Picc. & 5
p
5 3 3
5
œ
‰ J
147
Ob. & ∑ ∑ ∑ ∑ Ó Œ
∏
147
Cl. & ∑ ∑ ∑ ∑ ∑
B
147
Bsn. ∑ ∑ ∑ ∑ ∑
147
Hn. & ∑ ∑ ∑ ∑ ∑
147
Tpt. & ∑ ∑ ∑ ∑ ∑
?
147
Tbn. ∑ ∑ ∑ ∑ ∑
crotale
147
œ
∑ ∑ ∑ ∑ Ó Œ ‰ J
take bow arco
Perc &
π
*
147
B˙ µ˙ ˙ µ˙ Bœ µœ œ
&Ó œ bœ œ µœ Œ
15
µœ œ
5
π Âœ nœ bœ mœ m œ µ œ bœ µ œ n œ
&
3
Synth
6
147
Comp. & ∑ ∑ ∑ ∑ ∑
b˙ bœ
“bowed synth pluck”
˙ ˙ m˙ œ œ
&Ó œ bœ œ mœ Œ
15
mœ œ
5
π œ œ bœ mœ m œ n œ
∑ ∑ ∑ ∑ bœ m œ nœ
&
3
Kbd. 6
? ∑ ∑ ∑ ∑ ∑
*°
µœ œ
sul tasto.
Ȯ Oœ µ µOœ . Oœ Oœ r
b b Oœ .. Ȯ O O.
sul A
O O O ..
5
. Oœ Â µ Oœ Ȯ
147 sul A
Vn. & J µœ µ BOœ Oœ

