Chapter 3

Sound modeling: source-based approaches

Federico Avanzini

Copyright ⃝ c 2005-2019 Federico Avanzini

except for paragraphs labeled as adapted from <reference>
This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license. To
view a copy of this license, visit http://creativecommons.org/licenses/by-nc-sa/3.0/, or send a letter to
Creative Commons, 171 2nd Street, Suite 300, San Francisco, California, 94105, USA.

3.1 Introduction
It was 1971 when Hiller and Ruiz envisioned the possibility of using numerical simulations of the wave
equation for sound synthesis applications.
[. . . ] This is a completely new approach to electronic sound synthesis insofar as the starting point is
the physical description of the vibrating object [. . . ]

A decade later McIntyre, Schumacher, and Woodhouse published their classic study on the use of non-
linear maps for modeling the generation of self-sustained oscillations in musical instruments.
[. . . ] a fast minicomputer could produce results at a cycle rate in the audible range. The result would
perhaps have some novelty: an electronic musical instrument based on a mathematical model of an
acoustic instrument [. . . ]

Today the algorithms described by these authors can be easily implemented in real-time on general-
purpose hardware, and it is common practice to use the term physical modeling to refer to sound mod-
eling techniques in which the synthesis algorithms are designed based on a description of the physical
phenomena involved in sound generation.
Direct sound representations, that are merely based on a description of the sound waveform, do not
contain information about the way the sound has been generated and processed by the surrounding en-
vironment before arriving to the listener’s ear. Sampling in time the sound signal does not assume any
underlying structure, or process, or generative model, in sound representation. The symbolic descrip-
tion is extremely poor, and as a consequence very little interaction with the sound representations is
allowed. Although signal processing techniques can provide meaningful modifications (e.g. pitch shift,
time stretching), sampling is basically a static, low-level description of sound.
3-2 Algorithms for Sound and Music Computing [v.February 2, 2019]

High level representations of sound signals are necessarily associated with some abstract paradigms
that underlie sound production. As we have seen previously, when trying to develop a taxonomy of
sound synthesis methods a first distinction can be traced between signal models and source models. Any
algorithm which is based on a description of the sound pressure signal and makes no assumptions on the
generation mechanisms belongs to the class of signal models. Additive synthesis is a good example of
a signal model: as already mentioned, one major drawback of this technique is its enormous number of
control parameters: at least one amplitude and one pitch envelopes have to be specified for each partial.
Moreover, the sound representation has not a strong semantic interpretation, since these parameters do
not have a high-level meaning. Subtractive synthesis with its source-filter structure provides in a sense a
more semantic description of sound: in certain cases the two blocks can be given a physical interpretation
in terms of an exciting action and a resonating object, respectively. As an example, in the case of LPC
based speech synthesis the broadband input signal can be interpreted as a glottal source signal, and the
shaping filter represents the action of the vocal tract. However, in many other cases this interpretation
does not hold, and the control parameters in the model (e.g., the filter coefficients) do not have a high-
level meaning.
Source models aim at describing the physical objects and interactions that have generated an acoustic
event rather than the acoustic signal itself. This modeling approach often gives rise to rather complex de-
scriptions, that can lead to computationally expensive numerical algorithms. Several modeling paradigms
and techniques are available in the literature for deriving efficient implementations of such descriptions,
including lumped/distributed modeling, waveguide structures, finite difference methods, and so on. The
following sections describe in detail a few of these approaches. Here it is worth discussing another as-
pect, i.e. that of control. A direct consequence of assuming a source-based approach is that the resulting
control parameters have a straightforward physical interpretation: typical parameters in the models are
associated with masses, hardness/softness characteristics, blowing pressures, lengths: such a semantic
representation can in principle allow more intuitive interaction.
Source-based sound modeling paradigms are often grouped into two broad categories, namely lumped
and distributed models. Generally speaking, distributed models are more often used for describing vibrat-
ing bodies or air volumes where forces and matter depend on (and propagate along) both time and space.
One-, two- and three-dimensional resonators (such as strings, bars, acoustical bores, membranes, plates,
rooms, etc.) can be treated as continuous distributed systems, and mathematically described by means
of Partial Differential Equations (PDEs). One of the most popular distributed modeling approaches is
waveguide modeling, which will be discussed in detailed in Sec. 3.4.
Although waveguides are extremely successful in modeling nearly elastic mediums, where the D’Alembert
equation or some of its generalizations hold, they are not equally good in dealing with systems where
these hypotesis are not met. As an example, oscillations in a bar are governed by the so called Euler-
Bernoulli equation, for which no traveling-waves schematization can be assumed. One possible approach
for dealing with such systems is using finite difference or finite elements methods. These time-domain
techniques are based on direct discretization of the PDEs and consequently have high computational
costs. On the other hand, when properly applied they provide stable and very accurate numerical sys-
tems.
As opposed to distributed models, lumped models are used when a physical system can be conve-
niently described without explicitly considering its extension in space. As an example, a mechanical
resonating body may be described in terms of ideal masses or rigid elements, connected to each other
with spring and dampers, and possibly non-linear elements. Similar considerations may apply to elec-
trical circuits and even to certain acoustic systems. The resulting models are naturally described in the
time domain, in terms of Ordinary Differential Equations (ODEs). Sec. 3.5 discusses lumped modeling
approaches, and includes an introduction to modal synthesis. Defining modal synthesis as a lumped mod-
eling approach may be questionable, since the modal formalism incorporates a “spatial” representation

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-3

(e.g. it is possible to inject a force in a specific point of a modal resonator, or to measure its displacement
in a specific point). On the other hand, representing a resonator as a combination of a finite number of
modes correspons to approximating the resonator as a mesh of point masses connected with strings and
dampers, and in this sense modal synthesis may be regarded as a lumped modeling approach.

3.2 Physical structures and models

3.2.1 Simple vibrating systems and normal modes
Sound is produced by mechanical, acoustical, or electrical vibrations that ultimately generate an acoustic
pressure signal that reaches our ear. In this section we review the most elementary oscillating systems
and their properties.

3.2.1.1 Oscillators
The simplest physical oscillating system is the damped second-order (or harmonic) oscillator. A generic
oscillator of this kind is described by the following linear differential equation:

ẍ + 2αẋ + ω02 x = uext (t), (3.1)

where uext is an external driving signal. The general solution of the homogeneous equation (i.e., Eq. (3.1)
with uext = 0) is given by
x(t) = a0 e−αt cos(ωr t + ϕ0 ), (3.2)
√
where ωr = ω02 − α2 . The parameters a0 and ϕ0 are uniquely determined by the initial conditions
x(0), ẋ(0). In particular the impulse response of the system corresponds to initial conditions x(0) = 0
and ẋ(0) = 1, and is given by h(t) = e−αt sin(ωr t)/ωr .
An electrical system representing a damped harmonic oscillator is the RLC circuit (Fig. 3.1(a)).

d2 i di 1
L 2
(t) + R + i = 0, (3.3)
dt dt C
where i is the current in the circuit, L, R, C are the inductance, resistance, and capacitance of in the
circuit, respectively. This is an equation of the form (3.1), with α = R/2L, ω02 = 1/LC. Therefore it
has a solution of the form (3.2).
In the mechanical case, an instance of damped harmonic oscillator is the mass-spring-damper system
depicted in Fig. 3.1(b):
mẍ(t) + rẋ(t) + kx(t) = 0, (3.4)
where x is the displacement signal, m, r, k are the mass, mechanical resistance, and spring stiffness.
Again this is an equation of the form (3.1), with α = r/2m, ω02 = k/m. Therefore it has a solution of
the form (3.2).
In certain situations, acoustic systems can also be described in terms of lumped elements that are
equivalent to resistance, capacitance, and inductance. The variables involved in this case are air-flow
(or volume velocity) u(t), measured in m3 /s, and acoustic pressure p(t), measured in Pa. When the
dimensions of an acoustic element are much less than the sound wavelength, then the acoustic pressure,
p can be assumed constant over the element. In this case, the acoustic behavior of the element is, at least
at low frequencies, very simple. In particular, the Helmholtz resonator (depicted in Fig. 3.1(c)) behaves
to a good degree of approximation as a second order oscillator. We analyze this systems in terms of three
main elements: the opening, the neck, and the cavity.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-4 Algorithms for Sound and Music Computing [v.February 2, 2019]

m x(t)

S
R L
k r
V u(t)
C
O
i(t) L
(a) (b) (c)

Figure 3.1: Second order electrical, mechanical, and acoustic oscillators; (a) a RLC circuit; (b) a
mass-spring-damper system; (c) a Helmholtz resonator.

Resistive phenomena are observed during the passage of acoustic airflow through the opening, due
to a pressure difference ∆pop (t): the flow behavior is dominated by viscous and thermal losses and it is
reasonably assumed to be in phase with the acoustic pressure. Therefore the relation ∆pop (t) = Ru(t)
holds at the opening where the constant R is termed fluid-dynamic resistance. Simple inertial behaviors
are observed in the cylindrical neck. The air mass inside this tube is m = ρair SL (ρair being the air
density, S the cross-sectional area, and L the length). If a pressure difference ∆ptube (t) is applied at the
tube ends, the enclosed air behaves like a lumped mass driven by the force ∆ptube , and Newton’s law
implies
∆ptube (t) = ρair SL · v̇(t), (3.5)
where the relation u(t) = Sv(t) has been used, and v(t) indicates particle velocity. Finally, the cavity
has an elastic behavior. Consider the volume V (t) of air inside the cavity: the contraction dV (t) caused
by a pressure difference ∆pcav (t) is such that −ρair c2 · dV /V = ∆pcav . As a consequence, a new air
volume −dV can enter the cavity. By definition, this equals the integral of u(t) over time, therefore
∫ t
V
−dV (t) = u(t′ )dt′ = ∆pcav (t).
0 ρair c2
∫ (3.6)
ρair S 2 c2 t ′ ′
S∆pcav (t) = − v(t )dt ,
V 0

which represent a linear spring with stiffness ρair S 2 c2 /V . Both the air mass in the tube and the resistance
at the opening impede the same flow u, and are therefore in a “series” connection. This flow u enters
the cavity, so that the the volume is in series with the other two. The resulting equation for the particle
displacement x is
ρair S 2 c2
(ρair SL) · ẍ(t) + Rẋ(t) + x(t) = 0. (3.7)
V
Again equation of the form (3.1). Therefore solution of the form (3.2).

3.2.1.2 Impedance
The examples in the previous section show that in a large class of systems it is possible to construct pairs
of variables (often defined as Kirchoff variables) with the property that their product has the dimensions
of power (Kg m2 /s3 ). In electrical systems such a pair of variables is given by (v, i), voltage and current.
Integro-differential relations can be found that relate these two variables, in particular three elementary

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-5

Electrical Mechanical Acoustic

Current i (A) Velocity v (m/s) Flow u (m3 /s)
Voltage v (V) Force f (N) Pressure p (Pa)
(Resistance) R ( ) (Damping) r ( ) (Opening) R ( )
Kg·m2 Kg Kg
(Capacitance) 1/sC s (Spring) k/s s (Cavity) ρair c2 /V s m4 ·s
(Inductance) s/L (Mass) m · s (Bore) ρair Ls/S

Table 3.1: Summary of analogies in electrical, mechanical and acoustical systems.

relations define the fundamental quantities resistance R, inductance L and capacitance C. In the Laplace
domain, the integro-differential equations are turned into simple algebraic relations:

1
V (s) = R · I(s), V (s) = sL · I(s), V (s) = I(s). (3.8)
sC

These are particular examples of a more general relation in linear electric circuits:

V (s) = Z(s)I(s), (3.9)

where the quantity Z(s) is called impedance of the circuit and is defined as the ratio between the Laplace
transforms of voltage and current intensity. The inverse of Z(s) is called admittance, and it is usually
denoted as Γ(s) = Z(s)−1 .
Similar considerations apply to mechanical systems. Force f (Kg m/s2 ) and velocity v (m/s) are the
mechanical Kirchhoff variables, since their product is a power. Again, the ratio of these two variables
in the Laplace domain is defined as (mechanical) impedance, and its inverse is the (mechanical) admit-
tance. In the mechanical oscillator described above we have already introduced the three mechanical
equivalents of resistance, capacitance and inductance. The direct proportionality f (t) = rv(t) defines
ideal linear viscous forces, and by comparison with the first of Eqs. (3.8) r can be regarded as a mechan-
ical resistance. The inertial mass m of a non-relativistic body is defined as the ratio between the total
force acting on it and its acceleration, i.e. f (t) = ma(t) = mv̇(t), or F (s) = msV (s) in the Laplace
domain, and by comparison with the second equation in (3.8) m can be regarded as a mechanical induc-
tance. Finally, in an∫ ideal linear spring the elastic force is proportional to the elongation of the spring:
f (t) = kx(t) = k 0 v(t′ )dt′ . or F (s) = k/s V (s) in the Laplace domain, and by comparison with
t

the third equation in (3.8) the stiffness k can be regarded as a mechanical capacitance. Therefore the
aggregate impedance Z(s) of a second-order mechanical oscillator is Z(s) = ms + k/s + r.
As far as acoustic systems are concerned, acoustic pressure p (Kg/ms2 ) and volume velocity u (m3 /s)
are the acoustic Kirchhoff variables, since their product is a power. Again, the ratio of these two variables
in the Laplace domain is defined as (acoustic) impedance, and its inverse is the (acoustic) admittance.
In the Helmoltz resonator described above we have already introduced the three acoustic equivalents
of resistance, capacitance and inductance. More precisely, fluid-dynamic resistance is associated to
viscous and thermal losses at narrow openings: p(t) = Ru(t). Fluid-dynamic inductance is associated
to short, open tubes: p(t) = ρair L/S · u̇(t), or P (s) = ρair Ls/S · U (s) in the Laplace ∫ domain.
Fluid-dynamic capacitance is associated with enclosed air volumes: p(t) = ρair c /V · u(t′ )dt′ , or
2

P (s) = ρair c2 /(V s) · U (s) in the Laplace domain.

Table 3.1 summarizes the main analogies between electrical, mechanical, and acoustic systems, that
we have discussed throughout this section.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-6 Algorithms for Sound and Music Computing [v.February 2, 2019]

3.2.1.3 Coupled oscillators and modal decomposition

We have examined above the behavior of a single second-order oscillator. A way of describing more
complex oscillating systems is to represent them as combinations of the simple elements described so
far. As an example an oscillating mechanical system (a string, a membrane, etc.) can be described in
terms of point masses coupled through linear springs and dampers. Therefore the system is described as
a set of coupled second-order differential equations.
It is known that, under general hypotheses, one can find a change of variables such that the set
of coupled equations is turned into a set of uncoupled second-order equations. In order to clarify this
concept, let us look at the following simple mechanical example in which two point masses are connected
to each other and to the “walls” through three springs:

mẍ1 (t) + kx1 (t) + k(x1 − x2 ) =0,

(3.10)
mẍ2 (t) + kx2 (t) + k(x2 − x1 ) =0.

If we introduce a suitable set of variables q1,2 in place of x1,2 , the above equations can be decoupled, or
diagonalized:

q̈1 (t) = − ω02 q1 (t),

(3.11)
q̈2 (t) = − 3ω02 q2 (t),

with q1 = x1 + x2 , q2 = x1 − x2 , ω02 = k/m. The normal modes qi (i = 1, 2) are uncoupled and the xi
are linear combinations of the qi .
This simple example can be extended to more complicated systems, composed N masses coupled
through springs and dampers. One can in general reformulate the system in terms of normal modes
of oscillation, and the oscillation of each point mass can be seen as a linear combination of N normal
modes, each of which obeys the equation of a second-order (damped) harmonic oscillator. We will return
on these concept in Sec. 3.5.2, when discussing modal synthesis.

3.2.2 Continuous vibrating systems and waves

In the previous section we have examined oscillating systems constructed with lumped elements (e.g.
resistances, capacitances, inductances, and their mechanical and acoustic counterparts), and are therefore
represent by a finite and discrete set of points in space (e.g. a set of point masses). In this section
we examine vibrating systems that are distributed continuously in space, and are therefore described
by partial differential equations involving both space and time, rather than set of ordinary differential
equations in time.

3.2.2.1 The one-dimensional D’Alembert equation

Vibrational phenomena in an ideal elastic medium are described by the D’Alembert equation, whose
one-dimensional version is written as
∂2y 1 ∂2y
(x, t) = (x, t). (3.12)
∂x2 c2 ∂t2
This equation holds, for instance, in an ideal string of length L, linear mass density µ and tension T .
In this case the variable x ∈ [0, L] stands for position √ along string length and y stands for transversal
displacement of the string. The constant c has the value T /µ and has the dimensions m/s of a velocity.
A full derivation of Eq. (3.12) for the ideal string can be found in many textbooks: roughly speaking, the

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-7

r
r φ θ
O
O
z φ

(a) (b)

Figure 3.2: Illustration of (a) cylindrical and (b) spherical coordinates.

two main assumptions are that (i) the infinitesimal string segment dx moves only in the vertical direction,
so that its acceleration can be computed using only the transverse component of the tension as the acting
force; and (ii) the amplitude of the vibrations is very small.
There are interesting cases where acoustic disturbances can be assumed to be one-dimensional up
to a reasonable approximation. Propagation of acoustic pressure in a cylindrical or in a conical tube
is an example. Using cylindrical coordinates (see Fig. 3.2(a)), one can show that for cylindrical bores
one-dimensional longitudinal pressure waves in the z direction are described using Eq. (3.12), with z
in place of x and with y representing acoustic pressure. Using spherical coordinates (see Fig. 3.2(b)),
one can show that for conical bores one-dimensional spherical pressure waves are described through the
equation ( )
1 ∂ 2 ∂R 1 ∂2R
r (r, t) = (r, t), (3.13)
r2 ∂r ∂r c2 ∂t2
in which R(r) represents acoustic pressure, and the Laplacian operator is expressed in spherical coordi-
( 2∂) ( ) ∂2
nates as ∇2 = r12 ∂r
∂ 1
r ∂r + r2 sin ∂
θ ∂θ
∂
sin θ ∂θ 1
+ r2 sin 2 θ ∂ϕ2 . Using the substitution R = R̃/r, it is
easily seen that Eq. (3.13) reduces to the one dimensional D’Alembert equation (3.12) for the variable
R̃.

3.2.2.2 Traveling wave solution

A fundamental property of Eq. (3.12) is that it describes propagation phenomena. This statement can by
proved by factoring the equation as follows:
( )( )
∂ 1 ∂ ∂ 1 ∂
− + y = 0. (3.14)
∂x c ∂t ∂x c ∂t
From this factorization it is easily seen that generic solutions take the form

y(x, t) = y + (ct − x) + y − (ct + x). (3.15)

This is the solution to Eq. (3.12) originally proposed by D’Alembert himself. The two functions y ±
describe waveforms that translate rigidly with velocity c, in the right-going and left-going directions,
respectively. Their shape is determined by the boundary conditions (in space) and the initial conditions
(in time). As an example, if y represents the displacement of a vibrating string the initial conditions are
represented by an initial displacement and an initial velocity:
∂y
y0 (x) = y(x, 0), v0 (x) = (x, 0). (3.16)
∂t

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-8 Algorithms for Sound and Music Computing [v.February 2, 2019]

t t
(a) (b)

Figure 3.3: Boundary conditions and wave reflections; (a) fixed string end and negative wave reflection;
(b) free string end and positive wave reflection.

Boundary conditions impose constraints on the solution at the boundary of its domain. As an example,
if y represents the displacement of a vibrating string boundary conditions impose values for y and its
derivatives at the boundary points x = 0 and x = L. The two most common boundary conditions for
a string are the fixed end condition and the free end condition, which read as follows for the boundary
point x = 0 (similar equations can be written for the boundary point x = L)

Fixed end: y(x, t)|x=0 = 0, ⇒ y + (ct) = −y − (ct);

∂y (3.17)
Free end: (x, t) = 0, ⇒ y + (ct) = y − (ct).
∂x x=0

These equations show that boundary conditions imply “reflection” conditions on the traveling waves y ±
(see Fig. 3.3).

