0DWWL .

DUMDODLQHQ 9HVD 9lOLPlNL DQG

7HUR 7RORQHQ
3OXFNHG6WULQJ
Helsinki University of Technology 0RGHOV)URPWKH
.DUSOXV6WURQJ
Laboratory of Acoustics and
Audio Signal Processing
P.O. Box 3000
FIN-02015 HUT, Finland
$OJRULWKPWR'LJLWDO
http://www.acoustics.hut.fi/
[email protected], [email protected],
:DYHJXLGHVDQG
[email protected] %H\RQG

The emergence of what is called physical modeling (Smith 1987, 1992, 1997); however, the relation has
and model-based sound synthesis is closely related to never been explicated in full detail. The first aim of
the development of computational simulations of this article is to show how the more “physical”
plucked string instruments. Historically, the first waveguide model of a plucked string can be reduced
physical approaches (Hiller and Ruiz 1971a, 1971b; to an extended form of the Karplus-Strong type that
McIntyre and Woodhouse 1979; McIntyre, we call the single delay-loop (SDL) model. For a linear
Schumacher, and Woodhouse 1983) were followed and time-invariant (LTI) case, this reduction is
by the Karplus-Strong (KS) algorithm (Karplus and relatively straightforward, and results in a
Strong 1983). The KS algorithm was discovered as a computationally more efficient digital filter
simple computational technique that seemingly had structure. (Note that the historical order of the KS
nothing to do with physics. Soon thereafter, Julius algorithm and digital waveguides is the reverse of
Smith and David Jaffe showed a deeper their logical order, since the generalization was not
understanding of its relation to the physics of the developed until after the KS algorithms was
plucked string (Smith 1983; Jaffe and Smith 1983). designed. This article’s title reflects the historical
Later, Julius Smith generalized the underlying evolution: the “beyond” refers to recent
ideas of the KS algorithm by introducing the theory generalizations and extensions of both concepts.)
of digital waveguides (Smith 1987). Digital The second aim of this article is to discuss further
waveguides are physically relevant abstractions yet extensions to the basic SDL models that make them
computationally efficient models, not only for capable of simulating plucking styles, beats in string
plucked strings, but for a variety of one-, two-, and vibration, sympathetic vibrations, and resonant
three-dimensional acoustic systems (Van Duyne and strings. Such techniques have already been proposed
Smith 1993; Savioja, Rinne, and Takala 1994; Van and studied (Jaffe and Smith 1983; Smith 1993;
Duyne, Pierce, and Smith 1994). Further Karjalainen, Välimäki, and Jánosy 1993). Here we
investigations embodied these ideas in more detailed discuss them in the context of our recent
synthesis principles and implementations, resulting implementations of plucked-string models.
in high-quality and realistic syntheses of plucked
string instruments (Sullivan 1990; Karjalainen and
Laine 1991; Smith 1993; Karjalainen, Välimäki, and )XQGDPHQWDOV RI 6WULQJ %HKDYLRU
Jánosy 1993; Välimäki, Huopaniemi, Karjalainen,
and Jánosy 1996). A recent overview of research in The behavior of a vibrating string with a plucked
this field is given by Smith (1996). excitation can be described in terms of two traveling
The equivalence of Karplus-Strong and digital waves traversing the string in opposite directions
waveguide formulations in sound synthesis was al– and reflecting back at the string terminations
ready known when the waveguide theory appeared (Elmore and Heald 1969; Fletcher and Rossing 1991).
When we assume that during autonomous vibration
This article was originally published in Computer Music Journal,
the string is an LTI system, we can model it as shown
22:3, pp. 17-32, Fall 1998. This reprint is available at URL:
http://www.acoustics.hut.fi/~vpv/publications/cmj98.htm
in Figure 1 (Smith 1987, 1992). The two delay lines

Karjalainen, Välimäki, and Tolonen 1

Figure 1. A bidirectional
digital waveguide model
for a terminated string.

Propagation direction

Delay line

1/2
R (z ) R b(z )
f x (n) y (n)

Delay line

Propagation direction

can be interpreted as a digitized d’Alembert’s trip along the string: at the loop filter H l(z). When the
solution to the one-dimensional lossless wave loop filter is a two-point average y(n) = [x(n) + x(n –
equation. The two waveforms travel through the 1)]/2, and when the initial conditions (i.e., the initial
delay lines and reflect at reflection filters R f (z) and contents of the delay line) that are used to pluck the
R b (z) which produce phase inversion and slight string are taken to be random numbers, the well-
frequency-dependent damping. The input signal x(n) known Karplus-Strong algorithm for plucked-string
is summed into both delay lines just as output signal sounds is obtained (Karplus and Strong 1983; Jaffe
y(n) is taken as a sum of the wave-variable values in and Smith 1983). Note that the original algorithm
the two delay lines at the observation point. This uses no explicit input signal.
digital waveguide-modeling approach yields
efficient implementations for real-time sound
synthesis. For the digital waveguide of Figure 1, a )URP %LGLUHFWLRQDO 'LJLWDO :DYHJXLGHV
further assumption is needed: All signals to be WR 6'/ 0RGHOV
modeled must be bandlimited to below one-half of
the sampling rate. Owing to the LTI assumption, Above, we discussed two computational models for
string losses and dispersions can be commuted stringed musical instruments that provide the basis
between any driving or observation points (Smith for efficient real-time synthesis. The case of extreme
1992, 1997). This allows the use of ideal delay lines simplicity and efficiency, the KS model of Figure 2, is
that are computationally very efficient. certainly an oversimplification for anything but
In the string model of Figure 1, the input and rudimentary synthesis. More detailed models are
output signals can be of any wave-variable type, needed both for high-quality sound synthesis and for
such as displacement, velocity, acceleration, or slope theoretical understanding in physical modeling. In
(Smith 1992; Morse 1976). An interesting case is to this section, we derive the relations between the
select acceleration as the wave variable, since then an bidirectional digital waveguide and the SDL
ideal pluck corresponds to a unit impulse (Smith formulations in detail.
1983; Karjalainen and Laine 1991). Let us consider the relation of the two basic
Further using the above simplification principles, formulations, the bidirectional digital waveguide
it is possible to commute the elements of a model and the single delay-loop model. We will
terminated, dispersive, and lossy string into the form analyze two cases: a string with (1) a bridge output
illustrated in Figure 2, provided that the output and (2) a pickup output. An excitation—such as a
signal is taken to be a single traveling-wave pluck—in a real physical string initiates wave
component. In this extreme case, the losses and the components that travel independently in opposite
dispersion are lumped at a single point in the round directions. The output of the string—e.g., the force at