p P p
BO Ȯ Oœ Ȯ Oœ Ȯ .. Oœ ..
BŒ Œ µ œJ . µ µ Oœ Ȯ .. Oœ .. Oœ Ȯ Oœ .. .. b b Oœ Oœ
ø
Vla.
3 5
p 5
Vc.
? ∑ ∑ ∑ ∑ ∑
Cb.
? ∑ ∑ ∑ ∑ ∑
3/16/08
w œ œ œ. ˙ œ . µœ ˙ œ œ.
œ
& J µœ œ Âœ œ nœ J
152
bœ œ µœ . mœ . mœ µœ bœ ‰ Œ
5 3
Picc.
J J
3
∏ π
˙. œ µœ . ˙ œ œ œ œ œ œ. œ œ . µœ œ
152
œ . . µœ œ œ µœ œ œ nœ . œ œ œ œ
Ob. & R J
p
5 5 3
152 Bbœ ˙ œ bœ
& ∑ ∑ ∑ ∑ ‰
ø
Cl.
3
B
152
Bsn. ∑ ∑ ∑ ∑ ∑
152
Hn. & ∑ ∑ ∑ ∑ ∑
152
Tpt. & ∑ ∑ ∑ ∑ ∑
?
152
Tbn. ∑ ∑ ∑ ∑ ∑
152
œ œ. œ
Perc & Œ Ó ∑ ∑ ∑ ⋲J Ó
π
˙
c2
˙ Ó ˙
& µ œ œ bœ µ˙
152 3
œ µœ µ˙ µ˙
5
Bœ
15
œ µœ
6
( π) Âœ n œ bœ mœ mœ mœ Âœ œ
µ œ b œ µ œ n œ Bœ µ œ
& B œ µ œ Bœ œ Ó Bœ 5
Synth
µœ nœ œ
7
œ Bœ Bœ
? Ó Œ Ó b œ Bœ œ Bœ
µœ mœ
p œ
152
F
Comp. & ∑ ∑ ∑ ∑ ∑
π̇ π
˙ Ó ˙
& m œ œ bœ m˙ n˙
3
œ mœ m˙
5
œ
15
œ œ œ
6
( π) œ bœ mœ mœ œ bœ mœ nœ ( π) mœ œ mœ
& œ œ œ œ Ó ∑ œ mœ œ ∑ 5
Kbd.
œ œ œ bœ p
7
mœ nœ œ œ œ œ
? Ó mœ Œ Ó ∑ ∑
œ mœ
p F *°
œ bœ . œ Oœ Oœ Ȯ ..
m Oœ
ord.
µœ œ mœ . œ . µœ œ . . mœ œ r 5 j3
sul pont. sul tasto.
Âœ nœ
152 5
& œ œ œ . µœ œ œ Bœ j ‰ J
3
Bœ œ bœ .
ø
Vn.
P p
5
∏
5
œ . Bœ œ j
Oœ & µ œ œ
sul tasto.
B Oœ .. œ . µœ œ œ µ œ œ . Bœ œ . . B œ œ . Bœ œ B œ œ bœ Bœ œ B œ
sul tasto.
Oœ b bOœ Ȯ
sul pont.
œ œ. œ œ.
5
Vla.
J
3 p P
3
∏
Vc.
? ∑ ∑ ∑ ∑ ∑
Cb.
? ∑ ∑ ∑ ∑ ∑
3/16/08
œ œ µœ œ œ mœ œ œ Âœ . œ ˙ ˙ œ µœ œ
µœ œ µœ œ mœ . œ J
5 5
œ . œ µœ . µœ .
157
R J œ bœ œ ‰ ‰
5
& J
3
Picc.
p ∏ ∏
3 5 3
3
3 7
Bb œ . œ œ œ bœ . œ œ. œ ˙
r 5 µœ œ . µœ œ œ nœ . œ .
& mœ œ nœ . . œ œ œ . µ œ œ Bœ œ œ œ ⋲ J œ
157 5
5
œ 4 8
Ob.
p π π P P
5
157 ˙ œ œ. œ œ . Bœ œ œ Bb œ œ œ œ. Bbœ
œ Bœ œ œ bœ œ. n œ œ œ . Bœ œ Bœ ‰.
5
& Bœ bœ œ œ r
3
J
ø
Cl.
p p œ
5
3
∏
B
157
Bsn. ∑ ∑ ∑ ∑ ∑
157 con sord.
& ∑ ∑ Ó Œ
øœ ˙. 3 ˙.
Hn.
œ 7
157
p
& ∑ ∑ ∑ ∑ ∑
Tpt.
4 8
µœ ˙. œ ˙.
con sord.
?
157
∑ ∑ Ó Œ
ø
Tbn.
p
157
œ. œ
Perc & ∑ ∑ ∑ ‰ Ó ∑
p
5
˙
& ˙ œ
157
µ˙ 3 µ˙ 7
µ˙ Bœ Œ Œ µœ
15
µœ
3
mœ Âœ œ
Bœ œ œ œ œ Bœ œ œ œ œ µ œ Bœ
& µ œ Bœ œ Bœ Bœ œ Bœ Bœ
3
Synth bœ Bœ 6
4 8
5 7 5
? œ Bœ bœ œ œ
3
œ
mœ Œ Ó mœ ‰
157
Comp. & ∑ ∑ ∑ ∑ ∑
π̇
& ˙ m˙ n˙ œ ∑ Œ œ Œ m˙ nœ
15
œ
3
mœ œ œ
œ œ œ œ œ œ œ œ œ œ mœ nœ ( π) ∑
& mœ nœ œ œ œ œ œ
3
bœ œ œ
( π)
Kbd. 5 6
P
5 7
? œœ bœ œ œ
3
∑ ∑ œ
mœ Œ Ó ∑ mœ ‰
p
Ȯ .. Oœ O . O. Oœ Ȯ Oœ µ µ Oœ . Oœ Oœ Oœ Oœ Oœ Oœ µ µ Oœ j
. Ȯ Oœ m O O
157
& R mœœ
R
Vn.
3 P 7
5 5 3
µ µ Oœ Ȯ Oœ .. Oœ Ȯ
ord.
B ‰ J m m Oœ Oœ Oœ Oœ µ O . .
µ œ.. Oœ Oœ µ µ Oœ Oœ Oœ m m Oœ Oœ Oœ .. Oœ O Oœ ÂO Oœ
4 œ JÂ œ
ø J
Vla.
5 8
p 3 5
P 3
.
B ‰ BœJ ˙ œ ˙.
sul tasto.
? ∑ ∑ ∑
ø
Vc.
π
5
? m˙ . ˙. œ ˙.
arco
∑ ∑ Œ
ø
Cb.
3/16/08
œ œ mœ . œ œµ œ . œ. r
œ œ . m œ œ n œ . œ b œj ‰
take Flute
µœ œ
162
Bœ Œ ∑
5
µœ
5
&
6
Picc.
7 p 4 ∏
mœ œ. µœ œ œ nœ . œ . µœ œ. mœ œ œ nœ œ Bœ Bb œ Bœ
&8 œ mœ nœ J
162
œœ œ R
3
Ob.
RJ 4 œ Â œ n œ œ œ ⋲ 3
p π P
5 5
5
œ œ bœ œ œ œ Bœ . œ œ Bœ œ œ
J RJ J mœ œ œ nœ . œ Bœ œ bœ œ. œ œ œ œ bœ .
162 3
Cl. & R J J J œ œ Bœ Bœ
P
3 5
P
3 5
5
B
162
Bsn. ∑ ∑ ∑ ∑
Œ. ? Ó
162
& Œ ∑ ∑ Œ œ
ø ø
Hn.
7 œ 4
162
Bb œ . ˙ œ mœ ˙ œ nœ . . œ
&8 ∑ ∑ ⋲ J J
Tpt.
4
π
3
? œ
Œ.
162
Œ ∑ ∑ Ó µ˙
ø ø
Tbn.
162
œ. œ
Perc & ∑ ∑ ⋲ J Ó ∑
p
c3
162
7 4 ˙ m˙
3
œ mœ
& ‰ Bœ µ˙ Âœ nœ Ó
15
µœ mœ Bœ mœ µœ µœ Âœ µœ
Bœ mœ nœ Bœ œ µœ Âœ
&8 œ bœ mœ
Synth 4 5
µœ œ 5
? ‰ µ œj Œ Ó
œ
162
Comp. & ∑ ∑ ∑ ∑