3.2.2.3 Waves and modes

A different analysis of the wave equation was proposed by Fourier, who proved that the general solution
to Eq. (3.12) can be regarded as a superposition of a numerable set of so-called stationary waves.
We exemplify the Fourier analysis in the case of an ideal string with fixed-end boundary conditions
y(x, t)|x=0,L ≡ 0. We search for particular solutions y(x, t) = s(x)q(t), which we call stationary waves,
since they have a shape in space that is determined by the spatial function s(x) and is modulated in time
by the temporal function q(t).
By substituting the generic stationary wave solution into Eq. (3.12), one finds that the functions s
and q must satisfy suitable differential equations:

s′′ 1 q̈
(x) = 2 (t) ⇒ s′′ (x) = αs(x), q̈(t) = c2 αq(t), (3.18)
s c q

for some α ∈ R. This last equation follows from the fact that s′′ /s is a function of space only, while q̈/q
is a function of time only. Therefore these ratios must necessarily equal to a constant α.
Now look at the spatial equation. In order for the boundary conditions to be satisfied s has necessarily
to be a non-monotonic function and consequently the condition α < 0 must hold, so that s obeys the

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-9

equation of a second-order oscillator (otherwise s(x) would be an exponential function). Moreover, since
it has to be s(0) = s(L) = 0, only a numerable set of spatial frequencies are allowed for s:
√
2 nπ
s(x) = sin(kn x), with kn = , (3.19)
L L
√
where 2/L is just a normalization factor.
Once the spatial equation has been solved, the temporal equation gives
nπc
q(t) = A sin(ωn t + ϕ), with ωn = ckn = , (3.20)
L
where A and ϕ depend on initial conditions. Again, only a numerable set of temporal frequencies ωn =
ckn are allowed. Spatial and temporal frequencies are proportional to each other through the constant c.
In conclusion we have obtained the following stationary waves, or normal modes:
√
2
yn (x, t) = sin(ωn t + ϕn ) sin(kn x). (3.21)
L
The general solution to Eq. (3.12) can be expressed as a linear combination of these modes:

∑
+∞
y(x, t) = An yn (x, t), (3.22)
n=1

where An , ϕn are determined by the initial conditions. This latter equation re-states what we already
know: a periodic signal, such as the one generated in an ideal string with ideal boundary conditions, can
be expressed as a series of harmonically-related sinusoidal signals.
Note that the Fourier solution, expressed in term of normal modes, and the D’Alembert solution,
expressed in terms of traveling waves, are equivalent. In fact a standing wave yn (x, t) can be viewed
as a superposition of sinusoidal traveling waves. More precisely, using the Werner formulas1 a standing
wave can be written as
√
1
yn (x, t) = {cos[kn (ct − x) + ϕn ] − cos[kn (ct + x) + ϕn ]} . (3.23)
2L

Therefore a standing wave is the sum of two sinusoidal waves y ± that translate rigidly with velocity c, in
the right-going and left-going directions, respectively. This proves the equivalence of the D’Alemebert
and Fourier solutions.
Note however that normal-mode solutions are more general than traveling-wave solutions: already
a simple system like a one-dimensional bar, described by a 4th order PDE, does not admit a solution in
terms of traveling waves while its normal modes can be written analytically.

3.3 Delays and oscillations

3.3.1 The Karplus-Strong algorithm
This section reviews a sound synthesis algorithm which is relevant from many viewpoints. First, the
Karplus-Strong (KS hereafter) algorithm is a famous one and deserves to be studied. Second, it contains
many of the basic elements that are needed to provide a clear picture of what waveguide modeling is
1
2 sin α sin β = cos(α − β) − cos(α + β).

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-10 Algorithms for Sound and Music Computing [v.February 2, 2019]

25

20

Magnitude resp. (dB)

15

10

x[n] y[n]
0

g −5

−10

−15
0 0.5 1 1.5 2 2.5 3
z −m ωd (rad)

(a) (b)

Figure 3.4: A comb filter; (a) block scheme and (b) magnitude response.

all about, and yet it is structurally simple enough to be discussed in a limited amount of pages. Finally,
from a historical perspective it can be regarded as the first prototype of a waveguide approach to string
modeling: it is true that the original formulation of the algorithm did not contain any physical interpre-
tation. What is unquestionable, however, is that the KS algorithm is structurally identical to the simplest
waveguide models that we are going to examine in the next sections.

3.3.1.1 The comb filter

The basic computational structure underlying the KS algorithm is the IIR comb filter, which is represented
by the following difference equation (and transfer function):
1
y[n] = x[n] + gy[n − m], ⇒ H(z) = , (3.24)
1 − gz −m
with g ∈ R. The block structure of the filter is depicted in Fig. 3.4(a). The poles of H(z) are found from
√
z m = g. Therefore the filter has m poles pl = m gej2lπ/m for l = 0, . . . m − 1, equally spaced around
√
the circle of radius m g. In order for the filter to be stable, the condition | g | < 1 must be satisfied.
The corresponding magnitude response is plotted in Fig. 3.4(b). Each pole pl produces a peak in the
magnitude response. We can apply the analysis seen in Chapter Sound modeling: signal based approaches for the
second-order resonating filter in order to understand the relation between the poles pl and the peaks in the
response: in general, as g increases and grows closer to 1, each peak becomes higher and the associated
bandwidth narrows down. Note also that the filter produces a harmonic spectrum in which frequency
peaks are integer multiples of the “fundamental” frequency f0 = Fs /m Hz.
Figure 3.4(a) already provides us with an intuitive proto-physical interpretation: a disturbance (a
wave) in a medium is propagated through that medium, is confined within a certain length, bounces
back and forth due to some boundary conditions, has some energy dissipated at each bounce through the
coefficient g. Note that if the sign of the wave is inverted at each reflection, the resulting filter spectrum
is affected:
1
y[n] = x[n] − gy[n − m] ⇒ H(z) = (3.25)
1 + gz −m
√
In this case the poles are pl = m gej(2l+1)π/m for l = 0, . . . m − 1. This means that the corresponding
frequency peaks have all been shifted by an angle π/m with respect to the previous case: now the fre-
quency peaks are odd integer multiples of the “fundamental” frequency f0 = Fs /(2m) Hz. Section 3.4.3

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-11

will show that choosing a sign or another corresponds to describing two different boundary conditions
(e.g., an open termination versus a closed termination in an acoustical bore).

M-3.1
Write a function that computes the output of the comb filter of Fig. 3.4, given a desired fundamental frequency f0
and a factor g < 1.

M-3.1 Solution
function y = ks_simplecomb(f0,g);

global Fs;
m = round(Fs/f0); %length of the delay line
d= dline_init(m); % create a delay-line object

x=((rand(1, ceil(m))*2) -1)/2;% define random input vector of length m

x=[x zeros(1,round(-3*m/log10(g)) )]; % zero-pad x to hear sound tail
y=zeros(1, length(x)); % initialize output signal

for n = 1:length(x) % audio cycle

y(n) = x(n) + d.y*g; %read from delay line and update output
d=dline_compute(d); %update delay line
d.x = y(n);
end

The input signal x is defined in accordance to the KS algorithm specifications (see next section). Zero
padding of x is chosen in such a way that the ouput signal has time to decay by 60 dB (see Chapter
Sound in space). Matlab/Octave are very inefficient at computing long cycles, but we use this approach
for coherence with next examples; in particular we have used two auxiliary functions that initialize a
delay line structure

function f = dline_init(d); %initialize a dline structure of length d

% x is the current input value written into the line
% y=x(n-d) is the current output value read from the line
% in is a buffer containing d past input values

f.x = 0; f.y = 0;
if(floor(d) == d) f.d = d; % ok, d is a valid integer delay
else error(’Not a valid delay’);
end
f.in = zeros(1, d); % create buffer for past input values

and update the state of a delay line structure

function f = dline_compute(f);

f.y = f.in(1); % output is the first sample in the buffer

f.in = [f.in(2:length(f.in)), f.x]; % update buffer

3.3.1.2 Synthesis of plucked strings

The above observations suggest that the comb structure (3.24) may be employed to synthesize harmonic
sounds, such as those produces by an ideal string. However, in order to obtain a complete formulation of
the KS algorithm we still have to add some refinements to the structure. Specifically, what it is missing
is a mean to control the spectral tilt of the filter magnitude response (i.e. the rate at which the response
decays with increasing frequency), and to account for different decay rates for the sound partials.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-12 Algorithms for Sound and Music Computing [v.February 2, 2019]

4000

3500

3000

2500

f (Hz)
2000

1500

1000

500

0
0 1 2 3 4 5 6 7 8
t (s)

Figure 3.5: . Spectrogram of a plucked A2 guitar string. Note the harmonic structure and the decay
rates, which increases with increasing frequency.

In the real world a nylon guitar string is one of the closest relative of an ideal string and exhibits an
almost perfectly harmonic spectrum. Figure 3.5 shows the spectrogram of a a plucked guitar string: as
expected, a harmonic spectrum can be observed. However another relevant feature is that each harmonic
partial decays at a different rate, with lower partials surviving longer than higher partials.
On the other hand we have just seen that the IIR comb filter produces a spectrum in which all har-
monic peaks have the same magnitude, which means that the associated partials all decay in time at the
same rate. In order to simulate a frequency-dependent decay, one can insert a low-pass filter Hlp into the
feedback loop, as shown in Fig. 3.6(a): we call this structure a low-pass comb filter. Intuitively, at each
passage the high-frequency components are attenuated more strongly than low-frequencies components.
The simplest low-pass filter that can be employed is the first-order FIR already examined in Chapter
Fundamentals of digital audio processing:

1 1[ ]
y[n] = [x[n] + x[n − 1]] ⇒ Hlp (z) = 1 + z −1 . (3.26)
2 2
Figure 3.6(b) shows the frequency response of the low-pass comb structure after the insertion of Hlp :
as expected higher resonances are less peaked and have larger bandwidths, because now the filter poles
have frequency-dependent magnitudes.
However the insertion on a low-pass filter in the structure has also a second effect: it introduces an
additional half-sample delay, which can be observed if one looks at the phase response of Hlp (z) and is
qualitatively explained by the fact that this is filter averages the current sample with the previous one. A
consequence of this additional delay is that the fundamental frequency generated by the low-pass comb
structure is now f0 = Fs /(m + 1/2) Hz. Moreover, a closer analysis would also show that the upper
partials are not anymore integer multiples of f0 = Fs /(m + 1/2), due to the insertion of Hlp in the loop.
These deviations from the harmonic series can also be noticed from the plot in Fig. 3.6(b).
In many cases the deviations introduced by the low-pass filter are very small, especially for the lower
partials and for values of g that are close to 1. However they can still be perceivable. As an example, if
Fs = 44.1 kHz and m = 100, then a half sample delay corresponds to a delay in the order of 10−5 s: in
this case the IIR comb produces a fundamental at Fs /m = 441 Hz, while the low-pass comb produces a
fundamental at Fs /(m + 1/2) ∼ 439 Hz.

M-3.2
Find the response of the complete system given in Fig. 3.6 and plot magnitude and phase responses for various
values of g and m.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-13

25

20

Magnitude resp. (dB)

15

10

Hlp 5
x[n] y[n]
0

−5
g
−10

−15
z −m
0 0.5 1 1.5 2 2.5 3
ω (rad)
d

(a) (b)

Figure 3.6: Low-pass comb filter obtained through insertion of a low-pass element into the comb struc-
ture; (a) block scheme and (b) frequency response (the triangles mark the harmonic series lπ/L, l ∈ N).

The low-pass comb structure discussed so far is the core of the KS algorithm. However we have
not yet discussed what is the input signal to fed to the filter in order to obtain an output sound. Since
the impulse response of the filter is the signal that is resemblant of a plucked string sound, an obvious
choice is to inject the filter with an impulse. A second possible choice, originally suggested by Karplus
and Strong, is to impose a random initial state (m past values of y) to the filter: although this choice
has hardly any physical interpretation,2 it has the benefit of providing significant initial excitation in the
high-frequency region, with a consequent perceptual effect of an initial noisy transient followed by a
harmonic steady-state signal.

M-3.3
Write a function that implements the KS algorithm using the low-pass comb of Fig. 3.6, given a desired funda-
mental frequency f0 and a factor g < 1.

M-3.1 Solution
function y = ks_lpcomb(f0,g);

global Fs;
m = round(Fs/f0); %length of the delay line
d= dline_init(m); % create a delay-line object

x=((rand(1, ceil(m))*2) -1)/2;% define random input vector of length m

x=[x zeros(1,round(-3*m/log10(g)) )]; % zero-pad x to hear sound tail
y=zeros(1, length(x)); % initialize output signal
a_past=0; %initialize auxiliary variable (input to lowpass filter)

for n = 1:length(x) % audio cycle

a = x(n) + d.y*g; %read from delay line and sum to input
y(n) = 1/2 * (a + a_past); %update output through lowpass filter
a_past = a; % update auxiliary variable
d=dline_compute(d); %update delay line
d.x = y(n);
end

2
It would be like imposing initial random displacements to points of a string, as we shall see in the next sections.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-14 Algorithms for Sound and Music Computing [v.February 2, 2019]

3.3.2 Fine tuning and fractional delays

Sound pitch (which we assume to coincide with fundamental frequency)3 in the KS algorithm is quan-
tized: adding a unit delay in the comb filter modifies the fundamental period by 1/Fs , which is a rather
gross and perceivable quantization. In order to obtain a finer tuning of the delay loop, we need techniques
to simulate fractional delays.
An ideal delay of m samples is a filter with transfer function Hm (z) = z −m . Therefore its frequency,
magnitude, and phase responses are
( ) ( ) [ ( )]
Hm ejωd = e−jωd m , Hm ejωd ≡ 1, arg Hm ejωd = −mωd . (3.27)

We want to design a filter with the same characteristics, i.e. flat magnitude response and linear phase
response (equivalently, with constant and coincident phase and group delays). However we want the
slope of the phase response to be an arbitrary phase delay τph , and not limited to integer values m.
Moreover, since any real delay τph can be written as the sum of an integer delay ⌊τph ⌋ and and a fractional
delay 0 ≤ (τph −⌊τph ⌋) < 1, without loss of generality we restrict our attention to the design of fractional-
delay filters Hτph with 0 ≤ τph < 1).
Note that the impulse response of an ideal delay filter is
∫ +π
1
hτph [n] = e−jωd τph ejωd n dωd = sinc(n − τph ). (3.28)
2π −π

If τph = m ∈ N this reduces to h[n] = δ[n − m]. However, if τph is not integer then this is a non-
causal filter with infinite impulse response, i.e. a non-realizable filter. This remark makes it clear that
we will not be able to find exact realizations of fractional-delay filters, and we will have to look for
approximations.

3.3.2.1 FIR fractional delay filters

We first examine FIR fractional-delay filters, of the form

∑
N
Hτph (z) = bk z −k . (3.29)
k=0

Starting from this general form, we have to design of an N th order FIR filter approximating a constant
magnitude and linear phase frequency response. Several criteria can be adopted to drive this approxima-
tion problem. One approach amounts to minimizing some error distance between the FIR filter (3.29) and
the ideal fractional-delay filter defined previously. Possibly the most intuitive realization of this approach
is the minimization of the least squared (LS) error function, defined as the L2 norm the error frequency
response E(ejωd ) = Hτph (ejωd ) − e−jτph (i.e. E is the difference between the frequency responses of the
FIR filter and the ideal fractional-delay filter).
A different approach, that we describe in some more details, amounts to setting the error function
E(ejωd ) and its N derivatives to zero at ωd = 0:

dl E jωd
(e ) , l = 0, . . . N. (3.30)
dωdl ωd =0

3
In chapter Auditory based processing we will see that pitch perception is a complex phenomenon, and that the perceived pitch
does not necessarly coincide with the fundamental frequency.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-15

1 1

phase delay (samples)

ampl. response

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
ωd (rad) ωd (rad)

(a) (b)

Figure 3.7: Linear interpolation filters (N = 1) for τph = 0, 0.1, . . . , 1; (a) amplitude response and (b)
phase delay.

This is called the maximally flat design at ωd = 0, since it tries to make the error function as flat as
possible around the value 0, in the vicinity of zero frequency. Substituting Eq. (3.29) in these latter
N + 1 equations yields

∑
N
k l bk = τph
l
, l = 0, . . . N ⇔ V b = τ, (3.31)
k=0

N ], and V is a Vandermonde matrix with elements v

where b = [b0 , b1 , . . . , bN ], τ = [1, τph , . . . τph i,j =
(j − 1) . Since V is non-singular, the system has a unique solution which can be written in explicit
i−1

form as
∏N
τph − l
bk = , k = 0, . . . N. (3.32)
k−l
l̸=k;l=0

It is interesting to notice that the FIR filter coefficients obtained by this method are equal to those of the
Lagrange interpolation formula for equally spaced abscissas. In other words, the FIR filter determined
by these coefficients estimates the value x[n − τph ] by interpolating a polynomial of order N over the
N + 1 values x[n], x[n − 1], . . . x[n − N ]. This leads to Lagrange interpolation.4 For N = 1 one obtains
simple linear interpolation, b0 = 1 − τph , b1 = τph . For the case τph = 1/2 we reobtain the first-order
FIR low-pass filter.
Plots for N = 1 and different values of τph are shown in Fig. 3.7. The phase delay remains reasonably
constant up to high frequency values (and is exactly constant in the cases τph = 0, 1/2, 1). Note however
that the magnitude response has always a low-pass character. This is a drawback of these FIR filters:
high frequencies are attenuated due to non flat magnitude response. Using higher orders N allows to
keep the magnitude response close to unity and a phase response close to linear in a wider frequency
band. Of course, this is paid in terms of computational complexity.

M-3.4
Implement a fractional delay line using Lagrange interpolation.

M-3.4 Solution
4
We are not interested here in deriving the Lagrange interpolation method, which is reviewed in many textbooks of numerical
analysis.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-16 Algorithms for Sound and Music Computing [v.February 2, 2019]

Same approach as before. One function to initialize the line

function f = lagrangedline_init(d,N); %uses Nth order lagrange interpolation

f.x = 0; f.y = 0;
f.d = d; % set delay (not necessarily integer)
f.in = zeros(1, floor(d)+2); % create buffer for past input values
f.b=ones(1,N+1); %coefficients of the Lagrange interpolator
tau=d-floor(d); %fractional delay to be simulated
for k=1:length(f.b)
for l=1:length(f.b)
if (l˜=k); f.b(k)=f.b(k)*(tau-(l-1))/( (k-1)-(l-1)); end
end
end

and one to update the state

function f = lagrangedline_compute(f);

f.y = f.b * f.in(1:length(f.b))’; % output is lagrange interpolation of buffer

f.in = [f.in(2:length(f.in)), f.x]; % update buffer

These functions can be tested in the KS algorithm (examples M-3.1 and M-3.3) in place of the integer
delay lines.