π π
‰ ˙ m˙ n˙ Ó œ mœ
3
& œ œ œ
15
œ mœ œ mœ nœ mœ nœ œ
œ mœ nœ œ mœ nœ
& ∑ œ œ bœ mœ œ (π)
œ 5
P
Kbd. 5 6
? ∑ ∑ ∑ ‰ bœj Œ Ó
œ
*°
mœ œ
sul tasto.
œ œ
sul pont. sul tasto. ord.
162
Oœ Oœ µ O . j J Âœ œœ œ Âœ . œ nœ œ œ bœ œ œ Bœ µ œ œ
5
mO
Oœ ⋲ Bœ œ ⋲ ‰
5 sul A
& µ œ. Oœ µ µ Oœ Oœ R
ø
Vn.
3 J
p P ∏
5
7 4 3
r œ œ . m œ œJ œ nœ œ œ œ.
sul tasto.
œ œ œ
sul pont.
œ µ œ Âœ n œ Bœ O Oœ
6:5
Oœ ⋲ J œ mœ Âœ nœ ⋲ & œ B BOœ .. Oœ
ord.
B J
8 µBOœ J J
4 Bœ
ø
Vla.
p π
3 5
3
œ Bœ . . œ œ . Bœ œ
B œ œ œ ? BœJ œ œ œ. œ œ Bœ œ
bœ œ œ Bbœ
5
œ œ. œ mœ .
3
Vc.
J J J R µœ mœ nœ
P ∏
3 3 5 3
? ˙. ‰ ∑ Ó Œ ‰ œ w
J
ø ø
Cb.
3/16/08
µœ œ œ. œ œ œ mœ œ œ µœ œ nœ œ µœ œ µœ
Flute œœ œ
J J J R R
166
Fl. & Ó Œ 3 5
π p
3 3 5 3
2 3
œ œ œ mœ µœ œ µœ œ Âœ . .
nœ bœ œ œ J mœ . . nœ œ . bœ œ œ . mœ bœ œ bœ
166
& J J bœ œ B œ œ bœ Œ œ
œ 4 4
Ob.
P p π P
3 3 3
5 6
œ œ Bbœ . œ bœ œ Bœ œ Bœ nœ œ œ.
J J J mœ . nœ œ bœ . œ œ ⋲ J
166
R b œ œ Bœ Bœ
5
& Bœ
œ œ ⋲
ø bœ Bœ œ œ bœ ø
Cl.
P
3 3 3 5
∏
5
∏
œ. œ œ bœ œ bœ œ œ Âœ œ œ . œ . Âœ
B B Œ ⋲ J
166
∑
ø
Bsn.
p
5
5
? ˙
166
˙ ˙ Ó ∑ ∑
ø
Hn.
p 2 3
166
œ Bœ œ œ bœ œ œ œ. œ. œ œ œ œ
bœ
& œ œ. œ œ j 4 ‰ J
3
œ œ 4
Tpt.
p ∏ π
?
166 (remove mute)
˙ ˙ ˙. Œ ∑ ∑
ø
Tbn.
p
166
Perc & ∑ ∑ ∑ ∑
166
2 c4œ 3
& µœ bœ
5
Âœ nœ
15
mœ µœ µœ Âœ µœ
6
mœ n œ Bœ œ µœ Âœ mœ œ Bœ œ
& œ b œ Âœ Œ Ó œ bœ Âœ
Synth
œ œ œ œ 4 4
7
µœ µœ
6
? Œ ‰ Œ œ µœ mœ µœ
µœ œ œ
166
Comp. & ∑ ∑ ∑ ∑
& œ ∑ œ bœ
5
œ œ
15
mœ nœ mœ nœ œ
6
mœ nœ œ mœ nœ mœ œ œ π
& œ œ bœ œ Œ Ó œ œ bœ œ ∑
œ œ œ œ
P
Kbd.
7
6
? ∑ Œ mœ ‰ Œ ∑ mœ œ mœ mœ nœ
bœ œ œ
P
œ œ Bœ œ . Bb œ œ œ Bœ .
sul tasto.
mœ œ nœ œ. œ œ œ œ œ
& O OBBOœ .. Oœ Oœ .. µ µ Oœ Ȯ
166
œ. œ œ
J
Vn.
P p 2 3 3
O ® mœ . œ nœ .
sul tasto.
& œJ µ µ Oœ Oœ Oœ Oœ j œ
sul pont.
Oœ .. B Bb Oœ Oœ Oœ B BOœ .. j j œ œ bœ œ b œ
Oœ µ O bœ œ ⋲ B 4
3 5 3
Oœ
3
R µœ µµ Oœ Oœ BBOœ Oœ J 4
Vla.
p P p π
5
3 5
œœ µœ œ . n œ œ Bœ œ ? mœ
sul tasto.
? ‰ m Oœ Oœ Oœ n b Oœ .. Oœ Oœ .. Oœ Oœ Oœ b bOœ .. Oœ .. B µœ µœ œ
ord.
R R
ø
Vc.
5
p P p P
3 5
5
? w w œ Ó. Œ œ
ø ø
Cb.
3/16/08
bœ œ. Bœ œ œ µœ . . œ œ. µœ œ. µœ
170
œ. µœ œ Âœ . mœ J
& µœ bœ œ m œ n œ œ Bœ µ œ ‰
mœ
Fl.
5
3 p 5
π π
œ œ µœ . . µœ . Bœ œ œ . œ œ µœ œ µœ Bœ .
œ µœ Âœ R J œ
Â œ m œ Bœ b œ œ b œ œ œ ‰ .
170
& 4 œ œ µœ ‰ ⋲
5
Ob.
p π F p π F
5 3 5
5
œ Bb œ œ µœ œ Bœ œ. œ œ. œ
J œ œ. Bœ œ Bœ
J mœ ⋲ R
170
& J bœ œ bœ œ œ B œ µœ mœ B œ
œœ ø
Cl.
3
p F
3
5 3
π
bœ œœ œ œ bœ . œœ œ. Bœ œ µ œ
B œ µœ œ mœ µœ µœ ? œ µœ mœ nœ œ B J R œ mœ . µœ
170
J œ ⋲
Bsn.
π P
3
p
5
3 6
3
?
170 (remove mute)
Ó mœ ˙ œ ˙. œ Ó
ø ø
Hn.
3 p
170
œ. Bb œ œ œ. Bœ œ œ mœ œ bœ œ œ œ mœ . œ nœ .
5 5
&4 œœ
6
Tpt.
J œ œ œ mœ
P 3
? B
170
Tbn. ∑ ∑ ∑ ∑
œ.
marimba
œ œ œ
soft mallets
œ œ œ œ
⋲ J œ #œ .
170 5 6
& ∑ J bœ œ œ. bœ œ œ .
7
œ.
3
œ #œ œ
π F
3
Perc
? ∑ ∑ ∑ ∑
170
3 œ
& œ. Œ bœ œ.
15
µœ . œ. µœ . Bœ µœ . œ. µœ . œ
œ œ µœ œ œ œ
&4 µœ bœ œ bœ œ œ Bœ µœ
Synth 5:3 5
? m œ µœ µ œ
œ µœ mœ ‰
3
µœ œ
6
170
Comp. & ∑ ∑ ∑ ∑
p
& œ. Œ œ bœ œ. ∑
15
mœ . œ. mœ . œ mœ . œ. mœ . œ
œ œ œ œ œ œ
& ∑ œ bœ œ bœ œ œ œ
Kbd. 5:3
œ 5
? m œ mœ m œ
∑ ∑ ∑ œ mœ ‰
3
mœ nœ
œ
*° 6 π
œ.. œœ
µœ µœ . . Oœ Oœ bb Oœ
ord.
n m œœ œœ .. B œœ œœ b œ .
170 5
& œ.. mœ œ mœ m œ.. œ bœ œ Œ