3.3.2.2 All-pass fractional delay filters

We now examine IIR fractional-delay filters, of the form

z −N A(z) an + an−1 z −1 + . . . + a1 z −(N −1) + z −N

Hτph (z) = = . (3.33)
A(z −1 ) 1 + a1 z −1 + . . . + aN −1 z −(N −1) + an z −N

This is not the transfer function of a generic IIR filter. It represents the transfer function of a N th order
all-pass filter. According to the definition already given in Chapter Fundamentals of digital audio processing, an
all-pass filter is a filter with a perfectly flat magnitude response. The filter in the above equation satisfies
the property Hτph (ejωd ) ≡ 1 by construction: this property can be proved by noting that, since the
numerator polynomial is a mirrored version of the denominator polynomial A, the poles of a stable all-
pass filter are located inside the unit circle and its zeros are located outside the unit circle with the same
angle and with the inverse radius of the corresponding poles.
Since the above IIR filter satisfies by construction one of the two properties of an ideal delay filter
(flat magnitude response), we can now focus on the second one (linear phase response). The phase
response of an all-pass filter is found to be
[ ] [∑ ]
[ ( jω )] 1 N
k=0 ak sin(kω d )
arg Hτph e d = N ωd + 2 arg = N ωd + 2 arctan ∑N . (3.34)
A (e−jωd ) k=0 ak cos(kωd )

Therefore the phase response, the phase delay, and the group delay are all highly non-linear functions
of the filter coefficients. This means that one cannot expect as simple design formulas for the all-pass
filter coefficients as for FIR filters. Instead, one can almost exclusively find only iterative optimization
techniques for minimization of traditional error criteria.
Possibly the only design technique that has a closed-form solution is the maximally flat group delay
design. Let us start considering an all-pole low-pass filter with transfer function 1/A(z −1 ). It has been

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-17

0
1

phase delay (samples)

−0.5
phase response

0.8
−1
0.6
−1.5

−2 0.4

−2.5 0.2

−3
0
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
ωd (rad) ωd (rad)

(a) (b)

Figure 3.8: First-order Thiran allpass filters for τph = 0, 0.1, . . . , 1; (a) phase response and (b) phase
delay.

shown that the condition of maximally flat group delay at ωd = 0 for this filter yields the following
analytic solution:
( )∏N
k N 2τph + l
ak = (−1) , (3.35)
k 2τph + k + l
l=0
(N )
where k is the binomial coefficient. When τph > 0 then the filter is stable. This result can be applied to
our problem using Eq. (3.34): since the fractional phase delay of Hτph is twice those of 1/A, a maximally
flat all-pass filter with coefficients
( )∏N
k N τph + l
ak = (−1) , (3.36)
k τph + k + l
l=0

approximates the ideal delay filter with total delay N + τph . This is known as Thiran all-pass filter
approximation.
As an example let us look at the first-order all-pass filter
a1 + z −1
Hτph (z) = , (3.37)
1 + a1 z −1
with a1 < 1 for stability. The plots of its phase response and phase delay are shown in Fig. 3.8. In the
low-frequency region, the phase response can be approximated as follows:
[ ( )] sin ωd a1 sin ωd 1 − a1
arg Hτph ejωd ∼ − + ∼ −ωd , (3.38)
a1 + cos(ωd ) 1 + a1 cos ωd 1 + a1
i.e. the phase response is approximately linear with phase and group delay approximately equal to
(1 − a1 )/(1 + a1 ). Therefore given a desired phase delay τph one chooses
1 − τph
a1 = . (3.39)
1 + τph
This corresponds to the Thiran approximation with N = 1.
Thiran filters have complementary drawbacks with respect to Lagrange filters: although they provide
flat magnitude response, detuning of higher frequencies occurs due to phase non-linearity. In order to
have phase response approximately linear in a wider frequency range one has to use higher orders, at the
expense of higher complexities.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-18 Algorithms for Sound and Music Computing [v.February 2, 2019]

M-3.5
Implement a fractional delay line using Thiran filters.

M-3.5 Solution
Same approach as before. One function to initialize the line

function f = thirandline_init(d, N); %uses a Nth order Thiran filter

f.x = 0; f.y = 0;
f.d = d; % set delay (not necessarily integer)
f.in = zeros(1, floor(d)-N); % create buffer for past input values
% the Thiran filter account for the remaining N+(d-floor(d))
f.state=zeros(1,N); %state of the Thiran filter
f.a = zeros(1, N+1); % coefficients of the Thiran filter
tau = d-floor(d); %fractional delay to be simulated
for k = 0:N
ak = 1; for l=0:N; ak = ak * (tau+l)/(tau+k+l); end
f.a(k+1) = (-1)ˆk * nchoosek(N,k) * ak;
end

and one to update the state

function f = thirandline_compute(f);

[out,state] = filter(fliplr(f.a), f.a, f.in(1), f.state);

f.state=state;
f.y=out;
f.in = [f.in(2:length(f.in)), f.x];

These functions can be tested in the KS algorithm (examples M-3.1 and M-3.3) in place of the integer
delay lines.

3.3.2.3 Time-varying delays

3.4 Distributed models: the waveguide approach

This section introduces the basic concepts of waveguide modeling. Discussion is focused on one-
dimensional resonators, and no attention is devoted here to higher dimensional waveguide structures.
In their simplest form, waveguide models exploit the existence of an analytical solution to the
D’Alembert wave equation, which can be seen as a superposition of traveling waves (rigidly translat-
ing waveforms). Such a solution can be simulated in the discrete space-temporal domain using delay
lines, and the resulting numerical algorithms are extremely efficient and accurate. Moreover, physical
phenomena such as frequency dependent losses and dispersion can be included in the models by incorpo-
rating low-pass and all-pass filters in the delay line scheme. Again, careful design of such filters allows
for very accurate and relatively low-cost simulations.

3.4.1 Basic waveguide structures

3.4.1.1 Wave variables and wave impedance

So far, only displacement y (for a string) and acoustic pressure p (for a cylindrical bore) have been con-
sidered in the wave equation. However, alternative wave variables can be used in strings and acoustical

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-19

bores. As an example, the force acting on a string section dx is defined as

[ + ]
∂y ∂y ∂y − T [ ]
f (x, t) = −T (x, t) = −T (ct − x) + (ct + x) = ẏ + (ct − x) − ẏ − (ct + x) . (3.40)
∂x ∂x ∂x c

Therefore, using this equation force waves f ± can be defined as f ± := ∓ Tc ẏ ± . On the other hand, the
transversal velocity in the same string is given by

∂y
v(x, t) = (x, t) = ẏ + (ct − x) + ẏ − (ct + x). (3.41)
∂t
From this, velocity waves v ± are defined as v ± := ẏ ± . As we have seen in Sec. 3.2.1, the force-
velocity variable pair represent the mechanical Kirchhoff variables, in analogy with voltage and current
in electrical systems. From the previous equations it immediately follows that
√
f ± (ct ∓ x) = ±Z0 v ± (ct ∓ x), with Z0 = T /c = T µ. (3.42)

The quantity Z0 takes the name of wave (or characteristic) impedance of the string, and its reciprocal
Γ0 = Z0−1 is termed wave admittance. Note that using Z0 both the force f and the velocity v can be
related to the force waves f ± . Namely, the following relations hold:

1 [ + ]
f = f + + f −, v= f − f− ,
Z0
(3.43)
+ f + Z0 v − f − Z0 v
f = , f = ,
2 2
that transform the pair (f, v) into the pair (f + , f − ), and vice versa.
Wave impedance can be defined also in a cylindrical bore. In this case the Kirchhoff variables are
taken to be pressure p and flow u (volume velocity). These can be related through the wave impedance
Z0 : p± (ct ± x) = ±Z0 u± (ct ± x), where Z0 = ρair c/S and S is the constant cross-sectional area of
the bore. For conical geometries, the cross-section S is not constant and the definition of Z0 has to be
generalized. The wave impedance is then defined as a function Z0 (s) such that the relations P ± (r, s) =
±Z0 (s)U ± (r, s) hold in the Laplace domain. It can be seen that Z0 (s) = ρair c/S · [rs/(rs + c)].
In summary, Kirchhoff and wave variables in elastic media obeying the D’Alembert equation are
related through wave impedance and Eqs. (3.43). This results provide the basis for developing 1-D
waveguide structures.

3.4.1.2 Delay lines

Waveguide models exploit the existence of the solution (3.15) to the D’Alembert equation and discretize
this solution instead of the differential equation itself. This remark explains to a large extent why waveg-
uide structures are much more efficient than finite difference methods in simulating vibrations of elastic
media, at least in the 1-D case.
As a starting example, consider a pressure distribution p = p+ +p− inside an ideal lossless cylindrical
bore. We want to discretize p both in time and in space. If Ts is the sampling period, a suitable choice for
the spatial sampling step is Xs = cTs . Assume for simplicity that the length L is a multiple of the spatial
step, L = mXs . Then a discretized version of p is obtained through the variable substitution x 7→ mXs
and t 7→ nTs (with m, n ∈ N), and leads to

p(mXs , nTs ) = p+ (ncTs − mXs ) + p− (ncTs + mXs ) = p+ ((n − m)cTs ) + p− ((n + m)cTs ).

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-20 Algorithms for Sound and Music Computing [v.February 2, 2019]

p+[n] p+[n−m]
z− m

p [0,n] p [m,n]

z− m
L
p− [n] p− [n+m]
(a)

p+[n]
z− m

p [0,n] p [m,n]

ro r o−1 (L+ro) −1

L z− m
p− [n]

(b)

Figure 3.9: Lossless waveguide sections with observation points at position x = 0 and x = mXs = L;
(a) cylindrical section; (b) conical section.

Removing the constant sampling steps yields

p[m, n] = p+ [n − m] + p− [n + m]. (3.44)

The term p+ [n − m] in Eq. (3.44) can be thought of as the output from a digital delay line of length m,
whose input is p+ [n]. Analogously, the term p− [n + m] can be thought of as the input of a digital delay
line with the same length, whose output is p− [n]. This remark leads to the definition of a waveguide
section as a bidirectional delay line, as depicted in Fig. 3.9(a). The horizontal direction of this struc-
ture has a straightforward physical interpretation: it corresponds to the position x along the axis of the
cylindrical bore. In the example depicted in Fig. 3.9(a), two “observation points” have been chosen at
x = 0 and x = mXs = L. At these points, the pressure signal at time n is reconstructed by summing
the corresponding pressure waves p± .
A very similar structure can be outlined for numerically simulating a pressure distribution in an ideal
lossless conical bore. In this case, propagation is described by the one-dimensional equation (3.13),
whose general solution is given by

1
R(r, t) = [R̃+ (ct − r) + R̃− (ct + r)]. (3.45)
r
The conical waveguide is therefore defined as in Fig. 3.9(b). Observation points can be chosen analo-
gously to the cylindrical case.
At the beginning of this discussion we have assumed for simplicity that L = mXs . However this
quantization of the allowed lengths is too coarse for our purposes: with a sampling rate Fs = 44.1 kHz
and with a wave velocity c = 347 m/s (sound velocity in air at 20 C◦ ), the resulting spatial step is
Xs = 7.8 · 10−3 m. Length differences of this magnitude produce perceivable pitch variations in a wind
instrument. One way to overcome this limitation is to include in the structure a fractional-delay filter
(see Sec. 3.3.2) that provide fine tuning of the length of a waveguide section.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-21

y+ y+

1
−1

y− y−
(a) (b)

Figure 3.10: Ideal waveguide terminations: (a) positive reflection; (b) negative reflection.

3.4.1.3 Boundary conditions

Looking at Fig. 3.9 we immediately realize that we still one element in order to come out with a compu-
tational structure that describes e.g. a string with fixed ends or a cylindrical tube section with open ends:
boundary conditions.
In Sec. 3.2.2 we have briefly discussed fixed-end and free-end boundary conditions for the displace-
ment y(x, t)|x=0,L of a vibrating string. These can be immediately turned into reflection conditions for
both velocity waves and force waves. As an example, a fixed-end condition implies that the velocity is
0 at the boundaries, therefore the reflection conditions v + = −v − applies at both points. By looking at
Eq. (3.43), one also see that the 0 velocity condition translates into the reflection condition f + = f − at
both points. Therefore wave variables at the boundaries are multiplied by either 1 or −1 (see Fig. 3.10).
More in general, reflection conditions can be derived by formulating boundary conditions for Kirch-
hoff variables and then using Eq. (3.43) to relate Kirchhoff variables to wave variables. A second relevant
example is that of a cylindrical bore of length L, with a closed end at x = 0 and an open end at x = L.
The first condition implies u = u+ + u− = [p+ − p− ]/Z0 = 0 at x = 0 (no flow through a closed end),
which in turn implies the reflection conditions u+ = −u− and p+ = p− . The second condition implies
p = p+ + p− = 0 at x = L (p matches the atmospheric pressure at the open boundary), which in turn
implies the reflection conditions p− = −p+ and u+ = u− .
With these concepts in mind we can now go back to Sec. 3.3.1 and reinterpret the IIR comb structure
used to construct the KS algorithm. The IIR comb can be viewed as a pair of waveguide sections of
length m/2 samples in which traveling waves circulate and reflect at the boundaries according to some
reflection condition. If the coefficient g has a positive sign, as in Eq. (3.24), the corresponding condition
is that of a string fixed at both ends. The signal traveling into the filter can be interpreted either as
a velocity wave (two sign inversions at the boundaries) or as a force wave (no sign inversions at the
boundaries). As a result a harmonic spectrum is generated that contains all the partials. On the other
hand, if the coefficient g has a negative sign, as in Eq. (3.25), the corresponding condition is e.g. that
of a cylindrical bore with one open end and one closed end. The signal traveling into the filter can be
interpreted either as a flow wave or as a pressure wave (both with one sign inversion at the boundaries).
As a result a harmonic spectrum is generated that contains only the odd partials.

3.4.2 Modeling real world phenomena

As already mentioned, the waveguide structures introduced above describe ideal systems, i.e. ideally
elastic media, where the D’Alembert equation (3.12) or its spherical version (3.13) hold. Real systems
exhibit more complex behaviors. Two phenomena are particularly relevant for sound production: dissi-

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-22 Algorithms for Sound and Music Computing [v.February 2, 2019]

p+[n]
z− m

1 p [0,n] p [m,n]
1

Hdisp Hloss z− m
p− [n]

Figure 3.11: Waveguide simulation dissipation and dispersion phenomena through insertion of loss and
dispersion filters.

pation and dispersion. Both can be accounted for by adding proper time, space or time-space derivatives
of different orders to the ideal wave equation. Correspondingly the basic waveguide structure is modified
by inserting appropriate loss and dispersion filters in the loop, as in Fig. 3.11

3.4.2.1 Dissipation
Energy dissipation occurs in any real vibrating medium. In an acoustical bore this is due to air viscosity,
thermal conduction and wall losses. Dissipation in a string comes from internal losses related to elastic
properties of the material, energy transfer through terminations, and friction with air. For clarity, con-
sider the pressure distribution in a cylindrical bore. In the simplest approximation, all of the dissipation
phenomena can be incorporated in the D’Alembert equation by including an additional term proportional
to the first time derivative. As an example, a first-order approximation of a string with linear density µ,
tension T , and dissipation is given by the modified D’Alembert equation

∂2p ∂2p ∂p
µ (x, t) = T (x, t) − d1 (x, t). (3.46)
∂t2 ∂x2 ∂t
In the limit of small d1 , Eq. (3.46) still admits a traveling wave solution, which can be digitized with the
same procedure described in the ideal case:
d1 x d1 x
p(x, t) = e− 2c p+ (ct − x) + e 2c p− (ct + x), then
d1 Ts
(3.47)
p[m, n] = g m p+ [n − m] + g −m p− [n + m], with g = e− 2 < 1.

Thus the traveling waves are exponentially damped along the propagation direction, and this phenomenon
can be incorporated in the waveguide structure. In many real-world phenomena, however, losses increase
with frequency. As an example, the dissipative force exerted by the air on a moving string section is,
to a first approximation, directly proportional to the frequency of oscillation. Similar remarks apply to
the effects of internal material losses. A better approximation of dissipation phenomena in a string is
provided by the equation

∂2p ∂2p ∂p ∂3p

µ (x, t) = T (x, t) − d 1 (x, t) + d2 (x, t), (3.48)
∂t2 ∂x2 ∂t ∂t∂ 2 x
where d1 introduces frequency-independent dissipation and d2 introduces frequency-dependent dissipa-
tion. This frequency dependence can be accounted for by substituting the constant factor g with a loss
filter, which will have a low-pass characteristics. This is shown in Fig. 3.11, where losses have been
consolidated, or lumped, in a single loss filter Hloss (z) cascaded to the delay line. This filter summarizes
the distributed losses occurring in the spatial interval [0, 2mXs ].

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-23

With these concepts in mind we can go back again to Sec. 3.3.1 and reinterpret the the comb struc-
tures. In the simple IIR comb filter, the coefficient g < 1 plays the role of the loss factor g m , and
accordingly introduces equal decay times to all partials. In the low-pass comb filter, the low-pass trans-
fer function Hlp plays the role of the loss filter Hloss (z), and accordingly introduces frequency-dependent
decay times to the partials.

3.4.2.2 Loss filter design

There are many techniques for designing a loss filter Hloss to fit a real object. In this section we outline
a relatively simple approach to fit a lossy waveguide model to a real string sound.
First the sound of the target string has to be recorded and analyzed. This can be done using e.g. the
sinusoidal peak detection/continuation algorithms discussed in Chapter Sound modeling: signal based approaches.
As a result from the analysis stage, the frequencies fk and the decay times τk (k = 1, . . . , N ) of the
first N partials can be estimated. In particular τk is defined as the time required by the amplitude of the
kth partial to decay by 1/e with respect to its initial amplitude. A robust way of calculating the τk ’s
is fitting a line by linear regression on the logarithm of the amplitude envelopes derived from the peak
continuation algorithm.
The estimated parameters fk , τk specify the magnitude of Hloss over a set of N points:
( )
2πf
j F k − k
Hloss e s = e fk τk , k = 1, . . . , N. (3.49)

Given this magnitude

∑ specification,
( j2πf /Fa )common technique to design Hloss is through minimization of the
squared error N (H e k s − e −k/fk τk )2 . However one problem with these techniques is
k=1 loss
that one may find a filter whose magniture exceeds unity, which would result in an unstable waveguide
structure. Moreover, in order to avoid frequency dependent delay, Hloss (z) should be ideally a linear-
phase filter (and the length of the delay line should be reduced correspondingly, in order to obtain the
desired overall delay).
A more straightforward design approach amount to choose a first order IIR low-pass filter:

1+α
Hloss (z) = g , (3.50)
1 + αz −1

with −1 < α < 0 and g < 1. One can show that in this case the approximate analytical formulas for the
decay times are
( )2
1 2πfk α 2
≃a+b , with a = f0 (1 − g), b = −f0 , (3.51)
τk Fs 2(α + 1)

and where f0 is the fundamental frequency. Therefore the decay rate 1/τk is a second-order polynomial
of fk with even order terms. Consequently a and b can be straightforwardly determined by polynomial
regression from the prescribed decay times, and finally g and α are computed from a and b via the
inverse of Eqs. (3.51). In most cases, the one-pole loss filter yields good results. Nevertheless, when
precise rendering of the partial envelopes is required, higher-order filters have to be used.

M-3.6
Realize a complete loss filter design procedure, to be applied to a guitar sound. Use the spectral analysis tools
to estimate the decay times of the guitar string partials. Use Eq. (3.51) to design the filter (3.50).

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-24 Algorithms for Sound and Music Computing [v.February 2, 2019]

3.4.2.3 Dispersion
A second important phenomenon in natural wave propagation is dispersion. In a string, dispersion is
introduced by string stiffness, i.e. the phenomenon by which a string opposes resistance to bending.
Such a shearing force can be modeled as a fourth spatial derivative, which is introduced as an additional
term in the D’Alembert equation:

∂2p ∂2p ∂4p

µ (x, t) = T (x, t) − D (x, t), (3.52)
∂t2 ∂x2 ∂4x
where the dispersive correction term D is usually termed “bending stiffness” of the string, and is pro-
portional to the string Young’s modulus. If D is sufficiently small, its first-order effect is to increase the
wave propagation speed with frequency:
( )
Dω 2
c(ω) = c0 1 + , (3.53)
2T c20
√
where c0 = T /µ is now the wave propagation speed in the absence of dispersion. Equation (3.53)
states that a traveling wave is no longer a rigid shape that translate at constant speed. Instead, frequencies
“disperse” as they propagate with different velocities.5 As a consequence, the frequencies fk of the
allowed partials are not harmonic, instead they are stretched onto an inharmonic series according to the
equation √
fk = kf0 Ik , where Ik ≈ 1 + Bk 2 , (3.54)
and where B = π 2 D/T L2 . The quantity Ik is usually termed index of inharmonicity. Dispersion is
particularly important in piano strings, where the lower tones exhibit significant inharmonicity.
Having a non-uniform wave velocity c(ω) implies that it is not possible to define a sampling step as
Xs = c0 Ts . Instead, it can be said that a component with frequency f = ω/(2π) travels a distance c0 Ts
in the time interval c0 Ts /c(ω). As a consequence, each unitary delay z −1 in the waveguide structure
has to be substituted with an all-pass dispersion filter with unitary magnitude response and a non-linear
phase response approximates the frequency-dependent phase delay c0 Ts /c(ω).
Similarly to dissipative low-pass filters, these all-pass delays can be lumped into a single product
filter. Moreover, the linear and non-linear parts of the phase response can be treated separately. In
conclusion the dispersion filter that substitutes 2m unitary delays can be written as z −2m Hdisp (z), where
z −2m accounts for the linear part of the phase response and the all-pass filter Hdisp (z) approximates the
non-linear part. The resulting dispersive waveguide structure is then as in Fig. 3.11.