3
œ Bœ µœ mœ µœ
b œ. µœ ø J J
Vn.
3 F π p
Oœ
ord.
Oœ .. BO O Oœ µ O µœ
sul tasto.
mœ µœ œ . bœ œ œ œ bœ œ
sul pont.
µœ œ Oœ .. œ œ
Vla. B
4 J µœ Oœ Oœ O R
π p F
5 5 3
P
5
3
µ µ Oœ Oœ Oœ O . . O n Oœ . Oœ Â O . bœ . œ œ bœ . œ
ord. sul tasto.
œ
sul pont.
? . ÂOœ Â ÂOœ B B œ µœ mœ ?
sul pont.
‰.. Â œ. µœ µœ œ
5
µ œ m œ µœ
5
RÔ µœ m œ
œ
ø R
Vc.
5
p P p F
3
π
? ˙ œ œ œ œ ˙. ∑
ø
Cb.
p f
3/16/08
œ œ mœ œ œ œ bœ œ œ . œ
œ Bœ œœ œ µ œ µ œ Bœ œ œ bœ bœ Bœ µ œ Bœ
& J œ œ b œ Bœ Bœ ⋲J J œ œ bœ Bœ Bœ J
174
R µœ µœ ‰
œ œ
Fl.
P
7 6 3
π π P π p
3
3
œ œ. œ. œ œ bœ bœ B œ µœ B œ œ œ œ œ bœ bœ
µ œ œ Bœ . Bœ . m œ œ Bœ b œ J œ bœ B œ Bœ Bœ J
174 5
œ bœ œ œ œ ‰ Œ
5
Ob. &
p F p p
3 3
p
6
3
5
œ.. Bbœ œ µœ œ . Bœ œ b œ œ µ œ œ Bœ œ . bœ . œ œ. œ bœ bœ
174
R œ œ. J Bœ µœ Bœ
b œ B œ Bœ Bœ ⋲
3
&
œ mœ ø
Cl.
p F p F
5 7
5
π
7
bœ œ . Bœ œ. Bœ œ Bœ œ
B œ œ . µœ . ? nœ B Œ J œ œ Bœ j ? B
174
‰ ‰
5
J
7
œ œ œ œ. œ . bœ œ
3
Bsn.
µœ mœ mœ n œ œ R
p π P p
3
7
?
senza sord.
œ.
174
∑ ∑ ∑ & Ó ‰
Hn.
J
p5
174
Bb œ œ Bœ œ œ œ œ œ bœ . Bœ . œ œ. œ œ œ.
œ œ. ⋲ J bœ
5
& J œœ œ R
3
‰
3
œ œ œ
Tpt.
π p P π p F
3 5
5
œ ˙ Bœ j 3
senza sord.
B œ ? œ œ. µœ .
174
œ
3
J nœ B œ œ œ
J
Tbn.
π P 3
œ. œ. #œ œ #œ nœ œ œ bœ
J œ œ #œ . #œ #œ nœ œ bœ R # œ # œ #œ n œ
‰.
174
& Œ ⋲ œ bœ œ
5
œ œ œ
œœ
F
5
π F p π
5 3
3
? bœ œ bœ bœ œ Ó
Perc
∑ Œ œ bœ œ ‰ Œ ∑
5 3
174 c5
& œ mœ Ó œ mœ Œ
15
œ bœ œ œ bœ œ bœ œ œ bœ
œ b œ µœ B œ œ (F ) œ bœ
& bœ œ bœ œ œ œ b œ Bœ Bœ µ œ ‰
Synth
Bœ œ 3
F
6
œ Bœ œ
6 3
? µœ œ Œ Bœ œ µœ œ
œ 5:3
œ
174
Comp. & ∑ ∑ ∑ ∑
p
& ∑ œ mœ Ó œ mœ Œ
15
œ bœ œ œ bœ œ b œ mœ œ œ bœ œ œ bœ œ bœ
p œ
& bœ œ bœ œ œ ∑ œ bœ œ œ œ œ ‰
œ
P
3
Kbd. 6
3
œ œ œ
6
? ∑ mœ œ Œ ∑ œ œ mœ œ
œ œ
*° π 5:3
sul tasto.
O j œ œ œ. mœ œ µœ Bœ
B BOœ Oœ Oœ Oœ m Oœ Oœ .. œ œ œ bœ Bœ Bœ
174 5 3 6
& œ
5
Oœ Oœ .. bb Oœ Oœ Oœ Oœ
J
Vn.
3 P p F 5
œ. œ mœ .
sul tasto.
B œ Bœ µ œ m œ Bœ µ œ µ œ Bœ . µ œ . ⋲ m œJ Bœ µœ œ œ
ord. sul pont.
œ œ b œ Bœ Bœ Bœ B Bœ
5 5
‰ J J bœ
œ µœ mœ & œ µ œ mœ n œ &
3 5
Vla.
R œ J
π p F
5
3 3
7
bœ . œ œ. Bœ Bœ
sul tasto.
Bœ . m m Oœ ..
sul tasto.
Bœ œ œ œœ 7 ? µ BOœ Oœ nœ
sul tasto.
O
sul pont.
? µœ ‰ ⋲ BJ B œ œ µœ œ r ‰. O n Oœ
œ bœ œ
6
œ
Vc.
π π F π p F
3
? ∑ Œ ˙ œ ˙.
ø̇ ø
Cb.
3/16/08
q = 72
mœ œ bœ œ œ bœ b œ Bœ œ œ µœ œ bœ œ mœ
mœ œ bœ œ mœ nœ œ œ œ bœ œ m œ n œ œ
5 5
Bœ œ b œ œ ‰ .
178
& Bœ µœ
6 6
Fl.
5 œbœ œ bœ œ ⋲
f π P f p
3 5
3
œ Bœ µœ mœ œ bœ œ . µ œ . m œ œ œ bœ œ b œ mœ œ bœ œ µœ mœ
Bœ œ œ b œ Bœ 3 ‰ œ bœ
5
J œ œ bœ µ œ œ ‰.
178
& Bœ Bœ ‰ bœ œ bœ ‰
6
Ob.
f p p f p p f
7 4:5
5
œ . mœ œ bœ œ œµ œ m œ œ œ b œ œ m œ . œ œ b œ . œ.
œœ œ Bœ œBœ B œ Bœ œ bœ œ œ µœ µœ œ
178
bœ œ mœ nœ ⋲ ‰ 3
7
&
6
5
œ mœ ⋲ π
Cl.
œ
5
f ∏ π
5 7
6
π
6 7
bœ Bœ Bœ œ. bœ œ œ bœ œ µ œ µœ œ œ b œ œ œ . m œ . n œ œ