M-3.7
Implement the waveguide structure of Fig. 3.11, including the loss filter (3.50) and an all-pass filter to simulate
dispersion.

3.4.2.4 Dispersion filter design

Similarly to the discussion on loss filter design, the effects of dispersion in a real sound can be estimated
from analysis using e.g. the sinusoidal peak detection/continuation algorithms discussed in Chapter Sound
modeling: signal based approaches. The estimated series of partial frequencies fk provide an indication of the
degree of inharmonicity in the sound, and thus of dispersion.
5
Dispersion can be sometimes experienced when hiking on the mountains, by imparting an impulse on a long metallic cable
such as that of a cableway: after some seconds the impulse will bounce back and one will feel that it has “unraveled” into a
smoother step with high-frequency ripples running out ahead.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-25

In this section we outline one possible approach to the dispersion filter design. The total phase delay
over a waveguide loop of length 2m, with loss and dispertion filters is

kFs
τph (fk ) = = 2m + τloss (fk ) + τdisp (fk ). (3.55)
fk

With everything else known, this equation provide a phase delay specification for the dispersion filter:

kFs
τdisp (fk ) = − 2m − τloss (fk ). (3.56)
fk

Given L estimated partial frequencies {fk }k=1,...,L , one can then design an all-pass filter of order N < L
as follows. First, for each partial compute the quantities

1[ ]
βk = − τdisp (fk ) − 2N πfk , k = 1, . . . , L. (3.57)
2
Then, filter coefficients are computed by solving the system

∑
N
aj sin(βk + 2jπfk ) = sin βk , k = 1, . . . , L. (3.58)
j=1

This is an overdetermined system. It can be solved with a LS error criterion.

Note that this design approach is not based on the fitting of the relative positioning of the partials, but
on the absolute values of the fk ’s. Therefore the resulting all-pass filter accounts both for the simulation
of dispersion and for the fine-tuning of the string (fractional delay).

M-3.8
Realize a complete dispersion filter design procedure, to be applied to a piano sound. Use the spectral analysis
tools to estimate the frequencies of the piano partials. Use Eq. (3.58) to design the all-pass filter.

3.4.3 Junctions and networks

The last section has introduced the main concepts of waveguide modeling for a signal propagating in
a uniform medium. When discontinuities are encountered, the wave impedance changes and signal
scattering occurs, i.e. a traveling wave is partially reflected and partially transmitted.
Examples of non-uniform media are a cylindrical bore where the cross-sectional area changes abruptly,
or a string where the value of the linear mass density jumps changes discontinuously. In order to model
these discontinuities, appropriate junctions have to be developed, that connect two (or more) waveguide
sections. The boundary reflection conditions that we have examined at the end of Sec. 3.4.1 can be
regarded as special cases of junctions, as discussed in the following paragraphs.

3.4.3.1 The Kelly-Lochbaum junction

−1
Consider two cylindrical bores, with cross-sectional areas S1,2 and wave admittances Γ1,2 = Z1,2 =
S1,2 /ρair c, connected to each other. Analysis of this problem leads to the derivation of the well known
Kelly-Lochbaum junction.
The derivation is based on imposing appropriate physical constraints on the Kirchhoff variables p, u
at the junction. Specifically, continuity requires that pressures p1,2 have the same value pJ at the junction.
Moreover, the flows u1,2 from the two sides must sum to zero (simply said, the air entering one side of

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-26 Algorithms for Sound and Music Computing [v.February 2, 2019]

p1+ p2−

1−ρ

−ρ
ρ

p1− 1+ρ p2+

Figure 3.12: Kelly-Lochbaum junction for two cylindrical bores with different areas.

the junction and the air coming out from the other side must be the same). These two requirements lead
to the following conditions at the junction:

u1 + u2 = 0, p1 = p2 = pJ . (3.59)

Using the Kirchhoff analogy p ↔ v (voltage) and u ↔ i (current), Eqs. (3.59) can be regarded as
describing a parallel junction. If pressure wave variables are introduced as in Eq. (3.43) (with p+ and
p− denoting incoming and outgoing waves, respectively), and the junction pressure pJ is used, then the
relation p−
l = pJ − pl (for l = 1, 2) holds. Substitution in the first of Eqs. (3.59) yields
+

− − − −
1 + u1 ) + (u2 + u2 ) = Γ1 (p1 − p1 ) + Γ2 (p2 − p2 ) =
0 = (u+ + + +
(3.60)
1 − pJ ) + Γ2 (2p2 − pJ ).
= Γ1 (2p+ +

From this, the junction pressure pJ can be expressed in terms of the incoming pressure waves p+
1,2 as

Γ1 p+ +
1 + Γ 2 p2
pJ = 2 . (3.61)
Γ1 + Γ 2

Using this latter expression, the outgoing pressure waves p−

1,2 can be written as

Γ2 − Γ1 + 2Γ2
p−
1 = pJ − p1 = −
+
p + p+ ,
Γ2 + Γ1 1 Γ2 + Γ 1 2
(3.62)
2Γ1 Γ2 − Γ1 +
p−
2 = pJ − p2 = +
+
p+ + p .
Γ2 + Γ1 1 Γ2 + Γ 1 2
And finally

p−
1 = −ρ p1 + (1 + ρ)p2 ,
+ +
Γ2 − Γ1
with ρ≜ , (3.63)
p−
2 = (1 − ρ)p+
1 +ρ p+
2, Γ2 + Γ 1

These equations describe the Kelly-Lochbaum junction. The quantity ρ is called the reflection coefficient
of the junction. A scattering diagram is depicted in Fig. 3.12.
This junction has been extensively used in so-called “multitube lossless models” of the vocal tract.
These are articulatory models where the vocal tract shape is approximated as a series of concatenated
cylindrical sections. Pressure wave propagation in each section is then described using digital waveg-
uides, and interconnections are treated as Kelly-Lochbaum junctions. However this very same junction
can be used to describe not only acoustic, but also mechanical structures. As an example, consider two
strings with different densities, connected at one point: this can be thought of as a series junction, since
the physical constraints impose that velocity (i.e., “current”) has to be the same on the left and right

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-27

p+ p-
1 2
z - m1 z - m2
closed

open
1
-1
ρ

z - m1 z - m2
p+
p- 2
1
L1 L2

Figure 3.13: Example of use of the Kelly-Lochbaum junction: (a) a parallel junction of two cylindrical
bores; (b) realization with two waveguide sections and a Kelly-Lochbaum junction.

sides, and the sum of forces (i.e., “voltages”) from the two sides must be zero. Analogously to the above
analysis, a series Kelly-Lochbaum junction can be derived in this case.
Terminations of a waveguide model are an interesting particular case of junctions. Consider an ideal
cylindrical bore, closed at one end: this boundary condition corresponds to an infinite impedance Z2 =
∞ (i.e., S2 = 0), and thus to a reflection coefficient ρ = −1. In other words, complete reflection occurs
and the relation p− +
1 (0, t) = p1 (0, t) holds. Similarly, an ideally open end can be seen to correspond
to Z2 = 0 (i.e., S2 = ∞), and thus to ρ = 1: this is a second case where complete reflection occurs,
namely the relation p−1 (0, t) = −p1 (0, t) holds. These reflection conditions are identical to those derived
+

in Sec. 3.4.1 (similar considerations hold for string terminations).

Figure 3.13 shows an example where different junctions have been used and combined into a waveg-
uide model. Note that in this example the scattering junction between the two cylindrical sections is not
in the original Kelly-Lochbaum form; instead, a one-multiply scattering junction is used, which allows
more efficient implementation of Eqs. (3.63). Open- and closed-tube terminations are modeled according
to the above remarks.

M-3.9
Implement the waveguide structure of Fig. 3.13. Add a loss filter (3.50) to each WG section.

3.4.3.2 N-dimensional and loaded junctions

The result expressed in Eq. (3.63) can be readily extended to higher dimensions. Consider a parallel
junction of N acoustic bores. In this case a scattering matrix can be found, and Eq. (3.63) is generalized
to

p− = A · p+ , (3.64)

where p± are n-dimensional vectors whose elements are the incoming and outgoing pressure waves in
the n bores. The physical constraints expressed in Eq. (3.59) are also generalized as

p1 = p2 = . . . = pN = pJ ,
(3.65)
u1 + u2 + . . . + uN = 0.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-28 Algorithms for Sound and Music Computing [v.February 2, 2019]

Calculations analogous to those outlined for the Kelly-Lochbaum junction yield

 2Γ 
Γ
1
− 1, 2Γ2
ΓJ , · · · 2ΓN
ΓJ
 J 
  ∑
 2Γ1  N
A=  , 2Γ2
− 1, · · · 2ΓN

ΓJ ΓJ ΓJ , where ΓJ = Γl . (3.66)
 .. .. .. 
 . . .  l=1
2Γ1
ΓJ ,
2Γ2
ΓJ , ··· 2ΓN
ΓJ −1

As an example, a 3-dimensional junction can be used to model an acoustic hole in a wind instrument:
in this case, two waveguide sections represents the two sides of the acoustic bore with respect to the
hole, and the third one represents the hole itself. Note also that when N = 2 Eq. (3.64) reduces to the
Kelly-Lochbaum equations.
A second relevant extension of the Kelly-Lochbaum junction is the loaded junction, in which an
external signal is injected into the system. A simple example is that of a string that is excited (e.g.
hammered) at a given point. For continuity, the velocity of the string in this contact point will be the
same at both sides. Moreover, during the contact this velocity will be equal to the velocity of the hammer.
Finally, the sum of the forces at the contact point equals the hammer force. The following equations of
continuity are then derived:

v1 = v2 = vJ , f1 + f2 + fJ = 0. (3.67)

With the Kirchhoff analogies this is a series junction with an external load (the “currents” at the junction
are the same, and the potentials at the junction sum to the driving potential). Then

fJ = − f1 − f2 = . . . = −2Z0 (v1+ + v1− − vJ ),

1 (3.68)
⇒ vJ = v1+ + v2+ − fJ .
2Z0
This yields the scattering equations for the loaded junction:
1
v1− [n] =vJ [n] − v1+ [n] = v2+ [n] + fJ [n],
2Z0
(3.69)
1
v2− [n] =vJ [n] − v2+ [n] = v1+ [n] + fJ [n].
2Z0
The corresponding computational structure is shown in Fig. 3.14. This structure may be further extended
to the case of N -dimensional parallel or series loaded junctions.

M-3.10
Implement the waveguide structure of Fig. 3.14. Add a loss filter (3.50) to each WG section.

3.4.3.3 Non-cylindrical geometries

A final remark is concerned with junctions of conical elements. Generalizing the cylindrical case is not
straightforward, since the derivation of Kelly-Lochbaum equations is based on the implicit assumption of
plane wave propagation. This assumption permits imposition of the constraints (3.59) on a flat scattering
boundary, which is a wavefront for both p1 and p2 . But wavefronts in conical sections are spherical and
this circumstance makes it impossible to define a unique surface on which boundary conditions can be
applied: Fig. 3.15(a) shows that there is a region between the two spherical wavefronts which is within
neither conical segment. This ambiguity in the definition of the scattering boundary is usually overcome

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-29

v+
1 [n] v−
2 [n]
Waveguide Waveguide
string 1 string 2
v−
1 [n] v+
2 [n]

1/2Z 0
fJ [n]

Figure 3.14: Example of a loaded junction: a waveguide structure for a string excited by an external
force signal fJ [n] (e.g. a hammer).

Transition
volume

S2 S1
S1 S2

(a) (b)

Figure 3.15: Boundary regions for (a) non-convex and (b) convex conical junctions.

by assuming that the transition volume is small and thus pressure is constant inside the volume. Under
this assumption, continuity conditions analogous to (3.59) are imposed and the reflection coefficient ρ is
generalized to a first order filter R(s).
However, a second and more serious problem arises when one looks at the nature of R(s). This
filter turns out to be unstable (non-causal growing exponential) in the case of the convex configuration
depicted in Fig. 3.15(b). While this circumstance is physically consistent (in the continuous-time domain
the scattered waves can grow exponentially only for a limited time because they are cancelled out by
subsequent multiple reflections), in a numerical simulation the system can turn out unstable, due to the
approximations introduced by the discretization process and to round-off errors introduced by finite-
precision.

3.5 Lumped models and the modal approach

Lumped modeling approaches can be applied in variety of contexts where the physical system under
exam can be represented with ideal lumped elements. These include electrical circuits, with linear
elements like capacities, resistances, and inductances, connected in series and parallel; mechanical sys-
tems viewed as ideal point masses connected through springs and dampers, representing mechanical
resonators; acoustic systems viewed as networks of linear acoustic elements like bores, cavities, and
acoustic holes (like the Helmoltz resonator examined previously).
Lumped models are particularly suited for describing systems whose spatial dimensions are small
compared to acoustic wavelengths. As an example, pressure-controlled valves, such as single, double

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-30 Algorithms for Sound and Music Computing [v.February 2, 2019]

or lip reeds, can be conveniently described using the lumped modeling paradigm. Although these sys-
tems are quite complicated, due to their limited spatial extensions they can be modeled using lumped
elements, and it is widely accepted that such a simplified description captures the basic behavior of
pressure controlled valves. Similar remarks hold for hammers and mallets: during collision, they are de-
formed and subject to internal losses and non-linear restoring forces. However, interactions with strings
and bars have been modeled and efficiently implemented in sound synthesis algorithms by assuming the
hammer/mallet to be a lumped mass.

3.5.1 Numerical methods

Unlike waveguide structures, lumped models are developed in the continuous-time domain, and are de-
scribed through sets of ordinary differential equations (ODEs). In order to be implemented as numerical
algorithms for sound synthesis, the differential equations have to be discretized in an efficient and effec-
tive manner. In most cases, a trade-off has to be found between accuracy of the discretization technique
and efficiency of the resulting algorithms.

3.5.1.1 Impulse invariant method

When dealing with linear time-invariant systems, the most elementary technique to turn a continuous-
time system into a discrete-time one is sampling its impulse response.
If a continuous-time LTI system is described in terms of Kirchhoff variables, then it is possible to
define a transfer function which coincides with the admittance Γ(s) of the system. As an example, in
a mechanical lumped system this corresponds to defining the input as a driving force and the output as
the resulting velocity. The inverse Laplace transform γ(t) is the continuous-time impulse response. The
linear system can thus be digitized by defining the discrete response as γd [n] ≜ γ(nTs ), i.e. by sampling
γ(t). This technique is widely used in the context of digital filter design, and it is usually termed the
Impulse invariant method.
Assume that the continuous-time system has a rational transfer function Γ(s). This can be rewritten
using a partial fraction expansion (similarly to what we have done in Chapter Fundamentals of digital audio
processing for discrete-time systems):

∑M ∑
N
B(s) bk sM −k Kk
Γ(s) = = ∑k=0 , ⇒ Γ(s) = , (3.70)
A(s) N
k=0 ak s
N −k s − pk
k=1

where the pk ’s are the poles of the system. By taking the inverse Laplace transform of this latter equation,
one can see that the impulse response γ(t) is a combination of complex exponentials. This impulse
response is then sampled to obtain its digital counterpart:

∑
N ∑
N
( )n
γ(t) = Kk e pk t
, ⇒ γd [n] ≜ γ(nTs ) = Kk epk Ts . (3.71)
k=1 k=1

Finally, taking the Z-transform of γd yields:

∑
N
Kk Bd (z)
Γd (z) = −1
= , with pd,k = epk Ts . (3.72)
1 − pd,k z Ad (z)
k=1

This equation tells that the transfer function Γd (z) of the discretized system is still rational, with N poles
pd,k uniquely determined by the continuous-time poles pk .

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-31

One quality of the method is that stability is guaranteed at any sampling rate: if the continuous-time
system is stable, i.e. Re(pk ) < 0 for all k, then Eq. (3.72) tells that | pd,k | < 1 for all k, i.e. the discrete
time system is also stable. On the other hand, a drawback of the method is aliasing. Since γd [n] has been
obtained by sampling γ(t), then the discrete-time response Γd is a periodization of Γ:

∑
+∞ ( )
jω jω 2kπ
Γd (e ) = Γ +j . (3.73)
Ts Ts
k=−∞

As a consequence, aliasing can occur in Γd if the bandwidth of Γ exceeds the Nyquist frequency.

3.5.1.2 Finite differences and mappings “s-to-z”

An alternative approach to the discretization of ODEs amounts to replacing time derivatives with finite
differences, thus turning the differential equations directly into difference equations. Since in the Laplace
domain the derivation operator is turned to a multiplication by s, and since in the z-domain the unit delay
is turned into a multiplication by z −1 , approximating derivatives with finite differences corresponds to
finding appropriate s-to-z mappings. Let s = g(z) be such a mapping, then if the original continuous-
time system is LTI with impulse response Γ(s), the discrete-time response is found as Γd (z) = Γ (g(z)).6
The simplest possible mapping is obtained by replacing the derivative with an incremental ratio. Let
x(t) be a generic smooth function of time, then

dx x(nTs ) − x(nTs − h) x[n] − x[n − 1] 1 − z −1

(nTs ) = lim ≈ ⇒ s≈ ≜ g1 (z). (3.74)
dt h→0+ h Ts Ts

The mapping g1 (z) is known in numerical analysis as the backward Euler method. The adjective “back-
ward” is used because the first derivative of x at time n is estimated through the values of x at time n
and n − 1. Higher-order derivatives can be estimated through iterate application of Eq. (3.74). As an
example, the second derivative is computed as
[ ]
d2 x 1 x[n] − x[n − 1] x[n − 1] − x[n − 2] x[n] − 2x[n − 1] + x[n − 2]
(nTs ) ≈ − = . (3.75)
dt2 Ts Ts Ts Ts2

Alternatively, a centered estimate is also often used in combination with the backward Euler method. In
this case the second derivative is computed as:

d2 x[n + 1] − 2x[n] + x[n − 1]

2
x(tn ) ≈ . (3.76)
dt Ts2

A second, widely used s-to-z mapping is provided by the bilinear transform. Like the backward Eu-
ler method, it can be seen as a finite approximation of the time derivative, but in this case the incremental
ratio is assumed to approximate the value of ẋ(t) averaged on time instants nTs and (n − 1)Ts :

ẋ(nTs ) + ẋ((n − 1)Ts ) x[n] − x[n − 1] 1 − z −1

≈ , ⇒ s ≈ 2Fs ≜ g2 (z). (3.77)
2 Ts 1 + z −1
The mapping g2 (z) is known in numerical analysis as the one-step Adams-Moulton method.
Both the backward Euler method and the bilinear transform are implicit numerical methods. This
means that both methods turn a generic first-order differential equation ẋ(t) = f (x(t), t) into a difference
6
Note however that, unlike the impulse invariant method, finite differences do not assume linearity and time invariance of
the original system, and are therefore more general methods.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-32 Algorithms for Sound and Music Computing [v.February 2, 2019]

bilinear transform
1 Euler method

0.5

ω=0
Im(z)

−0.5

−1

−1 −0.5 0 0.5 1
Re(z)

Figure 3.16: Mapping of the vertical axis s = jω (solid circle lines) and of the left-half s-plane (shaded
regions) using the backward Euler method g1 and the bilinear transform g2 .

equation of the form x[n] = fd (x[n], x[n − 1], n), in which x[n] depends implicitly on itself through
the function fd . This is a source of problems for the resulting discrete-time system, since the difference
equation is not computable explicitely due to the instantaneous dependence of a variable on itself. Below
we discuss briefly this computability problem in the case of linear systems. Note that one advantage of
the centered estimate (3.76) is that when it is applied in conjunction with the Euler method to a second-
order ODE it leads to an explicit difference equation.