B œ mœ nœ J R J
178 3
Bsn. bœ œ ‰ 3
R bœ œ ⋲
p F p p f p
& œ. ?œ œœ bœ œ œ & ⋲ œJ . œ Bb œ .
178
bœ œ Œ œ Bb œ .
5
Bœ Bœ Bœ . R J
œ œ œ
Hn.
7
π p
Bœ . œœ bœ . œ œ mœ . nœ œ bœ . œ
& œ.
178
⋲ J œ bœ ‰
7 5
œ. œœ R œ
5
œ œ œ bœ œ bœ œ
7
œ œ
Tpt.
p p 5 5
f p p
µœ œ œ œ . œ œ
? œ BŒ J œ œ
178
j ‰ µœ œ œ.
œ. œ
Tbn.
p p P f
3
#œ #œ nœ œ bœ œ #œ nœ bœ
178
‰. œ #œ œ bœ œ bœ œ œ # œ œ bœ œ bœ œ
‰
5
& œ œ œ œ bœ œ œ bœ œ bœ
5 bœ œ bœ œ
p F p f
5
5:6
œœ
Perc
? œ œ œ bœ
œ ∑ ∑ œ ∑ ∑
p p
178 c6
& Ó ‰ œ b œ œ µ œ œ bœ Œ Œ ‰ œ bœ
3
œ mœ
15
6 7
œ µ œ µ œ œ bœ œ œ µœ œ bœ œ µœ
µ œ Bœ œ œ b œ mœ nœ œ µœ µœ œ µœ œ bœ œ mœ nœ œ
& B œ Bœ Œ Ó bœ œ µœ µ œ œ b œ
Synth
µœ œ m œ n œ µœ œ mœ nœ µœ
5
7 6
? Ó Bœ œ
5
œ Ó Ó œ‰ Ó
œ. 6
7
178
Comp. & ∑ ∑ ∑ ∑
& Ó ‰ œ b œ œ m œ œ bœ Œœ Œ ‰ œ bœ ∑
3
œ mœ
15
6 7
m œ m œ œ bœ œ m œ n œ œ mœ œ bœ œ mœ mœ œ bœ
mœ œ œ œ bœ p œ m œ n œ œ mœ m œ
& œ œ Œ Ó œ m œ m œ œ bœ
œ œbœ
° 5 F
Kbd.
œ œ F œ mœ nœ
7 6
m œ n œ mœ
5
? Ó œ œ œ Ó Ó œ‰ Ó mœ
œ. 7
p *°
6
q = 72
œ µœ œ œ œ bœ .
‰ œ bœ
ord.
œ µœ œ bœ œ µ œ µ œ œ b œ œ µœ œ bœ
J œ mœ œ µœ µœ
178 5
& œµ œ ‰
7
œ bœ œ mœ nœ
6
œ œ
Vn. 5
3 3
p f p p
3 3
π
.
Bœ ⋲ µœ œ. bœ . œ. µœ œ œ bœ œ bœ B
3 ord.
& œ B œ . Bœ . B œ Bœ mœ œ nœ œ µ œ µ œ œ bœ ⋲ ‰.
œ µ œ m œ n œ ⋲& J R
5 5
Vla.
œ bœ œ µ œ
π p 5
F 3
p p 5
Âœ µ œ œ bœ œ b œ
sul pont.
? Bœ œ µœ œ œ j mO . µB Oœ Oœ œ µœ m mOœ Oœ µ BOœ Oœ
r ⋲ ⋲ m œJ .
5 ord.
Oœ œ j‰
. R Oœ
œ
Vc.
œ
5
p p f 3 p p f
? Œ ‰ œ ˙ œ œ Œ ‰ & œJ œ. µœ . ?œ .
œ.
J
ø
Cb.
π f p F
3/16/08
q = 80
bœ œ œ µœ bœ œ œ µœ mœ
m œ Âœ n œ œ µ œ Âœ œ Bœ œ µ œ R œ mœ œ m œ œµœ Bœ
6
mœ œ
µœ bœ ‰ .
182 6
µ œ Bœ Bœ µ œ ⋲ ‰
5
Fl. & 3 œ Bœ µ œ œ
p f p f 4 p f
3
7
3
œ µ œ mœ œ œ 7
& œ Bœ R Âœ B œ œ œ œ ‰ m œ œ µ œ Bœ Bœ m œ œ µœ B œ
182 6:5
‰ bœ bœ œ ‰ Œ
4 Œ
7:5
Ob. œ µ œ Bœ b œ R µœ 8
p P f f
mœ œ œ µœ
œ œ bœ bœ œ bœ œ µœ J œ
5
mœ œ ⋲ mœ .
182
⋲ µœ
5
& µ œ Bœ b œ œ œ b œ Bœ B œ µ œ µ œ
8:5 5
œ bœ bœ œ œ bœ b œ œ m œ
Cl.
3
f f
3
œ bœ f
π p
3
‰ µœ . µœ bœ mœ ? œ µœ µœ mœ nœ œ µœ µœ bœ b œ Bœ œ µ œ µ œ b œ b œ B œ œ œ bœ
B Œ
182
mœ œ mœ mœ nœ œ ⋲
5
Bsn.
p f p p f
5
3 3
5
j
182
& œ ⋲ œ. Bb œ
3
bœ mœ œ. bœ . bœ œ b œ bœ
3 3
œ bœ œ œ bœ b œ ⋲
Hn.
f mœ nœ p 4 f 3
π
œæ
œ œ
& œ b œæ œæ
œ œ bœ mœ
182
œ œ œ. r ‰ œ
6
æ
3
œ œ mœ . œ. b œ bœ œ mœ nœ
R 4
æ 8
Tpt.
f p P f
µœ µœ bœ
5
B œ ?œ œ Bœ bœ m œ
182
Bœ œ bœ µœ . j ‰ Bœ bœ
µœ .
5
µœ Bœ
Tbn.
p p
3
b œ œ œ bœ
R œ œ œ œ œ #œ œ œ #œ
‰.
182 6
& Œ ⋲ Œ œ œ #œ œ bœ ‰ Œ
7 6
œ œ # œ bœ œœ œ
6
bœ œœœœ
p 5
f p p
? b œ œ œ # œ n œ œ b œ b œ bœ Œ bœ nœ
Perc
Œ Œ œ #œ nœ bœ œ œ b œ bœ
bœ ‰ Ó Œ bœ œ
6
7 7 6
‰. r 4 3
182
Ó Œ
c7 c8
& bœ
15
œ bœ œ œ
5
œ mœ œ œ µ œ œ Bœ mœ œ œ
& Œ m œ œ B œ Bœ
µ œ µ œ bœ µ œ µ œ µ œ µ œ Bœ µ œ Ó
5
5
Synth 3 4 8
µ œ µ œ m œ œ bœ
5
m œ Bœ µ œ œ b œ Bœ
7
? œ bœ ‰.
6
µ œ µ œ Bœ Œ Œ Bœ b œ Ó b œ œ bœ œ bœ
µ œ Bœ Bœ
œ 5
182
Comp. & ∑ ∑ ∑
& bœ Ó Œ ‰. r ∑
15
œ bœ œ œ
5
œ mœ œ œ mœ œ œ m œ œ œ
& Œ m œ m œ bœ m œ
5
p mœ œ œ œ œ mœ mœ œ
5
mœ
π
3
π
Kbd.
5
F
œ mœ mœ mœ nœ mœ œ bœ œ
7
? œ bœ œ bœ œ b œ ‰.
6
mœ mœ nœ Ó Ó b œ œ bœ œ bœ
mœ nœ œ
œ *°
*°
5
6
√
>
bœ œ œ œ µœ mœ œ
q = 80
sul E œ >œ
mœ œ m>œ
sul G
œ µ œ œ Bœ œ œ µœ Âœ nœ œ
182 5 5 sul A
& µ œ Bœ ‰ µœ µœ µ œ Âœ B œ µ œ
5
‰ œ J
µœ µœ µœ mœ bœ
Vn.
f p f p
7
p 4 3
œ œ œ œ
sul A
œ µœ µœ Bœ œ µœ Bœ µœ R
B µ œ Bœ Bœ µ œ m œ n œ Œ µœ Bœ bœ µœ ‰.
5 3
&
3
Vla.
œ bœ 4 œ bœ bœ 8
f p p f
5:6
œ µœ bœ œ. µœ .
? µœ œ Bœ µœ Âœ µœ mœ nœ mœ Âœ J µœ Bœ b œ Bœ µ œ œ b œ
b œ µ œ Bœ ⋲ ⋲ b œ Bœ b œ Bœ œ
6
Vc.
Bœ . . . µœ
.
5 5
p p f
3
bœ µœ œ b œ b œ Bœ b œ Bœ
? œ ? œ µœ µœ œ µœ 5 J œ. µ œ
&œ Bœ µœ ‰ . Bœ. B œ.
3
Cb. mœ nœ œ œ. 3
p f p
7 3
3/16/08
q = 100
>œ >œ µ>œ
Ÿ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ œ œ Bœ œ̆ b˘œ b˘œ m˘œ
œ̆ œ µœ œ Bœ
7
⋲ œ Bœ m œ
185
µ œ Bœ Bœ µ œ œ (mœ ) ˙ Ó
6
Fl. & œ œ J Bœ œ œ
p 3 Ï f 4
5
3 2
f œ
bœ œ µœ œ bœ bœ mœ œ
‰ . œ mœ
185
& 8 Bœ µ œ Œ ∑ ∑ µœ œ œ
Ob.
4 4 4
ƒ
7
6
b>œ µ>œ œ bœ bœ
r ‰. ⋲ Bœ Bœ œ œ œ Bœ mœ µœ œ œ œ
185
œ µ œ Bœ ‰ œ
3
& Bœ œ œ m œ
7
J Bœ m œ Bœ b œ mœ Bœ
3
> µ œ œ mœ
Cl.
œ f ƒ
p >œ Â œ
6
>œ µ œ œ > œ œ mœ µœ œ
? B µœ m œ B œ bœ ? ⋲ bœ œ œ œ mœ mœ Bœ œ b œ
185 6
mœ nœ bœ œ œ ‰ œ œ ⋲
6
bœ œ bœ œ œ œ mœ
œ œ >
Bsn.
f > > f 5
‰ ? b˙
185
& r ⋲ Œ Bœ bœ j Œ
3 œ œ  3 > bœ bœ >
Hn.
œ bœ bœ > œ
> > b>œ Í ƒ
2 4
p ƒ
œ̆ m˘œ n˘œ œ̆ ?
mœ œ œ œ
æ
185 5
‰ ∑
&8 æ
œ œæ
4 Œ œ ˙
Œ
3
bœ 4 > >
4
Tpt.
p ƒ p Í Ï
>œ µ>œ . m>œ . b>œ . >œ µ>˙
?
185
Bœ r ⋲ œ̆ Œ
5
> œ
> µ>œ œ
Tbn.
f f 3
ƒ Í ƒ
185
& ∑ ∑ ∑ ∑ ÷
f
œæ.
Perc
? ∑ ∑ ∑
185
3 2 c9
3 c10 4
&
15
mœ µœ Âœ ‰ œ œ
5
œ œ œ nœ Âœ œ Bœ œ œ Bœ Bœ œ œ
5
&8 ‰ 4 µœ mœ Bœ œ⋲ µœ œ µœ mœ Bœ 4 œ µœ œ µœ Œ ‰ 6 ‰
mœ µœ œ
Œ
4
Synth
m œ Bœ bœ œ m œ µ œ Bœ B œ œ t b œ Bœ œ b œ
7
? ‰. ‰ Œ œ Œ
5
µ œ œ µœ
7
œ µœ mœ œ Œ
7
Bœ œ µ œ œ 5
6
185
Comp. & ∑ ∑ ∑ ∑
& ∑ ∑ ∑ ∑
15
mœ nœ œ œ ‰ mœ nœ
5
œ œ mœ œ Œ œ œ œœ œ œ œ œ
5
& ‰ mœ œ œ œ œ mœ œ ⋲ mœ œ œ mœ œ œ ‰ 6 ‰ Œ
mœ nœ œ
f Ï
Kbd.
7
?
5
m œ œ bœ œ b œ
mœ nœ œ œ œ t œ œ bœ bœ
‰. ‰ Ó Œ œ Œ
7
œ mœ mœ œ Œ œ mœ
7
œœ *° mœ œ 5
6
q = 100
√>œ . >œ . m>œ .
sul D
>œ m>œ m>œ
sul G
>œ > r j
6
J J œ µ œ œ Bœ œ µ œ œ Bœ
185 sul E sul A
& œ œ ⋲ œ J J J ‰
mœ J œ œ œ m œ Bœ bœ b œ
6
mœ J
mœ
œ > mœ bœ œ bœ
Vn.
bœ
3 ƒ 2 ƒ 3 ƒ 3 4
>œ j
sul C
m>œ >œ . j œ.
sul G
µœ
sul D
œ œ. œ. œ Bœ œ
bœ ‰ ⋲ œJ B ⋲ ⋲ J mœ œ mœ
6
&8 4 R bœ œ> Bœ Bœ œ 4
4 > œ
Vla.
ƒ ƒ 5
ord. µ œ œ B œ
µ œ. µ œ. Bœ. b œ. ⋲ µ œ µ œ Bœ µœ œ µœ mœ
sul pont.
? ‰ œ œ ⋲ bœ œ œ œœ ⋲ B œ B œ œ bœ
6
Vc.
mœ mœ b œ Bœ œ b œ
œ
ƒ 5
œ œ bœ b œ Bœ œ b œæ µ œæ
pizz.
æ œæ
arco
? mœ
Cb.
œ.