3.5.1.3 Accuracy, stability, computability

A comparison between the first estimate in Eq. (3.77) and the first in Eq. (3.74), gives the intuition that the
bilinear transform provides a more accurate approximation than the Euler method. A rigorous analysis
would show that the order of accuracy of the bilinear transform is two, while that of the backward Euler
method is one.
Another way of comparing the two techniques consists in studying how the frequency axis s = jω
and the left-half plane Im(s) < 0 are mapped by g1,2 into the discrete domain. This provides information
on both stability and accuracy properties of g1,2 . As shown in Fig. 3.16, both the methods define one-
to-one mappings from s = jω, onto two circles. Therefore no frequency aliasing is introduced. Second,
both the methods are stable, since the left-half s-plane is mapped inside the unit circle by both g1 and
g2 . However we also see that both mappings introduce frequency warping, i.e. the frequency axis is
distorted. The bilinear transform g2 maps the axis s = jω exactly onto the unit circle z = ejωd , and the
mapping between the continuous frequency ω and the digital frequency ωd can be written analytically:
(ω ) ( )
2 1 − e−jωd 2j d ωTs
jω = = tan , ⇒ ωd = 2 arctan . (3.78)
Ts 1 + e−jωd Ts 2 2

At low frequencies ωd increases almost linearly with ω, while higher frequencies are progressively com-
pressed (warped) and ωd → ±π as ω → ±∞. This warping phenomenon is the main drawback of the
bilinear transform.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-33

u2 w u2 w

~
a y x u1
a b/(1-ab)
y b x u1

(a) (b)

Figure 3.17: A linear discrete-time system; (a) delay-free path, (b) equivalent realization with no delay-
free paths.

For the Euler method no analytic mapping can found from ω to ωd . The function g1 c “doubly” warps
the frequency axis: there is a progressive warping in the direction of increasing frequency (similarly to
the bilinear transform), and there is also warping normal to the frequency axis. Figure 3.16 also shows
that the poles of the discrete-time system obtained with g1 are more “squeezed” inside the unit circle than
those obtained with g2 . Furthermore, it can happen that continuous-time poles with positive real-part are
turned by g1 into discrete-time poles with modulus less than unity: in other words g1 can turn unstable
continuous systems into stable discrete systems. This numerical damping is a second major drawback of
the Euler method.
One more relevant aspect to discuss is the computability of the discrete-time systems obtained when
discretizing a system of ODEs with either g1,2 (or other mappings). As already stated, being these
implicit methods the resulting difference equations are implicit. In order to clarify this point, let us
consider the simple example depicted in Fig. 3.17(a). This system can be written as



w[n] = w̃[n] + y[n], with w̃ = u2 ,
x[n] = x̃[n] + ay[n], with x̃ = u1 + au2 , (3.79)


y[n] = bx[n], ⇒ y[n] = b[u1 [n] + au2 [n] + ay[n]],

where we have defined tilded variables w̃ and x̃ than only depend on the external inputs u1,2 , and are
therefore known at each time n.
The signals y and x are connected through a delay-free loop and the resulting set of difference
equations is implicit: in particular the last of Eqs. (3.79) shows that y[n] depends implicitly on itself. It
is easy, however, to rearrange the computation in order to solve this problem: the last of Eqs. (3.79) can
be inverted, yielding
b
y[n] = [u1 [n] + au2 [n]]. (3.80)
1 − ab

This new equation relates y to the computable vector x̃. Therefore, an equivalent realization of the system
is obtained as shown in Fig. 3.17(b). The key point in this example is that the discrete-time system is
linear, which allows explicit inversion of the last equation in (3.79).
This simple example is an instance of the so-called delay-free loop problem. In the linear case the lit-
erature of digital signal processing provides techniques for the restoring computability by rearrangement
of the structure.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-34 Algorithms for Sound and Music Computing [v.February 2, 2019]

3.5.1.4 Wave digital filters

The bilinear transform finds application in the theory of Wave Digital Filters (WDFs). These structures
are digital equivalents of the lumped circuit elements described in Sec. 3.2.1. WDF theory has been
developed primarily for electric circuits but can be applied as well to mechanical and acoustic systems
using Kirchhoff analogies.
Wave digital filters are constructed in two steps. The first step amounts to converting the continuous-
time lumped circuits in wave variables. Here the definition of wave variables is identical to that used for
waveguides models (see Eq. (3.43), namely:
f + Z0 v f − Z0 v
f+ = , f− = , (3.81)
2 2
where the mechanical Kirchhoff variables force f and velocity v have been used for clarity. The only
and fundamental difference with Eq. (3.43) is that here Z0 is a reference impedance that can be given
any value and has no direct physical interpretation. The variables f ± themselves do not have a clear
physical interpretation since in a lumped model they cannot be interpreted as traveling waves. Therefore
Eqs. (3.81) have to be regarded as a mere change of coordinates.
Using wave variables, circuit elements can be converted into one-port elements. Given one of the
elementary lumped elements analyzed in Sec. 3.2.1 (resistance, inductance, capacity) and its associated
impedance Z(s), the new variables f ± are related to each other through a reflectance R(s):

Z(s) − Z0
F (s) = Z(s)V (s), ⇒ F − (s) = R(s)F + (s), with R(s) ≜ . (3.82)
Z(s) + Z0

The circuit element can then be visualized as a black box with a port consisting of two terminals, with
a port voltage applied across them, and an associated flowing current, as in Fig. 3.18(a). A linear sys-
tem can then be modeled through series and parallel connections of one-port elements: as an example,
Fig. 3.18(b) visualizes a series connection of two ports representing the mechanichal system

mẍ(t) + kx(t) = f (t). (3.83)

The second step in WDF design is the discretization of R(s). The equivalent wave digital filter is
obtained using the bilinear transform as R(g2 (z)). Note that since the reference impedance Z0 can be
given any value, this provides an additional degree of freedom in the design. In particular, Z0 can be
chosen such that the WDF has no delay-free paths from input to output, therefore guaranteeing com-
putability when connecting more than one element. As an example, consider the three elementary me-
chanical impedances Zmass (s) = ms, Zspring (s) = k/s, Zloss (s) = r. For the mass, the reflectance is
Rmass (s) = (ms − Z0 )/(ms + Z0 ), therefore the equivalent WDF is

(2Fs − Z0 /m) − (Z0 /m + 2Fs )z −1

Rmass (z) = , ⇒ Rmass (z) = z −1 with Z0 = 2Fs m. (3.84)
(2Fs + Z0 /m) − (Z0 /m − 2Fs )z −1

Therefore choosing Z0 = 2Fs m leads to the interesting result that no delay-free path is present in the
corresponding WDF. Similarly, one can prove that Rspring (z) = z −1 with Z0 = k/2Fs , and Rloss = 0
with Z0 = r.
This brief section has shown that WDFs can be used to digitize lumped element networks using
wave variables and adapted impedances in such a way that delay-free computational loops are avoided in
the resulting numerical structure. We have shown a single example of a series connection between two
elements. The concept of connection is generalized in WDF theory with the concept of adaptors, which
are N -port elements that model interconnection between arbitrary numbers of elements.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-35

+ ms −Z 0,1
f(t) R1 (s) R1 (s)=
− ms+Z 0,1
f +(t)
+
+ k/s − Z 0,1
f(t) v(t) R(s) R2 (s) R2 (s)=
− k/s +Z 0,1
−
f −(t)

(a) (b)

Figure 3.18:

3.5.2 Modal synthesis

Modal synthesis is conceptually simple: the sound of a resonating object is represented as a linear com-
bination of the outputs of N second order oscillators, each of which represents one mode of oscillation of
the object excited by a driving force or acoustic pressure: in this sense modal synthesis can be regarded
as a lumped physical modeling approach, and can also be interpreted a source-filter approach in which
the source is the driving signal and the filter is bank of second-order resonators.
Understanding the mathematical and physical basis of modal theory is a bit less straightforward. In
the next sections we sketch the main concepts for both discrete systems (e.g. discrete networks of masses,
springs, and dampers) and continuous systems (i.e. partial differential equations in space and time). We
show that modal theory is fundamentally the same for these classes of systems.
Therefore the power of modal synthesis is that it is a very general technique that can be applied to
a large class of sounding physical systems (while e.g. waveguide techniques are suited only for elastic
systems that obey some perturbed version of the D’Alembert equation).

3.5.2.1 Normal modes in finite dimensional systems

In Sec. 3.2.1 we have studied the simple example of two coupled mechanical oscillators, and we have
seen that the resulting system can be viewed as the combination of two uncoupled oscillators, whose
frequencies depend on those of the original ones. This approach can be extended to a generic network of
N linear undamped oscillators:
M ÿ(t) + Ky(t) = f ext (t). (3.85)

In this equation y is a vector containing the displacements of the N points of the network, while M is
the mass matrix: typically (but not necessarily) it is diagonal and contains the masses ml (l = 1 . . . N )of
each point of the network. K is the stiffness matrix and is in general not diagonal because the points are
coupled through springs.
Now we consider the homogeneous equation (f ext ≡ 0) and look for a factorized solution of the
form y(t) = s · sin(ωt + ϕ). By substituting this into Eq. (3.85), one finds

Ks = ω 2 M s. (3.86)

This is a generalized eigenvalue problem for the matrix K: more precisely, ω 2 is an eigenvalue of
M −1 K and s is the associated eigenvector. In general one will find N distinct eigenvalues and eigen-
vectors ωi2 and si (for simplicity we consider normalized si ’s). The key property of these eigenvectors

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-36 Algorithms for Sound and Music Computing [v.February 2, 2019]

is that they are orthogonal with respect to the mass and the stiffness matrix:

sTj M si = δi,j mi , sTj Ksi = δi,j ki , (3.87)

where mi and ki = ωi2 mi are real positive scalars. The orthogonality condition also implies that the
modal shapes si are linearly independent. The si ’s can be used to define a modal transformation, i.e. a
change of spatial coordinates that transforms system (3.85) into a set of N uncoupled oscillators:

y = Sq q = S T y, with S = [s1 |s2 | . . . |sN ] . (3.88)

Substituting this into Eq. (3.85) and premultiplying by S T yields

Mq q̈ + Kq q = S T f ext (t), with Mq = S T M S, Kq = S T KS. (3.89)

By virtue of the orthogonality property, the matrices Mq and Kq are diagonal and contain the elements
mi and ki on their diagonals, respectively. Therefore this is a system of uncoupled oscillators with
frequencies ωi , the quantities mi and ki represent the masses and the stiffnesses of these modes.
The matrix S T of the modal shapes defines how a driving force f ext acts on the modes: as a
particular case, consider a scalar force acting only on the lth point of the network, i.e. f ext (t) =
[0, . . . , fext (t), . . . , 0]T (where the only non-null element is in the lth index). This force is applied to
the generic ith mode, scaled by the factor si,l , i.e. the shape of the ith mode at the lth point of the net-
work. If this factor is 0, i.e. if the ith mode has a node at the lth point of the network, then no force is
transmitted to the mode.
The oscillation yl (t) of the system at the lth spatial point will ∑be the sum of the modal oscillation
N
weighted by the modal shapes, according to Eq. (3.88): yl (t) = i=1 si,l qi (t). Again, if the ith mode
has a node at the lth point of the network, that mode will not be “heard” in this point. In conclusion the
motion of the network is determined by the motion of N second-order mechanical oscillators and by the
transformation matrix S.
This formalism can be extended to systems that include damping, i.e. where we add a term Rẏ in
Eq. (3.85).

3.5.2.2 Normal modes in PDEs

Now look at the concept of normal modes from a different perspective: a distributed object is not mod-
eled as a network of lumped elements, but instead as a partial differential equation that describes the
displacement y(x, t) as a continuous function of space and time. We can reformulate the modal descrip-
tion even in this case: as we will see, there are strict analogies with the case of finite dimensional systems
outlined above.
We use a concrete example, a string with fixed ends, to derive the modal formulation in the case
of continuous systems. In analogy with the case of finite dimensional systems, we now state that a
normal mode is a factorized solution y(x, t) = s(x)q(t). For the example under exam, we already know
that the D’Alembert equation with fixed boundary conditions admits the factorized solutions yn (x, t) =
sn (x)qn (t) of the form (3.21). If a force density fext (x, t) is acting on the string, the equation is
∂2y ∂2y
µ (x, t) − T (x, t) = fext (x, t). (3.90)
∂t2 ∂x2
If one substitutes in this equation the mode yn (x, t), and then multiplies by sn (x) and integrates over the
string length, the following equation is found:
[ ∫ L ] [ ∫ L ] ∫ L
′′
µ sn (x)dx q̈n (t) − T
2
sn (x)sn (x)dx qn (t) = sn (x)fext (x, t)dx. (3.91)
0 0 0

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-37

The second integral can be integrated by parts to obtain

[ ∫ L ] [ ∫ L ] ∫ L
s2n (x)dx q̈n (t) − T s′n (x)sn (x) 0 − [s′n (x)]2 dx qn (t) =
L
µ sn (x)fext (x, t)dx, (3.92)
0 0 0

where the term s′n (x)sn (x)|L 0 is identically zero for fixed (or even free) boundary conditions. Therefore
∫L
the equation for the nth mode is that of a second-order oscillator with mass mn = µ 0 s2n (x)dx and
∫L
stiffness kn = T 0 [s′n (x)]2 dx. For the ideal string the modal shapes are simply sn (x) = sin(nπx/L),
therefore mn = µL/2 and kn = T L/2.
The shape also defines how a driving force acts on the mode. As a particular case, consider a force
density that is ideally concentrated in a single point xin of the string, i.e. fext (x, t) = δD (x − xin )u(t)
(where the function δD (·) is the Dirac delta): then the force acting on the nth mode is sn (xin )u(t), and
if xin is a node of the mode then no force is transmitted to it. We already know that the oscillation
y(xout , t) of the system at the∑spatial point xout will be the sum of the modal oscillations weighted by the
modal shapes: y(xout , t) = +∞ n=1 sn (xout )qn (t). Again, if the nth mode has a node at the point xout ,
that mode will not be “heard” in this point.
This analysis can be extended to include dispersion and dissipation. As an example, we know that
for a dissipative string we have to add the terms d1 ∂y/∂t − d2 ∂/∂t(∂ 2 y/∂x2 ) on the left-hand side of
Eq. (3.90). Again, by substituting the mode yn (x, t) in the[equation, and then multiplying by s]n (x) and
∫L ∫L
integrating over the string length, one finds that the term d1 0 s2n (x)dx + d2 0 [s′n (x)]2 dx q̇(t) has
to be added to Eq. (3.92), which represents a viscous damping term for the second order oscillator.

M-3.11
Compute modal parameters for a string with linear dissipation.

M-3.11 Solution
function [omega,alpha,m,s]=modal_string(L,T,mu,d1,d2,N,M);

xstep=L/(M-1); xpoints=0:xstep:L;

s=zeros(N,M); omega=zeros(1,N); alpha=omega; m=omega;

for i=1:N % i is mode number
s(i,:)= sin(i*pi*xpoints/L); %spatial shape
m(i) = mu*xstep*sum(s(i,:).ˆ2); %=mu*L/2; modal mass
dsdx= i*pi/L*cos(i*pi*xpoints/L);
k = T*xstep*sum(dsdx.ˆ2); %=T*L/2*(i*pi/L)ˆ2; modal stiffness
omega(i)=sqrt(k/m(i)); %=i*pi*c/L; modal frequency
alpha(i)=(d1*xstep*sum(s(i,:).ˆ2)-d2*xstep*sum(dsdx.ˆ2) )/(2*m(i)); %loss
end

Parameters are functions of the string tension T , linear density µ, loss factors d1,2 . One can choose
the number N of modes to compute, and the number M of spatial points for the shape computation.
Two remarks. First, we are using the ideal spatial shapes: this is not correct for the dissipative string,
but is acceptable for small d1,2 values. Second, we are computing the integrals numerically, although
for the ideal string shapes these have analytical solutions: in Sec. 3.5.3 we will examine less trivial
shapes.

In conclusion the modal representation of continuous systems described by PDEs is in strict analogy
with that of discrete systems described as networks of masses and springs. Here we have obtained similar
equations, where the discrete spatial index l = 1 . . . N indicating the points of the network (3.85) has
become a continuous spatial variable x, sums over l have become integrals over x, and a numerable

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-38 Algorithms for Sound and Music Computing [v.February 2, 2019]

Mode: 1 2 3 4 5 ... 10
N= 1
2
3
4
5 ..
.
... ...
10 ... ...

(a) (b)

Figure 3.19: Analogies between continuous and discrete systems: (a) approximation of an ideal string
with a mass-spring network; (b) modes of the discrete system for different numbers N of masses.

infinity of modes has been found instead of a finite set of N modes. These strict analogies reflect the fact
that continuous systems can be seen as the limit of discrete systems when the number of masses becomes
infinite. As an example, a string can be approximated with the discrete network of Fig. 3.19(a) made of
N masses and N + 1 springs. Figure 3.19(b) shows that for a given N the system has N modes, whose
shapes resemble closely those of the first N modes of the continuous string. Moreover the approximation
grows closer and closer as N increases. One could also show that the modal frequencies of the continuous
system are underestimated by those of the discrete system, due to the spatial discretization.

3.5.2.3 Discrete-time mechanical oscillators

We have seen that each mode of either a discrete or a continuous system is a second order oscillator:
1
q̈(t) + 2αq̇(t) + ω02 q(t) = fmode (t),
m
(3.93)
m−1
Q(s) = H(s)Fmode (s), with H(s) = .
s2 + 2αs + ω02

The frequency ω0 = k/m and the loss factor α = r/m depend on the geometry and the material of the
object. The force fmode that is “felt” by a single mode depends on the modal shape and on the spatial
force distribution, and is scaled by the modal mass m. The displacement y(x, t) at a certain point x of
the structure is a linear combination of the modes q(t), where the coefficients of the linear combination
are the modal shapes s(x) at the point x. This is true whether we have a discrete set of points xi or a
continuous domain, although in practice the spatial domain will be always discretized.
In order to construct a modal synthesizer, the first step to perform is to construct a discrete-time
equivalent of the second order oscillator (3.93). We can discretize the differential equation with the
numerical methods examined previously in Sec. 3.5.1. The impulse invariant method yields:
[ ( −αT ) ]
Ts emωr sin(ωr Ts ) z −1
s

H(z) = . (3.94)
1 − [2e−αTs cos(ωr Ts )] z −1 + e−2αTs z −2

The presence of a z −1 factor at the numerator indicates that this is an explicit numerical scheme (there is
no instantaneous dependence on the input). The backward Euler method yields
1
m(Fs2 +2αFs +ω02 )
H(z) = 2Fs (α+Fs ) Fs2
. (3.95)
1− z −1 + z −2
2
Fs +2αFs +ω0 2 2 2
Fs +2αFs +ω0

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-39

0 0 0 0
amplitude resp. (dB)

−10 −10 −10 −10

−20 −20 −20 −20

−30 −30 −30 −30

−40 −40 −40 −40

0 5000 10000 0 5000 10000 0 5000 10000 0 5000 10000
f (Hz) f (Hz) f (Hz) f (Hz)

(a) (b) (c) (d)

Figure 3.20: Amplitude responses of a second order oscillator with constant mass and quality factor,
and ω0 = 2, 4, 8 kHz: continuous-time responses (solid lines) and discrete-time responses (dashed lines)
with (a) impulse invariant method, (b) backward Euler method, (c) backward Euler method with centered
scheme, (d) bilinear transform.