Œ œ µœ œ
Œ æ æ ææ
ƒ
f
3
3/16/08
,
q = 69
jet
F whistle
‰ œj œ B˙
189
& j Œ ˙. ∑ ∑ Ó
mœ . mœ ˙ œ
ø ø ø ø
Fl.
4 Ï 3f 2 3 4
189
j
∑ ∑
&4 4Œ ⋲œ. œ
4˙ 4 ˙. 4 w
ø
Ob.
p F
189 ,^ Ÿ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ˙
& j ‰ Œ Ó ‰ j Œ Ó
ø
b œ ( bœ )˙ . . ˙ ˙
ø ø
Cl.
mœ
Ḟ
, Ï^
?
189
Ó ∑ ∑ ∑ ∑
m˙ ˙ ˙
ø
Bsn.
ƒ
?
189 ,^
Œ ∑ ∑ & Œ
˙ w
4 bƒœ ˙. ˙.
ø
Hn.
3 2 3 p 4
, j
? Œ B œj œ
189
œ ‰ Ó ∑ ∑ &4 Œ
(pedal)
4 ˙ ˙ œ
3
4 w
ø 4 4
Tpt.
ƒ p
?
,
‰ µœ ˙ w
189
j‰ ∑ ∑
J
m˙ ˙ ˙.
ø
œ
Tbn.
ƒ p
189
,
÷ Œ Ó ∑ ∑ ∑ ∑ ∑
bass drum
œ
>
Perc
f
189
4 3 2 3 4
&
15
&4 4 4 4 4
Synth
t
mw
189
Comp. & ∑ ∑ ∑ ∑ ∑ ∑
& ∑ ∑ ∑ ∑ ∑ ∑
15
& ∑ ∑ ∑ ∑ ∑ ∑
,ƒ
Kbd.
t ∑ ∑ ∑ ∑ ∑
mw
,
F q = 69
189
Ÿ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
sul pont.
≥
& j ‰ Œ Ó ∑ ‰j
mœ
3 mœ ˙ ( )ø 2˙ 3 ˙f . 4 m˙ . œ
Vn.
œ
4Ï
Ÿ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
sul pont.
≥
B ∑ ∑
4 ‰ œj ˙ (Bœ )
ø
Vla.
4 4˙ 4 ˙. 4 ˙. œ
f
,
?
sul pont. ord.
Œ j
œ œæ æ̇ æ̇. æ̇ b˙ œ. œ ˙
n˙ . œ œ
ø ø
Vc.
Ï p f ƒ
? , ≤ sul pont. ord.
œ œæ æ̇ æ̇. æ̇ j
m˙ . œ œ m˙ . œ. œ ˙
Cb.
ÿ p ƒ
Ï p f
3/16/08
q = 66
>
& œ ˙. ⋲ Bœ . w ˙.
195
∑ Ó Œ
ø J
Fl.
f 3 4 Íp 3 4
j > .
195
& œ ‰ Œ Ó ∑
4 Ó Œ bœ w
4˙
ø 4 4
Ob.
Íp
œ ˙. ‰. r
195
& ∑ Ó mœ œ w ˙.
ø >
Cl.
F Íp
?
195
m>œ œ w ˙.
Bsn. ∑ ∑ Ó ‰ J
Íp
195
⋲ j Ó ∑
air only
& ˙
ø ≠.
ø ≠
3 ƒ≠ ≠ ≠ µ˙ w
Hn.
p 4 Íp
3 4
m>œ >
® mœ œ
195 con sord.
Ó ∑ ∑
take mute
& ˙ nœ . ˙ œ. 4 nœ . ˙
ø
Tpt.
4 4 4
f P 5
≠air. only ≠ ≠ ≠ ≠
? ˙
195
⋲ J J ‰ µw w ∑
ø ø
Tbn.
ƒ p Íp
195
Perc ÷ ∑ ∑ ∑ ∑ ∑ &
3 4 bœ 3 4
µ m œœœœ
195
&
15
m b µ œœœœœ Â m œœœœ
& Ó
Synth
ø 4 4
F
4 4
t
195
Comp. & ∑ ∑ ∑ ∑ ∑
∑ ∑ bœ Œ Ó ∑ ∑
&
15
m m œœœœ
& ∑ ∑ ∑ ∑ ∑
F
Kbd.
t ∑ ∑ ∑ ∑ ∑
q = 66
Ȯ Ȯ Oœ µ µ wO
195
& Œ ∑ ∑
œ ø
Vn.
P 3p 4 p 3 4
B Œ µ µ Ȯ Ȯ Oœ wO ∑ ˙
4Œ
œ
ø 4 4 4
Vla.
P p p p
? j
∑ ‰ Bœ ˙ . w ˙.
b˙ . œ
Vc.
> Íp
p
? ∑ & wo ∑ ∑ ?
˙. œ
Cb.
> p
p
3/16/08
Â>œ ˙. ˙.
flz.
µ ˙æ. æ̇ æ̇
˙.
200
& Œ ‰ Œ
ø
Fl.
Íp ƒ
3
4 3 4
w> ˙. B˙> ˙ w
& 4 ˙.
200
Œ
Ob.
4 4
Íp Ï p
200
& ˙. œ . Âœ w ˙. œ œ µœ Œ Ó ∑
Cl.
> œ mœ œ
Íp mœ
f fl
Ï
?
200
˙ m>˙ w ˙. µœ mœ nœ œ mœ w
Bsn. m œ bœ œ Œ Œ
Íp fl p
f Ï
j
200
& Ó Œ µ˙ . œ ‰ ∑
Bœ w ˙.
>
Hn.
4 Íp 3 4 ƒ p
mœ .
200
&4 œ Œ Ó Ó Œ nœ 4 ˙ . ∑ Ó Œ ⋲ j
4 bœ .
Tpt.
p p
π
B>œ œ w ˙.
?
200
Tbn. Ó ‰ J j ‰ ∑
Íp >˙ . œ
Sf p
#œ
crotales
œ.
⋲J Œ
200
Perc & Œ Ó Œ œ Ó Œ ∑ ∑ ÷
p
200
4 µœ 3 4
& Bœ Bµ µ œœœ
15
Bœ
µ œœ Bœœœ
&4 4 4ƒ œ
F mÂm œœœ
Synth
t
b œœœ
200
Comp. & ∑ ∑ ∑ ∑ ∑
& m œœ Œ Ó ∑ ∑ ∑ ∑
15
œ
m œœ Œ Ó ∑ ∑ ∑ ∑
&
F
Kbd.
ƒ
t ∑ ∑ ∑ Œ Ó ∑
b œœœ
æ æO
200
µ µ wO w
Vn. & ∑ ∑ ∑
4 3 4 ƒ
œ > ˙. æ
wO
æ
wO
B Œ Œ ‰ b œJ w &4
4
ø 4
Vla.
Íp ƒ
æ æ
Vc.
? ∑ ∑ ∑ & bb wO wO
ƒ
> ≥
sul pont.
? ⋲ µ œJ . ˙ . w ˙.
ord.
Cb.
œœ œœ œœ ˙˙ ww
Íp ƒ F
3/16/08
æ̇.
205
& µœ w w œ Œ Ó ∑ ∑
ø ø
Fl.
F 2 4
œ œ ˙. w µw w w
205
& ∑
ø
Ob.
4 4
π
205
& ∑ Œ j ‰ ‰ j ∑
m˙ .
ø
œ ˙ w
ø
Cl.
w œ π
π P
? w œ œ ‰ Ó
205
Bsn.
J ∑ ∑ ∑ ∑
π
205
∑ ∑ ∑ Œ ∑ ∑
air only
Hn. & ≠ ≠ ≠
p ƒ 2 4
> B>œ B>œ .
œ Bœ ˙ .
205
& j 3 œ ∑ Ó Œ
œ . œ mœ œ œ. œ œ œ œ 4˙
5
˙ œ œ 4
Tpt.
3
F p p 5
≠ ≠ ≠
air only
?
205
Tbn. ∑ ∑ ∑ Œ ∑ ∑
p ƒ
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ >r
Ó. Ó. Ó ‰.
sand
205
Ó Ó
÷ Œ ∑ ∑
blocks
Perc œ
J ‰ Œ Ó &
π P sus cym. – brass mallet
π f
205
2 4
&
15
& 4 4
Synth
205
Comp. & ∑ ∑ ∑ ∑ ∑ ∑
& ∑ ∑ ∑ ∑ ∑ ∑
15
& ∑ ∑ ∑ ∑ ∑ ∑
Kbd.
t ∑ ∑ ∑ ∑ ∑ ∑
205
µ µ wO Ow
Vn. & ∑ ∑ ∑ ∑
p 2 4
O B˙ w
sul pont.
& w Ó ∑ ∑ ∑ B
ø
Vla.
4 4
p F
O Ow ?
Vc. & w ∑ ∑ ∑ ∑
p
.
? j ‰ œj Œ œ ⋲ÂœJ
sul pont. ord. pizz.
ww w w œ ‰ Œ Ó Ó Œ J
ø
Cb.
f p F
3
3/16/08
q = 60
œ̆ 12"
R ‰.
211
Fl. & Œ Ó ∑ ∑ ∑ ∑
4 S
12"
211
Ob. &4 ∑ ∑ ∑ ∑ ∑
12"
j
211 air only
& ∑ Œ ‰ ≠ ≠ ≠ ≠ Œ Ó ∑ ∑
ø ø
Cl.
f
12"
? r‰ .
211
Œ Ó ∑ ∑ ∑ ∑
œ
Bsn.
fl
211
S 12"
∑ Ó Ó ∑
air only
& ≠ ≠ ≠ ≠ ≠ ≠≠≠≠ ≠ ≠ ≠ ≠ ≠ ≠ ≠ ≠
ø ø ø
Hn.
4 fl fl fl fl fl fl fl fl fl fl fl fl
f f
j
12"
211
&4 œ ‰ Œ Ó Œ Ó ∑
air only
≠ ≠ ≠ ≠ ≠ ≠≠≠≠≠≠≠≠ ≠ ≠ ≠ ≠ ≠ ≠ ≠ ≠
ø ø˘ ˘ ˘ flf˘ fl˘ fl˘ fl˘ fl˘ fl˘ fl˘ fl˘ fl˘ fl˘ fl˘ fl˘ fl˘ fl˘ fl˘ øfl˘
Tpt.
f
≠ ≠ ≠≠≠≠≠≠≠≠≠≠≠≠ ≠ ≠ ≠ ≠ ≠ ≠ ≠ ≠
12"
?
211
∑ Ó Œ Ó ∑
air only
ø ø
Tbn.
f
crotales medium yarn sus. cym. 12"
œ
211
& Œ Ó ÷Ó Œ œ ∑ ∑ ∑
dome
Perc
f p
4 œ
12"
211
& Œ Ó
15
&4
fade out
Synth
Ï 12"
t
mœ
211 °
Comp. & ∑ ∑ ∑ ∑ ∑
12"
& ∑ ∑ ∑ ∑ ∑
15
& ∑ ∑ ∑ ∑ ∑
ƒ
Kbd.
t Œ Ó ∑ ∑ ∑ ∑
mœ
q = 60
h@ h@ h@ h@
bow directly on bridge 12"
211 (noise only)
& Ó Œ Ó ∑ ∑
h h h h
ø
Vn.
4 p
m h@h
@ h@h h@h
B Ó Œ Ó ∑ ∑
(noise only)
hh
4
ø
Vla.
hh@.. hh@..
? Ó Ó Œ Œ ∑ ∑
(noise only)
ø
Vc.
p
12"
?
arco
Œ Ó ∑ ∑
mw w œ
ø
Cb.
>
ƒ
March 14, 2008
New Haven, Connecticut
3/16/08