This instead is an implicit numerical scheme. The backward Euler method with centered scheme yields
Ts2 −1
[ m z]
H(z) = . (3.96)
1 + ω02 Ts2 + 2αTs − 2 z −1 + [1 − 2αTs ] z −2

Like in the impulse invariant case, this is an explicit numerical scheme. By looking at the poles of this
discrete-time system one can see that it can become unstable depending on the mechanical parameters
and on the sampling period: the scheme is not unconditionally stable. Finally, the bilinear transform
yields [ ]
1
m(4Fs2 +4αFs +ω02 )
(1 + 2z −1 + z −2 )
H(z) = 2(ω02 −4Fs2 )
. (3.97)
−1 + 4Fs −4αFs +ω0 z −2
2 2
1 + 4F 2 +4αF +ω 2
z 4F 2 +4αF +ω 2 s s 0 s s 0

Like in the case of the backward Euler method, this is an implicit numerical scheme.
The resulting amplitude responses are shown in Fig. 3.20. As expected, the impulse invariant method
exhibits aliasing, the Euler method exhibits warping and numerical damping, the Euler method with
centered scheme tends to become unstable for high omega0 values, and the bilinear transform exhibits
warping (but not numerical damping).

M-3.12
Write a function that computes the filter coefficients of the mechanical oscillator discretized with (a) the impulse
invariant method, (b) the Euler method g1 (z), (c) Euler method with the centered estimate (3.76), and (d) the
bilinear transform. Compare the frequency responses of the resulting discrete-time systems.

M-3.12 Solution
function [B,A]=modal_oscillator(m,alpha,omega,method)

global Fs; Ts=1/Fs;

switch method
case ’impinv’
omegar=sqrt(omegaˆ2-alphaˆ2); eaTs=exp(-alpha*Ts);
B= [0 Ts*eaTs/(m*omegar)*sin(omegar*Ts) 0];

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-40 Algorithms for Sound and Music Computing [v.February 2, 2019]

A= [1 -2*eaTs*cos(omegar*Ts) eaTsˆ2];
case ’euler’
delta= Fsˆ2+2*alpha*Fs+omegaˆ2;
B= [1/(m*delta) 0 0];
A= [1 -2*Fs*(alpha+Fs)/delta Fsˆ2/delta];
case ’eulercenter’
B= [0 Tsˆ2/m 0];
A= [1 (omegaˆ2*Tsˆ2 +2*alpha*Ts -2) (1-2*alpha*Ts)];
case ’bilin’
delta=4*Fsˆ2 +4*alpha*Fs +omegaˆ2;
B= [1/(m*delta) 2/(m*delta) 1/(m*delta)];
A= [1 2*(omegaˆ2 -4*Fsˆ2)/delta (4*Fsˆ2 -4*alpha*Fs +omegaˆ2)/delta];
otherwise error(’unknown numerical method’);
end

3.5.2.4 A modal synthesizer

A simple modal synthesizer can be constructed as a parallel connection of N numerical oscillators. By

choosing a different center frequency ω0 and damping factor α for each oscillator, it is possible to account
for a set of partials and decay times of the resonator spectrum. Moreover, the modal shapes determine
both how a force signal is injected into the modal oscillator and how the modal oscillations are combined

M-3.13
Write a function that computes the output of a modal resonator given an input force signal.

M-3.13 Solution

function y = modal_synth(x,omega,alpha,m,s,in,out,method);

global Fs;
N=length(omega);% it must be size(omega)=size(alpha)=size(m)=N
% it must be size(s,1)=N; 0<in<size(s,2); 0<out<size(s,2);
y=zeros(1,length(x));
for i= 1:N
[B,A]=modal_oscillator(m(i),alpha(i),omega(i),method);
y_i=filter(B,A,s(i,in)*x);
y=y +s(i,out)*y_i;
end

We have assumed that the force distribution is concentrated in a single point, represented by the index
in. We “pick-up” the resonator signal at another point, represented by the index out (like we were
using a contact mike attached to the object at the point out).

The input modal parameters can be chosen to match those of an arbitrary object. Moreover, morphing
between different shapes and material can be obtained by designing appropriate trajectories for these
parameters.

M-3.14
Synthesize the sound of a dissipative string using the modal approach.

M-3.14 Solution

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-41

global Fs; Fs=44100;

slength=8; %sound length (s)

%%%%% Physical parameters for a E3 guitar nylon string %%%%%

mu=5.25e-3;
T = 60; % string tension (N)
L= 0.65; %string length (m)
d1=mu*.65;
d2=-T*9e-8;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

fmax=30; % impulsive force (N)

x=[fmax zeros(1,round(slength*Fs))]; %force signal zero-padded to sound length

N= 120; M=round(L/5e-3); %no. of modes and spatial points (incl.ends)

in=round(M/20); out=round(M/20); %input and output points
[omega,alpha,m,s]=modal_string(L,T,mu,d1,d2,N,M); %compute modal parameters
y=modal_synth(x,omega,alpha,m,s,in,out,’impinv’); %compute output signal

3.5.3 Modal analysis

The modal synthesizer that we have constructed needs to know the modal parameters for the specific
resonator under exam. The question is then how to determine these parameters.
In the case of a discrete system of N point masses with linear interaction forces, modal parameters
are exactly found through standard matrix calculations. Most systems of interest of course do not fit these
assumptions. For some distributed systems, particularly for symmetrical problems with simple boundary
conditions, the partial differential equation describing the system can be solved analytically, giving the
modal parameters. Alternatively, either accurate numerical simulations (e.g. wave-guide mesh methods)
or “real” physical measurements can be used.

3.5.3.1 Simple 1-D shapes

The simplest tractable case is the ideal string: we have already discussed the modal solution in this
case. There are other tractable cases: one interesting example is the ideal bar, with various boundary
conditions. Bars are almost as relevant as strings for musical applications: mallet percussion instruments,
such as the marimba, the xylophone, the vibraphone, and so on, are based on the oscillations of bars.
Transverse vibrations in a bar are due to internal elastic force generated when the bar is bent. One can
show that for a bar with constant cross-section the equation governing the bar transversal displacement
y is the Euler-Bernoulli equation:

∂2y EK 2 ∂ 4 y
(x, t) = − (x, t), (3.98)
∂t2 ρ ∂x4

where E is the Young modulus of the material, K is the radius of gyration,7 and ρ is the volume density.
Note that the fourth-order term is the one that we used to describe a stiff (and dispersive) string. The
modal solutions y(x, t) = s(x)q(t) are in this case

y(x, t) = [A cosh kx + B sinh kx + C cos kx + D sin kx] cos(ωt + ϕ), with k = cω, (3.99)
∫ ∫
7
This would need some explanation. In short: K 2 = S1 z 2 dS, where S = dS is the total cross-section of the bar and
z is the distance from the neutral axis, i.e. the axis along the bar which does not change its length when the bar is bent (at one
side of the neutral axis there is elongation, at the other side there is compression). Everything clear??

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-42 Algorithms for Sound and Music Computing [v.February 2, 2019]

free−free

clamped−free

hinged−hinged

(a) (b)

Figure 3.21: Modal description of the ideal bar: (a) ideal bar with various boundary conditions and (b)
corresponding modes.

√
and where c2 = ωK E/ρ. This modal solution cannot be interpreted in terms of traveling waves,
therefore waveguide methods fall short here, while modal synthesis can be successfully employed.
The constants A, B, C, D as well as the allowed frequencies are determined depending on four
boundary conditions (two at each end). The conditions for a free end are ∂ 2 y/∂x2 = ∂ 3 y/∂x3 = 0
(no torque and no shearing force); those for a supported (hinged) end are y = ∂ 2 y/∂x2 = 0 (no dis-
placement and no torque); and those for a clamped end are y = ∂y/∂x = 0 (no displacement and zero
slope). Three notable examples are shown in Fig. 3.21(a). For these cases, numerical solution of the
equations resulting from boundary conditions yields
√
π2K E [ ]
(free-free) {ωn } = 2
3.0112 , 52 , 72 , . . . , (2n + 1)2 , . . . ,
4L ρ
√
π2K E [ ]
(clamped-free) {ωn } = 2
1.1942 , 2.9882 , 52 , . . . , (2n − 1)2 , . . . , (3.100)
4L ρ
√
2π 2 K E 2
(hinged-hinged) {ωn } = n .
L2 ρ
Note that in the first two cases the frequencies are strongly inharmonic, while in the third case they are
harmonically related: the corresponding lowest modes are shown in Fig. 3.21(b). Mallet percussions
most typically use bars with (approximately) free-free conditions. However in many cases their bars do
not have constant cross-sections, instead their are cut with an arch on the underside in such a way that
the theoretical partials of the free-free series in Eq. (3.100) are shifted and aligned to an almost harmonic
series.

M-3.15
Compute modal parameters for a bar with the three boundary conditions examined here, and with linear dissipa-
tion.

M-3.15 Solution
Like Example M-3.11, but using the modal shapes of the ideal bar.

3.5.3.2 Simple 2-D shapes

The first example of a musically relevant 2-D shape is a rectangular membrane with fixed ends, like the
one depicted in Fig. 3.22(a). The ideal membrane obeys the 2-D D’Alembert equation:
[ 2 ]
∂2z ∂ z ∂2z
σ 2 (x, y, t) = T (x, y, t) + 2 (x, y, t) = T ∇2 z(x, y, t), (3.101)
∂t ∂x2 ∂y

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-43

z
Lx (1,1) (2,1) (1,2) (3,1) (3,2) (3,3)
Ly x
y ... ...

(a) (b)

(0,1) (1,1) (2,1) (0,2) (3,1) (1,2)

z
(4,1) (2,2) (0,3) (5,1) (3,2) (6,1)
x
a
y

(c) (d)

Figure 3.22: Modal description of ideal membranes: (a) ideal rectangular membrane with fixed ends
and (b) corresponding modes; (c) ideal circular membrane with fixed ends and (d) corresponding modes.

where z is the membrane vertical displacement and the constants T and σ are the membrane surface
tension (in N/m) and surface density (in Kg/m2 ). The symbol ∇2 = ∂ 2 /∂x2 + ∂ 2 /∂y 2 stands here for
the 2-dimensional Laplacian operator. Modal solutions z(x, y, t) = s(x) (x)s(y) (y)q(t) are found with
the same procedure used for the ideal string:
√ √ ( ) ( )
2 2
zn,m (x, y, t) = sin kn(x) x sin km
(y)
y cos (ωn,m t + ϕn,m ) ,
Lx Ly
√[ ] [ ]
(3.102)
nπ mπ (x) 2 (x) 2
with kn(x) = , km (y)
= , ωn,m = c kn + kn .
Lx Ly
Note that the modal frequencies ωn,m are not harmonically related in this case. The modal shapes
(x) (y)
sn,m (x, y) = sn (x)sm (y) have straight nodal lines: the lowest modes are shown in Fig. 3.22(b).
A second example, even more relevant for musical applications, is the circular membrane with
fixed ends, like the one in Fig. 3.22(c). In this case the 2-D D’Alembert equation is more conve-
niently written in circular coordinates x = r sin θ and y = r cos θ and the laplacian becomes ∇2 =
∂ 2 /∂r2 + 1/r(∂/∂r) + 1/r2 (∂/∂θ). Accordingly, one looks for modal solutions of the form z(r, θ, t) =
s(r) (r)s(θ) (θ)q(t).
Substituting this into the 2-D D’Alembert equation results in two differential equations for s(r) and
(θ) (r)
s(θ)( ) the angular shapes sm (θ) = cos(mθ). Then for each m, the radial shapes are sn (r) =
. One finds
(r) (r)
Jm km,n r , i.e. they are the first-kind Bessel functions of order m, with radial frequencies km,n . The
(r) (r)
allowed values for km,n are found as usual by imposing that sn = 0 at the fixed boundary, therefore are
determined by the nth zero of Jm . In conclusion the m, n mode has m nodal diameters (determined by
(θ) (r)
the function sm ) and n nodal circles (determined by the function sn ). The lowest modal frequencies
ωn,m are
2.405c
{ωn,m } = [1, 1.594, 2.136, 2.296, 2.653, 2.918, 3.156, 3.501, 3.6, 3.652, 4.06, 4.154] , (3.103)
a
and are highly inharmonic. The corresponding modes are shown in Fig. 3.22(d).
M-3.16
Compute modal parameters for a rectangular and a circular bar with fixed boundary conditions, and with linear
dissipation.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-44 Algorithms for Sound and Music Computing [v.February 2, 2019]

M-3.16 Solution
Like Example M-3.11, but using the modal shapes of the ideal rectangular and circular membrane.

3.5.3.3 Experimental estimation

When the modal solution cannot be written analytically, modal parameters can still be estimated. One
approach is to extract modal data from a recorded audio signal. Various methods are known that can
estimate resonances (center frequencies and quality factors) from a signal: these include linear prediction
techniques and partial tracking techniques examined in Chapter Sound modeling: signal based approaches. However
there are various problems to deal with. First, modal frequencies are often very closely spaced and one
needs high-resolution methods that are able to discriminate nearby resonances. Second, estimates derived
from analysis of a single sound lack information about the spatial shapes of the modes. Third, there are
many inaccuracies related to technical difficulties in the recording of object responses: ideally one should
record the impulse response of an object, for many different excitation points and many different pick-up
points. In practice one will strike the object and record the response in air, with consequent spatially
distributed interactions, and sound radiation through air.
Modal shapes may be observed through more sophisticated measurement devices, e.g. by using
holographic interferometry. A relatively simple experimental technique amounts to place some finely
divided material (e.g. fine sand or flour) on the resonating object (e.g. a plate of arbitrary shape), and
then setting the object into forced oscillation (most typically through mechanical or electromechanical
means) with a sinusoidal driving signal which is tuned to the frequency of the desired mode. As a
consequence one will observe the sand on the object bouncing and moving about, and only at or near
the nodal lines of the mode the sand will be stationary. Thus the sand is either bounced off the object
or else collects at the nodes, forming so-called Chladni patterns (from the name of the german physicist
and musician who first observed nodal patterns through this technique). Variations of this technique have
been commonly used by acoustic instrument makers, especially for the design and construction of the
resonating bodies of violins, guitars, cellos, etc.
An alternative “experimental” approach amounts to simulate the response of an object with finite
difference or finite element methods. This implies spatial discretization, which means that only a finite
amount of modes can be estimated. Moreover, modal data obtained in this way suffers from underesti-
mation of modal frequencies, due to errors introduced by spatial discretization.

3.6 Non-linear physical models

So far in this Chapter we have examined linear models, mostly employed to simulate physical resonators.
However musical oscillators are often strongly non-linear.
Non-linearities must be present for a system to reach stable self-sustained oscillations. As an ex-
ample, self-sustained oscillations in the acoustic bore of a woodwind or brass instrument can only be
explained in terms of a non-linear, persistent excitation mechanism. More precisely, the valve (a single
or double-reed, or the player’s lips) at the bore termination acts as a non-linear element that injects en-
ergy into the system. A very similar description holds for bowed string instruments, where the bow and
its non-linear friction force is the exciting element. In other cases the instrument is non-linearly excited
only for a limited amount of time: a struck string or bar interacts with the hammer or mallet through a
non-linear contact force. Values for the contact time are typically a few milliseconds, and after this short
excitation the system evolution is linear and the oscillations decay away.
Generalizing from the above examples, we may schematize a musical instrument (or any sound-
producing physical system) by means of two main functional blocks, as in Fig. 3.23. The resonator is

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-45

Exciting EXCITER RESONATOR

actions Out
Non−linear Linear
Dynamic Dynamic
System System

Modulating
actions

Figure 3.23: Exciter-resonator interaction scheme for a musical instrument.

where the oscillations actually take place (an acoustic bore, a string, a bar, etc) and is therefore related
to such sound attributes as pitch, spectral envelope, and so on. The exciter controls the way energy is
injected into the system, thus initiating and possibly sustaining the oscillations, and relates in partic-
ular to properties of attack transients. A simple yet striking demonstration of the effectiveness of the
exciter/resonator schematization is provided by mounting a clarinet mouthpiece on a flute.8 The bore
boundary conditions are changed from open-open to closed-open, so that it plays one octave lower, and
the resulting instrument is perceived as a bad sounding clarinet. In other words, the excitation mecha-
nism defines sound identity (“it’s a clarinet”), while the resonator is mostly associated to sound quality
(“it’s a bad clarinet”).
The interaction between the two blocks is a two-way interaction, where the state of each block influ-
ences the other. As an example, the impact force between a hammer and a string depends on the displace-
ments and velocities of both hammer and string, and affects both. Clearly there are also examples where
non-linearities in the excitation are negligible: plucked string instruments can be conveniently treated
as linear systems (strings and instrument body), where the “pluck” is described as a non-equilibrium
initial condition (i.e., the pluck gives a string a non-zero displacement distribution and a null velocity
distribution).
Finally, note that non-linearities are not necessarily related to excitation mechanisms only: even res-
onators, that are assumed to be linear in a first approximation, can exhibit non-linear behaviors. As an
example, when a string vibrates outside the limit of small oscillations its length cannot be anymore as-
sumed to be constant, but varies (together with string tension) during an oscillation cycle: this length- and
tension-modulation mechanism can produce perceivable pitch glides in the sound. Similar considerations
apply to other systems (e.g. non-linear circuit elements).

3.6.1 Non-linear circuits

3.6.1.1 Non-linear capacities

Consider the well known Chua-Felderhoff electrical circuit: this is a RLC circuit, made of a series con-
nection of a resistor R, an inductor L and a capacitor C. The elements R and L are constant, while this
is not the case for C. More precisely, the characteristic of the capacitance is a function of the voltage v,

8
The author has enjoyed a live demonstration with such a “flarinet”, performed by Joe Wolfe while giving a seminar in
Venice, 2000.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-46 Algorithms for Sound and Music Computing [v.February 2, 2019]

−10 −11
x 10 x 10
1.5 25 non−linear
linearized
20

1 15

q [C]
C [F]

10

5
0.5
0

−5
0
0 2 4 6 8 10 12 0 2 4 6 8 10 12
v [V] v [V]
(a) (b)

Figure 3.24: Non-linear behavior of (a) capacitance C(v) and (b) charge q(v) in the Chua-Felderhoff
circuit.

so that the system is described as follows:

( √ )
1 C0
v(q) = 2 2 2 2
q + q q + 4v0 C0 , ⇔ C(v) = √ ,
2v0 C02 v
1+ (3.104)
v0
v(q) + Rq̇(t) + Lq̈(t) = ve (t), (v > v0 ).

The variable q(t) stands for the charge on the capacitor, and ve (t) is an applied voltage. Note that
C(v) ∼ C0 when v → 0, i.e. the system is a linear RLC circuit in the limit of small oscillations.
However, for larger voltage v this approximation does not hold, and C(v), q(v) behave as depicted in
Fig. 3.24(a) and (b), respectively. There is no easy way to translate the non-linear relation (3.104) into
the Laplace domain, because the definition of impedance given in Sec. 3.2.1 assumes linearity of the
circuit elements.
The Chua-Felderhoff circuit has been extensively studied and is one of the classical systems used
for exemplifying transition to chaotic behavior: when the peak of the voltage generator is increased, the
behavior of the charge q(t) on the capacitor undergoes successive bifurcations.

3.6.1.2 Vacuum tubes

....................................

3.6.2 Mechanical interactions

3.6.2.1 Impacts
Several musical and non musical classes of sounds are produced by a single impact of two objects, one of
which (at least) resonates as a consequence of the collision. Moreover, impact is at the basis of other more
complex mechanical contacts: as an example, scraping and rolling can be seen as temporal sequences of
micro-impacts between non-smooth surfaces.
The ideal impact is a force signal shaped like a Dirac delta in time. It imparts to the resonator an
ideal force impulse in an infinitesimal time. If the resonator is initially at rest, such force impulse imparts
to the resonator initial conditions given by null initial displacement and a non-zero initial velocity whose
magnitude depends on the magnitude of the delta.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-47

4 1400
compression velocity (m/s)

1200
2

contact force (N)

1000
0
800

−2 600

400
−4
200

−6 0
−2 −1 0 1 2 0 0.2 0.4 0.6 0.8 1 1.2
compression (m) −4
x 10 compression (m) x 10
−4

(a) (b)

Figure 3.25: The non-linear impact model (3.106): (a) phase portrait of a point mass hitting a hard
surface; (b) the corresponding non-linear force during impact.

In a less ideal impact model one would assume that the impact force is non-null over a finite duration
or time (the contact time between the colliding objects) and takes finite values. The force magnitude is
related to the impact energy (e.g. the impact velocity of the hammer hitting the resonator), while the
contact time is related to the hardness of the impact. A simple signal model of the impact force is the
following:
{
fmax [ ]
2 1 − cos( 2πt
τ ) , 0 ≤ t ≤ τ,
f (t) = (3.105)
0, otherwise,

where τ is the contact time and fmax is the maximum force value.
More complex models must take into account other effects. There is dissipation of energy during
contact. The contact force itself is a function of the relative compression x(t) between the two contacting
objects (which may be thought as the difference between the displacements of the two objects during the
contact), and also of the compression velocity v(t) = ẋ(t). Accordingly, a more physically-based model
of the impact force is the following:
{
kx(t)α + λx(t)α v(t), x > 0,
f (x(t), v(t)) = (3.106)
0, otherwise,

where k is the force stiffness, λ is the force damping weight, and the exponent α depends on the local
geometry around the contact area. As an example, according to Herz theory of contact an ideal impact
between two spherical objects obeys this equation with α = 3/2 and λ = 0.
Figure 3.25(a) depicts the simulation of a point mass hitting a rigid surface with the impact model (3.106:
the phase portrait shows that due to dissipation the mass velocity after the impact is always lower in
magnitude than the initial impact velocity, and converges to a limit value. Figure 3.25(a) shows the
corresponding impact force: it has a non-linear characteristics that depends on the exponent α, and it
exhibits a hysteresis effect that is associated to to the dissipative component λxα v. This plot is qualita-
tively resemblant of what one would observe by measuring the contact force during a real impact of a
small mass againts a rigid surface.

M-3.17
Simulate a modal oscillator excited by the non-linear impact force (3.106).

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-48 Algorithms for Sound and Music Computing [v.February 2, 2019]

0.3 fs

0.2
friction force (N)

0.1 fc string
0

−0.1

−0.2

−0.3

bow
−0.2 −0.1 0 0.1 0.2 0.3
relative velocity (m/s)
(a) (b)

Figure 3.26: Stick-slip friction: (a) example of parametrization of a kinetic (static) friction curve; (b)
Helmoltz motion resulting from stick-slip ideal string-bow interaction.

3.6.2.2 Stick-slip friction

Stick-slip friction is a second relevant mechanical interaction in sound production. A typical musical
example is the interaction between bow and string in a violin. A non musical example is the sound
produced by a finger rubbing on a moisty window or on a glass.
We know from physics that static friction is higher than dynamic friction: the simplest model as-
sumes that the friction force is proportional to the normal force fN between two contacting objects, but
the coefficient of proportionality is higher if there is no relative motion and is lower if there is relative
motion. More refined models define the coefficient of proportionality as a function of the relative veloc-
ity. These are called kinetic models (as the friction force is assumed to be a function of velocity only), or
static models (since the force-velocity dependence is derived under stationary conditions). One possible
parametrization of a kinetic friction force model is:
{ [ ]
sgn(v) fc + (fs − fc )e−(v/vs ) , fN > 0,
2

f (v(t)) = (3.107)
0, otherwise,

where fc , fs are the Coulomb force and the stiction (short for static friction) force respectively, while
vs is named Stribeck velocity. The Coulomb force and the stiction force are related to the normal force
through the equations fs = µs fN and fc = µd fN , where µs and µd are the static and dynamic friction
coefficients. If fN ≤ 0 this means that there is no contact. The dependence of the friction force on
velocity, as given in Eq. (3.107), is shown in Fig. 3.26(a).
When two objects in relative motion interact through a friction force of this kind, a stick-slip phe-
nomenon is generated in which the two objects remain in static contact for a certain amount of time (the
“stick” phase) and suddenly detach (the “slip” phase). Sound generation occurs when this alternation
of stick and slip phases occurs in an almost periodic fashion and with an audio rate, typically locked to
some of the proper resonance frequencies of the interacting objects.
An example of stick-slip interaction is the Helmoltz motion occurring in an ideal, rigidly terminated,
bowed string (see Fig. 3.26(b)). Assuming the bow to be perfectly rigid and to be in contact with the
string in a single point, the string motion at the contact point is a sawtooth signal in which the string
remains stuck to the bow hair for a considerable fraction of each vibratory cycle, and slips back abruptly
when its displacement becomes large enough, to begin the next cycle. In normal playing condition the
resulting frequency of oscillation is almost coincident with the first-mode frequency of the string. Further

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-49

y(x,t)+y’(x,t)dx

dL y’(x,t)dx
y(x,t)

dx

Figure 3.27: Length dL of a string at point x over the segment dx.

analysis of this Helmoltz motion would reveal that at every instant the shape of the string consists of two
line segments joined by a corner, and the corner travels on an envelope composed of two parabolas.

M-3.18
Simulate a modal oscillator excited by the non-linear friction force (3.107).

More refined, dynamic friction models include some “memory”. The dependence of friction on the
relative sliding velocity is modeled using a differential equation. These models are able to take into
account presliding behavior, where the friction force increases gradually for small displacement val-
ues. Static and dynamic friction models have the same behavior at high or stationary relative velocities,
but dynamic models provide more accurate simulation of transients, which is particularly relevant for
realistic sound synthesis.

3.6.2.3 Tension modulations

The phenomenon of tension modulation is qualitatively different from the previous examples. This non-
linear effects is not generated from an external excitation force. It is a non-linear correction to the
D’Alembert equation when the limit of small oscillations of the elastic medium is not valid.
The simplest example of tension modulation is encountered in a vibrating string with fixed ends.
When the string is significantly displaced from equilibrium, its length and therefore also its tension are
increased. When it returns closer to its equilibrium state, its length and tension are decreased. Clearly
the rate of this tension modulation is twice the rate of the transversal vibration, since minimum tension
occurs at equilibrium, and maximum tension occurs at both extreme diplacements.

∂2p ∂2p ∂4p ∂p ∂3p

µ (x, t) − T [y(x, t)] (x, t) + EI (x, t) + d 1 (x, t) − d2 (x, t) = 0, (3.108)
∂t2 ∂x2 ∂4x ∂t ∂t∂ 2 x
where T [y(x, t)] is the string tension and is now a function of the string displacement. More precisely, it
is proportional to the string length, which in turns depends on y(x, t). From the √ theorem of Pithagoras,
the length dL at point x over the segment dx is (see Fig. 3.27) dL[y(x, t)] = dx2 + (y ′ (x, t)dx)2 .
Then the total string length deviation ∆L from the length L0 at equilibrium is
∫ L0 ∫ L0 √
∆L[y(x, t)] = L[y(x, t)] − L0 = dL[y(x, t)] − L0 = 1 + y ′ (x, t)2 dx − L0 . (3.109)
0 0

Then the tension is

EA∆L[y(x, t)]
T [y(x, t)] = T0 + , (3.110)
L0
where A is the string section.
Tension modulation in a waveguide model can be simulated by using all-pass filters with time-varying
coefficients that account for length modulation. Tension modulation in modal synthesis can be simulated

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-50 Algorithms for Sound and Music Computing [v.February 2, 2019]

mouth

y
L
p to
m p bore
u
y max

Reed
lip
h
L

Figure 3.28: Schematic representation of the reed-mouthpiece system.

by finding the modes of Eq. (3.108). As an example, in the simple case where d2 = E = 0 in Eq. (3.108),
one can find the modes
[ ]
∑
+∞
q̈i (t) + d1 q̇(t) + i2 ω02 + ω12 l2 ql2 (t) qi (t) = 0, i = 1, . . . , +∞, (3.111)
l=1

that can be interpreted as mechanical oscillators in which the frequency of oscillation depends on the
modal displacement. In the general case of Eq. (3.108) including dispersion and frequency-dependent
dissipation, a similar modal description can still be found.

3.6.3 Acoustic interactions

3.6.3.1 Jets
.....................

3.6.3.2 Quasi-static reeds

Reeds are acoustic valves that oscillate due to pressure differences at the two sides. The simplest example
is the single reed, schematically represented in Fig. 3.28. The reed dimensions are small with respect to
typical wavelengths in the resonator, thus pressure can be thought of as constant along the reed surfaces;
under normal playing conditions, the first mode of oscillation of the reed is well above the main frequency
components of the pressure signal in the resonator. Oscillations occur mainly in the vertical direction,
and a single degree of freedom can be reasonably assumed, i.e. the vertical displacement yL of the reed
tip from the equilibrium. These considerations justify the choice of a lumped modeling approach for the
reed.
The simplest posssible lumped model regards the reed as a system with stiffness only, neglecting in-
ertia and damping properties. In this approximation the reed moves in phase with the pressure difference
∆p(t) = pm (t) − p(t) across the reed:

kyL (t) = Sd ∆p(t) ⇒ ka yL (t) = ∆p(t), (3.112)

where pm is the pressure inside the performer’s mouth, p is the (oscillating) acoustic pressure inside
the instrument bore, k is the effective reed stiffness, Sd is an effective driving surface on which the
pressure ∆p acts, and ka = k/Sd is the stiffness per unit area. Equation (3.112) is called a quasi-static
approximation since it can be determined experimentally in static conditions where a constant pressure

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-51

−4
x 10
1.2
1000
1

p+ [Pa]
0.8 500
u [m3/s]

0.6
0
0.4

0.2 −500
0
0 500 1000 1500 2000 2500 3000 3500 −600 −400 −200 0 200 400 600 800
∆ p [Pa] p− [Pa]
(a) (b)

Figure 3.29: Quasi-static approximation of a single reed; (a) u versus ∆p and (b) rotated mapping p+
versus p− .

difference δp is injected into the system and the corresponding constant displacement yL is measured
after an initial transient.
As far as aerodynamics is concerned, the relation between the reed opening h(t), the airflow u(t)
through the slit, and the pressure drop ∆p(t) can be approximated through the equation

ρair | u(t) | 2
∆p(t) = f (u(t), h(t)) = sgn[u(t)] , (3.113)
2 wh(t)

where w is the reed width. This equation is derived from the Bernoulli law.9 Using Eq. (3.112), the
reed opening h is computed as h = ymax − yL = ymax − ∆p/ka , and by substituting this relation into
Eq. (3.113) one finds
 ( )√
w sgn[∆p(t)] y − ∆p(t) 2| ∆p(t) |
, ∆p < ka ymax ,
max ka ρair
u(t) = (3.114)
0, otherwise.

Figure 3.29(a) shows the plot of the resulting relation between u and ∆p. For low ∆p values, u increases
until a maximum is reached at ∆p = ka ymax /3. For higher ∆p values, the flow starts to drop due to reed
closure, and reaches the value u = 0 at ∆p = ka ymax . Beyond this value the reed is completely closed.
This non-linear map can be used to construct a quasi-static reed model. If wave variables p± are
introduced in the cylindrical bore, i.e. p = p+ + p− and u = p+ − p− , then these relations can be
substituted into Eq. (3.114). As a consequence this non-linearity can be turned in a new one in which
p+ depends on p− through a non-linear reflection function Rnl , i.e. p+ = Rnl (p− ). This is depicted in
Fig. 3.29(b).
Despite its simplicity, the quasi-static model is able to capture the basic non-linear mechanisms
of self-sustained oscillations in a single reed instrument. Due to its compactness and low number of
parameters, this model has been also used for sound synthesis purposes.

M-3.19
9
The Bernoulli law holds for incompressible non-viscous fluids in stationary conditions, and states the relation u = A · x ·
∆p1/2 sgn(∆p) between the flow u and the pressure difference ∆p through an aperture of width x. Some authors adopt for the
single reed the generalized equation u = [A · x∆p1/2 sgn(∆p)]1/α , with an experimentally determined value α = 3/2.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-52 Algorithms for Sound and Music Computing [v.February 2, 2019]

Write a function that computes the pressure wave p+ [n] reflected into the bore from the wave p− [n] arriving from
the bore, according to the quasi-static model (3.114). Implement a quasi-static clarinet model in which the quasi-
static reed is coupled to a waveguide cylindrical bore and driven by a mouth pressure signal pm . The bell can
be modeled as a low-pass filter, that radiates frequencies above its cut-off (typically around 1500 Hz) and reflects
low frequencies back inside the bore.

M-3.19 Solution
Further refinements to this model should include propagation losses, fractional-delay filters in order
to allow for fine tuning of the bore length, and acoustic holes modeled as scattering filters connected
through 3-port junctions to the main waveguide structure.

3.6.3.3 Dynamic reeds

More refined reed models need to take into account the dynamics of the reed. A reasonably accurate
description is obtained through a second-order mechanical oscillator, driven by the pressure drop ∆p
between mouth and mouthpiece:
{
mÿL (t) + rẏL (t) + k(yL (t) − y0 ) = Sd ∆p(t), yL < ymax ,
(3.115)
yL (t) = ym and ẏL (t) = 0, yL ≥ ymax ,
where m and r represent the reed mass and damping, while other parameters and variables are defined
as before. The constant y0 represents the reed displacement at rest.
This modeling approach is reasonable for the same reasons mentioned before: small reed dimensions
compared to typical wavelengths in the resonator, reed oscillation mainly in the vertical direction, and
high frequencies of the transversal reed modes (only the first mode is relevant). Note that in Eq. (3.115)
the phenomenon of reed beating (i.e. complete closure of the reed) is here incorporated in the lumped
model in a non-physical way, by imposing an ideal “stop” when the reed tip reaches its maximum al-
lowed displacement ym . Note also that the quasi static approximation examined in the previous section
corresponds to approximating the transfer function of this system with its value at 0 frequency.
Another refinement amounts to taking into account an additional component affecting the total flow
inside the instrument: the reed motion generates the flow Sd ẏL (t), proportional to the reed tip velocity.
If we now call u the flow inside the instrument and uf the flow entering from the slit, these are related
through the following equation:
u(t) = uf (t) + ur (t), with ur (t) = Sr ẏL (t). (3.116)
Incorporating this dynamics into the model results in more convincing sound synthesis, especially as
far as transients are concerned. Realistic effects can be obtained, such as transitions to high regimes of
oscillation. Both the resonance and the damping of the reed oscillator (3.115) and g play a role in helping
transition to the second register of a single reed instrument. As an example, the clarion register in the
clarinet plays a twelfth above the fundamental register and is usually obtained with the aid of a register
hole. Howver the clarion register can be produced also without opening the register hole if the reed
resonance matches a low harmonic of the playing frequency and the damping is small enough. Another
playing regime in single reed instruments is the so-called reed regime (“squeaks”): this can be obtained
by imparting an extremely low damping to the reed oscillator, so that the oscillation is governed by the
reed resonance.

M-3.20
Implement a dynamic clarinet model in which the dynamic reed is coupled to a waveguide cylindrical bore and
driven by a mouth pressure signal pm .

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-53

Finally, additional degrees of freedom must be taken into account when simulating other types of
reeds. Double reeds (such as those found in oboe and bassoon) are composed of two reeds that oscillate
independently, and even if one assumes perfect symmetry of oscillation the flow model differs from the
one examined previously, due to the smallness of the aperture. In so-called lip reeds the role of the reed
is taken by the performer’s lips, that are constrained into the mouthpiece and vibrate at the fundamental
frequency: at least two degrees of freedom are needed to simulate lip vibration.

3.6.4 Computability issues

We have examined in Sec. 3.5.1 the concept of delay-free loop in the case of linear systems, and have
mentioned some strategies for dealing with it. However, more severe computability problems can arise
when simulating non-linear elements.

3.6.4.1 Non-linear systems and delay-free loops

It should be clear that in the non-linear case one cannot perform a rearrangement such as in (3.79),
because a non-linear equation is not always analytically invertible. The question is then how to deal with
the delay-free loop problem in the non-linear case.
One can use an explicit numerical method, that produces a system of difference equations in which
there are no delay-free loops. This choice solves the computational problem but can introduce more
severe artifacts in the numerical system: explicit methods have lower orders of accuracy with respect
to implicit methods, and more importantly are not unconditionally stable, i.e. are not stable for any
sampling frequency Fs and for any values of the system parameters.
A rudimentary solution, that is nonetheless often encountered in the literature of physical modeling,
amounts to inserting fictitious delay elements z −1 in the computational scheme. In practice this is a
variant of the previous approach: instead of using an explicit method from the beginning, one makes
the computation explicit a posteriori, through the insertion of delay elements. While this “trick” can be
acceptable at significantly high sampling rates, the insertion of delay elements can again deteriorate the
accuracy and stability properties of the numerical system. Even worse, in this case one cannot determine
analytically the stability range of the system.

3.6.4.2 Iterative methods

Numerical analysis provides iterative methods to find solutions of non-linear systems of algebraic equa-
tions: examples of such methods include fixed-point iteration and Newton iteration, and each of them
requires specific hypothesis on the non-linear system to hold.
Using an iterative solver is advantageous over the previous approaches in that one can exploit the
accuracy and stability properties of an implicit method without introducing additional numerical errors
in the system. One major drawback, however, is that one does not know in advance the number of
iterations that are needed for the solver to converge to the solution y[n]: this can be a problem for
real-time applications, where one wants to know the time needed to compute one sound sample.
See [Fontana and Avanzini, 2008] for details about Newton-Raphson and fixed-point iteration for the
simulation of non-linear systems.

3.6.4.3 Sheared non-linearities

In many practical cases the delay-free loop problem takes the form of the implicit dependence

y[n] = f (x̃[n] + Ky[n]), (3.117)

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-54 Algorithms for Sound and Music Computing [v.February 2, 2019]

original (f)
k= 1.5 1.2 0.9 0.6 0.3 sheared (h)
1

0.8

y 0.6

0.4

0.2

0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
x

Figure 3.30: Shear transformation of f (x) = e−x for various k values.

2

where f is a non-linear function, and x̃[n] is a vector of variables that are known at time n. The variables
y[n] depend instantaneously onto themselves in the above equation. If one could turn this implicit de-
pendence into a new explicit dependence y[n] = h(x̃[n]), this would solve the delay-free loop problem.
This is achieved using the implicit mapping theorem. Define the function g as

g(x̃, y) = f (x̃ + Ky) − y, (3.118)

and assume that there is a point (x̃0 , y 0 ) such that g(x̃0 , y 0 ) = 0. Moreover, assume that the following
condition holds [ ]
gi
det[Jy (g)(x̃0 , y 0 )] = det (x̃0 , y 0 ) ̸= 0, (3.119)
yj i,j

where Jy (·) denotes the Jacobian matrix with respect to the y variables. From the definition of g, it is
seen that Jy (g) = Jx (f )K − I. Therefore, condition (3.119) implies that the matrix [Jx (f )K − I]
must be non-singular at the point (x̃0 , y 0 ). If these conditions are fulfilled, then the implicit mapping
theorem states that a function h(x̃) exists locally (i.e. for points x̃ in a neighborhood of x̃0 ), with the
properties
h(x̃0 ) = y 0 and g(x̃, h(x̃)) = 0. (3.120)
If the above conditions are fulfilled globally rather than in a neighborhood of (x̃0 , y 0 ), then h is defined
globally.
A few geometrical considerations can help understanding the shape of the new function h. Consider
the coordinate transformation [ ] [ ] [ ]
x̃ I −K x
= · . (3.121)
y 0 I y
This defines a shear that leaves the y axes unchanged and distorts the x axis into the x̃ axis. The plot of
the function y = f (x) “lives” in the (x, y) space. Then the plot of y = h(x̃) is obtained by applying
the coordinate transformation (3.121), and is therefore a sheared version of the former.
In order to understand this shear effect, consider the following example with a scalar function f :
R → R:
y[n] = f (x[n]) = e−(x[n] ) ,
2
with x[n] = x̃[n] + ky[n]. (3.122)
Condition (3.119) translates in this case in the condition f ′ (x) ̸= 1/k, which has a straightforward geo-
metrical interpretation: the shear transformation defined in Eq. (3.121) is such that the vector [x, y]T =

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-55

[k, 1]T (i.e. a point with tangent 1/k) is transformed into the vector [x̃, y]T = [0, 1]T (i.e. a point with
vertical tangent). This explains why the derivative of f cannot equal 1/k.
Figure 3.30 shows the original function f (x), together with the sheared one h(x̃), for various k val-
ues. It can be seen that the horizontal coordinate is distorted when applying the shearing transformation.
Moreover, note that for k = 1.5 the new function h(x̃) cannot be defined globally, because the condition
f ′ (x) ̸= 1/k is not fulfilled globally in this case.