Klingbeil Dissertation Web

Uploaded by

Copyright:

Available Formats

Klingbeil Dissertation Web

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Klingbeil Dissertation Web

Uploaded by

Copyright:

Available Formats

Spectral Analysis, Editing, and Resynthesis:

Methods and Applications

Michael Kateley Klingbeil

Submitted in partial fulfillment of the

Spectral Analysis, Editing, and Resynthesis: Methods and Applications

Michael Kateley Klingbeil

This document describes the design, development, and applications of cross-platform

variation of the McAulay-Quatieri technique of peak interpolation and partial tracking.

limited undo/redo. Hundreds of simultaneous partials can be synthesized in real-time

1.1 Timbre and the Computer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Synthesis Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.3 Theory Versus Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1.3.1 Implementation History . . . . . . . . . . . . . . . . . . . . . . . . . . 10

1.3.2 Present Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.1 The Frequency Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.1.2 The Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.1.3 The Discrete Fourier Transform . . . . . . . . . . . . . . . . . . . . . 16

2.1.4 The Short-Time Fourier Transform . . . . . . . . . . . . . . . . . . . . 18

2.2 STFT Analysis and the Phase Vocoder . . . . . . . . . . . . . . . . . . . . . . 22

2.2.1 Putting Phase to Good Use . . . . . . . . . . . . . . . . . . . . . . . . 24

2.2.2 Resynthesis and Transformation . . . . . . . . . . . . . . . . . . . . . 25

2.2.3 Time Expansion/Compression . . . . . . . . . . . . . . . . . . . . . . 26

2.3 Spectral Envelopes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

2.4.1 STFT Frequency Interpolation . . . . . . . . . . . . . . . . . . . . . . 32

2.4.2 Partial Tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

2.4.3 Improved Partial Tracking . . . . . . . . . . . . . . . . . . . . . . . . . 37

2.4.5 Limitations and Extensions . . . . . . . . . . . . . . . . . . . . . . . . 41

3 Sinusoidal Analysis and Resynthesis 42

3.1 Analysis Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3.1.1 Analysis Window . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3.1.2 Amplitude Thresholds . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

3.2 Partial Tracking Using Linear Prediction . . . . . . . . . . . . . . . . . . . . 46

3.3.1 Inverse FFT Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

3.3.2 IFFT Synthesis Implementation . . . . . . . . . . . . . . . . . . . . . . 52

3.3.3 Oscillator Bank Resynthesis . . . . . . . . . . . . . . . . . . . . . . . . 53

3.3.4 Phase Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

3.4 Data Storage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

3.4.1 Time-span Frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

3.5 Transient Sharpening . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

3.6 Noise Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

3.6.1 Bandwidth Enhanced Sinusoids . . . . . . . . . . . . . . . . . . . . . 62

4 Editing and Transformation 64

4.1 Graphical User Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

4.2 Selection Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

4.2.1 Rule-based Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

4.3 Editing Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

4.4 Time Expansion and Contraction . . . . . . . . . . . . . . . . . . . . . . . . . 70

4.5 Frequency Modifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

5.1 Spectral Composition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

5.2 Sonic Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

5.2.1 Timbre Hybridization . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

5.2.2 Spectral Tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

5.3 Software Composition Environments . . . . . . . . . . . . . . . . . . . . . . 85

5.3.3 Common Music . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

5.4 Compositional Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

5.5 Performance Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

A Tear of the Clouds score 107

A.1 Program Note . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

A.2 Score . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108