M-3.21
Simulate a modal oscillator excited by the non-linear impact force f (x(t) = kx(t)α (i.e. the impact model (3.106)
with λ = 0) as follows: use an implicit numerical scheme (e.g. the bilinear transform), find the implicit dependence
in the form (3.117), and construct the corresponding sheared non-linear function.

3.7 Commented bibliography

Sound modeling techniques can be classified according to many criteria. Two general references that
address these issues are [De Poli, 1991, Smith III, 1991]. Specifically, the taxonomy based on signal
models and source models, and their subclasses, proposed at the beginning of this chapter is based on
[De Poli, 1991].
Seminal ideas that eventually lead to the definition of physically-based sound modeling techniques
are to be found in research on musical instrument acoustics. Some classic papers in this area are [Hiller
and Ruiz, 1971a,b, Schumacher, 1981, McIntyre et al., 1983]. In particular, the two citations at the
beginning of the Introduction are taken from Hiller and Ruiz [1971a], McIntyre et al. [1983], respectively.
A book that covers the topic of musical acoustics exhaustively is [Fletcher and Rossing, 1991]. In
particular our discussion of the analogies between electrical, mechanical, and acoustic systems, given in
Sec. 3.2.1 is based on an analogous discussion in [Fletcher and Rossing, 1991, Ch.1].
A general overview on approaches and techniques used in physical modeling, with an emphasis on
structural and computational aspects, is provided by De Poli and Rocchesso [1998]. Figure 3.23 in
this chapter (typical block scheme of a musical instrument model) is based on an analogous scheme
in [De Poli and Rocchesso, 1998]. Two more recent and very complete tutorial papers on the topic of
physical modeling are [Smith III, 2004] and [Välimäki et al., 2006].
About waveguide modeling approaches. The theory of 1-D waveguide models is now well estab-
lished. An exhaustive introduction to the topic is given by Smith III [1998], who provides full derivations
of waveguide structures and examples of musical instrument modeling, together with a vast bibliography.
A more recent and even more exhaustive overview is given by the same author in [Smith III, 2008].
The basic principles of waveguide models were already present in the work of Kelly and Lochbaum
[1962] on speech synthesis, where a so-called “transmission-line modeling” approach was used to simu-
late the human vocal tract through delay lines and scattering junctions. The definition of “digital waveg-
uide modeling” was introduced later by Smith III [1985] in the context of musical applications, because
of an analogy to the concept of waveguide that has been used, for example, in microwave technology.
The Karplus-Strong algorithm, which we have regarded as the first step toward the development of digi-
tal waveguide structures, was originally proposed by Karplus and Strong [1983]. Fractional-delay filters:
detailed discussion is provided by Laakso et al. [1996]. Modeling of dissipation and dispersion: Bank
[2006] discusses the topic at length, with application to physically-based synthesis of the piano. In par-
ticular, the frequency-dependent dissipation model reported in Eq. (3.48) was first proposed by Bensa
et al. [2003], although in the context of finite-difference simulations, as an improvement of the dissipa-
tion model by Hiller and Ruiz [1971a]. About waveguide junctions. Many textbooks on digital speech
processing discuss multitube lossless models of the vocal tract, which are basically cylindrical waveguide
sections connected by Kelly-Lochbaum junctions: see e.g. [Deller et al., 1993]. We have not addressed

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-56 Algorithms for Sound and Music Computing [v.February 2, 2019]

the topic of higher dimensional (2- and 3-D) waveguide structures: seminal ideas were first presented by
van Duyne and Smith III [1993].

About lumped modeling approaches. Numerical and computational aspects: most of the techniques
described in Sec. 3.5.1 are found in DSP textbooks: see e.g. [Mitra, 2005]. In the field of numerical
analysis, a comprehensive discussion on numerical methods for ordinary differential equations is given
by Lambert [1993]. The example illustrated in Fig. 3.17 about delay-free computational paths in linear
systems is adapted from [Mitra, 2005, Sec. 6.1.3, Fig. 6.5]. A classic reference to the theory of Wave
Digital Filters (WDF) theory is [Fettweis, 1986].

Finite difference schemes have been applied to also to the explicit numerical simulation of partial dif-
ferential equations, e.g. for modeling idiophones [Chaigne and Doutaut, 1997] and single reed systems
[Stewart and Strong, 1980]. A recent book about the applications of finite difference methods to numer-
ical sound synthesis is [Bilbao, 2009], which discusses the fundamentals of finite differences and shows
how they can be employed to simulate strings, bars, plates, membranes, acoustic tubes. Among other
lumped modeling approaches, in the early nineties Cadoz and coworkers have introduced the CORDIS-
ANIMA model [Florens and Cadoz, 1991], which describes vibrating bodies as a set of interconnected
mass-spring-damper cells.

Modal synthesis. A classic presentation of modal synthesis techniques is [Adrien, 1991]. Cook
[1997] developed a series of “physically-informed” approaches to the modeling of percussion sounds,
which are based on a modal description. The use of modal sound synthesis to virtual reality applica-
tions is discussed in [van den Doel and Pai, 2004]. A corpus of relevant contributions in this field has
been provided by Rabenstein and coworkers [Trautmann and Rabenstein, 2003], who have proposed
the so-called functional transformation method (FTM): in essence, the method exploits the existence of
an analytical form of the modal parameters for a set of relevant multidimensional differential systems,
including strings and membranes with various boundary conditions. Our examples of modal analysis
for simple 1-D and 2-D shapes is based on [Fletcher and Rossing, 1991, Ch.2-3]. The same book also
shows experimental results of modal analysis on several musical instruments, including modal shapes
and Chladni patterns. In addition to linear prediction techniques and partial tracking methods, already
discussed in Chapter Sound modeling: signal based approaches, a method for high-resolution estimate of modal
parameters from sound analysis has been proposed in [Esquef et al., 2003].

About non-linear physical models. The non-linear impact model of Eq. (3.106) was first proposed
by Hunt and Crossley [1975]. . Concerning stick-slip friction models, an overview of traditional models
in the context of sound synthesis applications (bowed strings) is provided by Serafin [2004]. More com-
plex dynamic stick-slip models, typically used in the literature of automatic control, have been recently
applied to sound synthesis by Avanzini et al. [2005]. We have seen that the reed mechanism is that of
pressure-controlled valves: a classic paper on the topic is [Fletcher, 1993]. The quasi-static single reed
examined in Sec. 3.6.3 was first studied by Schumacher [1981] and has been used extensively in the
literature. Other types of reeds: for the double reed see [Guillemain, 2004], for the lip reed see [Adachi
and aki Sato, 1996]. Lip reeds have some similarities with vocal fold functioning: a classic example of a
vocal fold model applied to voice synthesis is [Ishizaka and Flanagan, 1972].

We have seen that new problems are encountered when non-linear elements are present in the delay-
free computational path: Borin et al. [2000] provides a discussion of these issues, together with a pro-
posed non-iterative solution (in brief, a set of hypotheses and techniques to pre-compute a “sheared”
non-linear function that makes the numerical scheme computable), and applications to the simulation of
acoustic systems.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Chapter 3. Sound modeling: source-based approaches 3-57

References
Seiji Adachi and Masa aki Sato. Trumpet Sound Simulation Using a Two-dimensional Lip Vibration Model. J. Acoust. Soc.
Am., 99(2):1200–1209, Feb. 1996.

Jean-Marie Adrien. The missing link: Modal synthesis. In Giovanni De Poli, Aldo Piccialli, and Curtis Roads, editors,
Representations of Musical Signals, pages 269–297. MIT Press, Cambridge, MA, 1991.

Federico Avanzini, Stefania Serafin, and Davide Rocchesso. Interactive simulation of rigid body interaction with friction-
induced sound generation. IEEE Trans. Speech Audio Process., 13(6):1073–1081, Nov. 2005.

Balasz Bank. Physics-based Sound Synthesis of String Instruments Including Geometric Nonlinearities. PhD thesis, Budapest
University of Technology and Economics, Dep. of Measurement and Information Systems, Budapest, 2006.

Julien Bensa, Stefan Bilbao, Richard Kronland-Martinet, and Julius O. Smith III. The simulation of piano string vibration:
From physical models to finite difference schemes and digital waveguides. J. Acoust. Soc. Am., 114(2):1095–1107, Aug.
2003.

Stefan Bilbao. Numerical sound synthesis - Finite difference schemes and simulation in musical acoustics. John Wiley & Sons,
Chichester, 2009.

Gianpaolo Borin, Giovanni De Poli, and Davide Rocchesso. Elimination of delay-free loops in discrete-time models of nonlin-
ear acoustic systems. IEEE Trans. Speech Audio Process., 8(5):597–606, Sep. 2000.

Antoine Chaigne and Vincent Doutaut. Numerical Simulations of Xylophones. I. Time-domain Modeling of the Vibrating Bar.
J. Acoust. Soc. Am., 101(1):539–557, Jan. 1997.

Perry R. Cook. Physically informed sonic modeling (PhISM): Synthesis of percussive sounds. Computer Music J., 21(3):
38–49, 1997.

Giovanni De Poli. A Tutorial on Digital Sound Synthesis Techniques. In Curtis Roads, editor, The Music Machine, pages
429–447. MIT Press, 1991.

Giovanni De Poli and Davide Rocchesso. Physically Based Sound Modelling. Organized Sound, 3(1):61–76, Apr. 1998.

John R Deller, John G. Proakis, and John. H.L. Hansen. Discrete-Time Processing of Speech Signals. Macmillan, New York,
1993.

Paulo A. A. Esquef, Matti Karjalainen, and Vesa Välimäki. Frequency-zooming arma modeling for analysis of noisy string
instrument tones. EURASIP Journal on Applied Signal Processing, 2003(10):953–967, 2003.

Alfred Fettweis. Wave Digital Filters: Theory and Practice. Proceedings of the IEEE, 74(2):270–327, Feb. 1986.

Neville H. Fletcher. Autonomous Vibration of Simple Pressure-Controlled Valves in Gas Flows. J. Acoust. Soc. Am., 93(4):
2172–2180, Apr. 1993.

Neville H. Fletcher and Thomas D. Rossing. The physics of musical instruments. Springer-Verlag, New York, 1991.

Jean Luc Florens and Claude Cadoz. The physical model: modeling and simulating the instrumental universe. In Giovanni
De Poli, Aldo Piccialli, and Curtis Roads, editors, Representations of Musical Signals, pages 227–268. MIT Press, Cam-
bridge, MA, 1991.

Federico Fontana and Federico Avanzini. Computation of delay-free nonlinear digital filter networks. Application to chaotic
circuits and intracellular signal transduction. IEEE Trans. Sig. Process., 56(10):4703–4715, Oct. 2008.

Philippe Guillemain. A digital synthesis model of double-reed wind instruments. EURASIP Journal on Applied Signal Pro-
cessing, 2004(1):990–1000, Jan. 2004.

Lejaren Hiller and Paul Ruiz. Synthesizing Musical Sounds by Solving the Wave Equation for Vibrating Objects: Part I. J.
Audio Eng. Soc., 19(6):462–470, June 1971a.

Lejaren Hiller and Paul Ruiz. Synthesizing Musical Sounds by Solving the Wave Equation for Vibrating Objects: Part II. J.
Audio Eng. Soc., 19(7):542–551, July 1971b.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
3-58 Algorithms for Sound and Music Computing [v.February 2, 2019]

Kenneth H. Hunt and F. R. Erskine Crossley. Coefficient of restitution interpreted as damping in vibroimpact. ASME J. Applied
Mech., 42:440–445, June 1975.

Kenzo Ishizaka and James L. Flanagan. Synthesis of voiced sounds from a two-mass model of the vocal cords. Bell Syst. Tech.
J., 51:1233–1268, 1972.

Kevin Karplus and Alexander Strong. Digital Synthesis of Plucked String and Drum Timbres. Computer Music J., 7(2):43–55,
1983.

John L. Kelly and Carol C. Lochbaum. Speech synthesis. In Proc. 4th Int. Congr. Acoustics, pages 1–4, Copenhagen, Sep.
1962.

Timo I. Laakso, Vesa Välimäki, Matti Karjalainen, and Unto K. Laine. Splitting the Unit Delay Tools for Fractional Delay
Filter Design. IEEE Signal Processing Magazine, 13(1):30–60, Jan. 1996.

John D. Lambert. Numerical Methods for Ordinary Differential Systems. John Wiley & Sons, 1993.

Michael E. McIntyre, Robert T. Schumacher, and James Woodhouse. On the Oscillations of Musical Instruments. J. Acoust.
Soc. Am., 74(5):1325–1345, Nov. 1983.

Sanjit K Mitra. Digital Signal Processing. McGraw-Hill, third edition, 2005.

Robert T. Schumacher. Ab Initio Calculations of the Oscillations of a Clarinet. Acustica, 48(2):71–85, 1981.

Stefania Serafin. The sound of friction: real-time models, playability and musical applications. PhD thesis, Stanford University,
Center for Computer Research in Music and Acoustics, Stanford, 2004.

Julius O. Smith III. A new approach to digital reverberation using closed waveguide networks. In Proc. Int. Computer Music
Conf. (ICMC’85), pages 47–53, Vancouver, 1985.

Julius O. Smith III. Viewpoints on the History of Digital Synthesis. In Proc. Int. Computer Music Conf. (ICMC’91), pages
1–10, Montreal, Oct. 1991.

Julius O. Smith III. Principles of digital waveguide models of musical instruments. In Mark Kahrs and Karl-Heinz Brandenburg,
editors, Applications of Digital Signal Processing to Audio and Acoustics, pages 417–466. Kluwer Academic Publishers,
New York, Mar. 1998.

Julius O. Smith III. Virtual acoustic musical instruments: Review and update. Journal of New Music Research, 33(3):283–304,
Autumn 2004.

Julius O. Smith III. Physical Audio Signal Processing: for Virtual Musical Instruments and Digital Audio Effects, December
2008 Edition. http://ccrma.stanford.edu/˜jos/pasp/, 2008. Accessed 15/12/2008.

Stephen E. Stewart and William J. Strong. Functional Model of a Simplified Clarinet. J. Acoust. Soc. Am., 68(1):109–120, July
1980.

Lutz Trautmann and Rudolf Rabenstein. Digital Sound Synthesis by Physical Modeling Using the Functional Transformation
Method. Kluwer Academic, New York, 2003.

Vesa Välimäki, Jyri Pakarinen, Cumhur Erkut, and Matti Karjalainen. Discrete-time modelling of musical instruments. Rep.
Prog. Phys., 69(1):1–78, 2006.

Kees van den Doel and Dinesh K. Pai. Modal Synthesis for Vibrating Objects. In Ken Greenebaum, editor, Audio Anecdotes.
AK Peters, Natick, MA, 2004.

S. A. van Duyne and J. O. Smith III. The 2-D Digital Waveguide Mesh. In Proc. IEEE Workshop on Applications of Sig.
Process. to Audio and Acoustics (WASPAA’93), pages 177–180, New Paltz (NY), Oct. 1993.

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>
Contents

3 Sound modeling: source-based approaches 3-1

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-1
3.2 Physical structures and models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-3
3.2.1 Simple vibrating systems and normal modes . . . . . . . . . . . . . . . . . . . 3-3
3.2.1.1 Oscillators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-3
3.2.1.2 Impedance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-4
3.2.1.3 Coupled oscillators and modal decomposition . . . . . . . . . . . . . 3-6
3.2.2 Continuous vibrating systems and waves . . . . . . . . . . . . . . . . . . . . . 3-6
3.2.2.1 The one-dimensional D’Alembert equation . . . . . . . . . . . . . . . 3-6
3.2.2.2 Traveling wave solution . . . . . . . . . . . . . . . . . . . . . . . . . 3-7
3.2.2.3 Waves and modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-8
3.3 Delays and oscillations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-9
3.3.1 The Karplus-Strong algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-9
3.3.1.1 The comb filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-10
3.3.1.2 Synthesis of plucked strings . . . . . . . . . . . . . . . . . . . . . . . 3-11
3.3.2 Fine tuning and fractional delays . . . . . . . . . . . . . . . . . . . . . . . . . . 3-14
3.3.2.1 FIR fractional delay filters . . . . . . . . . . . . . . . . . . . . . . . . 3-14
3.3.2.2 All-pass fractional delay filters . . . . . . . . . . . . . . . . . . . . . 3-16
3.3.2.3 Time-varying delays . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-18
3.4 Distributed models: the waveguide approach . . . . . . . . . . . . . . . . . . . . . . . . 3-18
3.4.1 Basic waveguide structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-18
3.4.1.1 Wave variables and wave impedance . . . . . . . . . . . . . . . . . . 3-18
3.4.1.2 Delay lines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-19
3.4.1.3 Boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-21
3.4.2 Modeling real world phenomena . . . . . . . . . . . . . . . . . . . . . . . . . . 3-21
3.4.2.1 Dissipation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-22
3.4.2.2 Loss filter design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-23
3.4.2.3 Dispersion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-24
3.4.2.4 Dispersion filter design . . . . . . . . . . . . . . . . . . . . . . . . . 3-24
3.4.3 Junctions and networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-25
3.4.3.1 The Kelly-Lochbaum junction . . . . . . . . . . . . . . . . . . . . . . 3-25
3.4.3.2 N-dimensional and loaded junctions . . . . . . . . . . . . . . . . . . 3-27
3.4.3.3 Non-cylindrical geometries . . . . . . . . . . . . . . . . . . . . . . . 3-28
3.5 Lumped models and the modal approach . . . . . . . . . . . . . . . . . . . . . . . . . . 3-29
3.5.1 Numerical methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-30
3.5.1.1 Impulse invariant method . . . . . . . . . . . . . . . . . . . . . . . . 3-30
3.5.1.2 Finite differences and mappings “s-to-z” . . . . . . . . . . . . . . . . 3-31

3-59
3-60 Algorithms for Sound and Music Computing [v.February 2, 2019]

3.5.1.3 Accuracy, stability, computability . . . . . . . . . . . . . . . . . . . . 3-32

3.5.1.4 Wave digital filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-34
3.5.2 Modal synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-35
3.5.2.1 Normal modes in finite dimensional systems . . . . . . . . . . . . . . 3-35
3.5.2.2 Normal modes in PDEs . . . . . . . . . . . . . . . . . . . . . . . . . 3-36
3.5.2.3 Discrete-time mechanical oscillators . . . . . . . . . . . . . . . . . . 3-38
3.5.2.4 A modal synthesizer . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-40
3.5.3 Modal analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-41
3.5.3.1 Simple 1-D shapes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-41
3.5.3.2 Simple 2-D shapes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-42
3.5.3.3 Experimental estimation . . . . . . . . . . . . . . . . . . . . . . . . . 3-44
3.6 Non-linear physical models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-44
3.6.1 Non-linear circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-45
3.6.1.1 Non-linear capacities . . . . . . . . . . . . . . . . . . . . . . . . . . 3-45
3.6.1.2 Vacuum tubes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-46
3.6.2 Mechanical interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-46
3.6.2.1 Impacts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-46
3.6.2.2 Stick-slip friction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-48
3.6.2.3 Tension modulations . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-49
3.6.3 Acoustic interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-50
3.6.3.1 Jets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-50
3.6.3.2 Quasi-static reeds . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-50
3.6.3.3 Dynamic reeds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-52
3.6.4 Computability issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-53
3.6.4.1 Non-linear systems and delay-free loops . . . . . . . . . . . . . . . . 3-53
3.6.4.2 Iterative methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-53
3.6.4.3 Sheared non-linearities . . . . . . . . . . . . . . . . . . . . . . . . . 3-53
3.7 Commented bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-55

This book is licensed under the CreativeCommons Attribution-NonCommercial-ShareAlike 3.0 license,

⃝2005-2019
c by the authors except for paragraphs labeled as adapted from <reference>