2 Computer Music Journal

Figure 2. A simplified linear string model as a single
delay-loop (SDL) structure. The Karplus-Strong
algorithm is a special case in which the excitation is
given as the initial state of the delay line.

x (n) y (n)
Loop filter Delay line

H ( z)
l
z-L I

the bridge of an acoustic instrument, or the pickup In Figure 3, we redefine the dual delay-line
voltage in an electric guitar—reacts to both wave waveguide model for an ideally plucked string with
components. The effects of the excitation and pickup transversal bridge force as an output. This situation
positions are easily simulated in the waveguide is applicable to the simulation of the acoustic guitar,
model that is based on a dual delay line, as depicted for example. In our notation, H A,B(s) refers to the
in Figure 1. However, for sound-synthesis purposes, transfer function from point A to point B. Note that
the SDL realization, such as in Figure 2, is more we have divided the pluck excitation X(s) into two
efficient. Also, it is interesting from a theoretical parts, X 1(s) and X 2(s), such that X 1(s) = X 2(s) = X(s)/2.
point of view to formulate an SDL model that We can first simplify the model by deriving an
includes the effects of the excitation and pickup equivalent single excitation at point E1 that
positions. It has been shown that an ideal corresponds to the net effect of the two excitation
acceleration or velocity input into a string model components at points E1 and E2. When we assume
(corresponding to plucking or striking the string, that the bridge termination point (R1, R2) is to the
respectively) can be approximated by a unit impulse right of the input point (E1, E2) as in Figure 1, the
(Smith 1992). Thus, by assuming linearity and time equivalent single excitation at E1 can be expressed as
invariance, we can naturally think in terms of
impulse responses, and interpret the string model as XE1,eq(s) = X1(s) + HE2,L2(s)R f(s)HL1,E1(s)X2(s)
a linear filter. = ½[1 + HE2,E1(s)]X(s) = HE(s)X(s), (1)
where subscript “eq” stands for “equivalent,” and
Plucked String with Bridge Output X E2,E1(s) is the left-side transfer function from E2 to
E1 consisting of the partial transfer functions from E2
In the discussion to follow, we describe the transfer to L2 and L1 to E1, and the reflection function R f(s).
functions of the model components in the Laplace Thus, H E(s) is the equivalent excitation transfer
transform domain. The Laplace transform is an function.
efficient tool in linear continuous-time systems The output signal of interest is the transverse force
theory. In particular, time-domain integration and F( s) at the bridge. It can be elaborated as
derivative operations transform respectively into
division and multiplication by the Laplace variable s. F(s) = F+(s) + F–(s) = Z(s)[V+(s) – V–(s)]
We can replace the complex variable s with jω
= Z(s)[A1(s) – A2(s)]/s, (2)
(where j is the imaginary unit −1 , ω is the radian
frequency, which is equal to 2 π f, and f is the where the “+” and “–“ subscripts denote the two
frequency in Hz) to derive the corresponding opposite propagation directions. Equation 2 states
representation in the Fourier transform domain, i.e., that the bridge force is the bridge impedance Z(s)
the frequency domain. Later, we approximate the times the difference of the string velocity
continuous-time system by a discrete-time system in components V +(s) and V –(s) at the bridge. The
the Z-transform domain. For more information on acceleration difference A 1(s) – A 2(s) is integrated
Laplace, Fourier, and Z-transforms, see a standard (operator 1/s) to yield velocity difference V +(s) – V –
textbook on signal processing, such as that by (s). Hence
Oppenheim, Willsky, and Young (1983).

Karjalainen, Välimäki, and Tolonen 3

Figure 3. Dual delay-line
waveguide model for a
plucked string with out-
put at the bridge.

H L1,E1( s) H E1,R1 ( s)
L1 E1 R1 A 1( s)

Delay line Bridge

force
1/2 X 1 ( s)
R ( s) Z ( s)
f X ( s) R b( s) F ( s)
s
X 2 ( s)

Delay line

L2 E2 R2 A 2 ( s)

H E2,L2 ( s) H R2,E2 ( s)

F(s) = Z(s)[A1(s) – A2(s)]/s transfer function from excitation to bridge output as

= Z(s)[A1(s) – R b(s)A1(s)]/s F ( s)
H E,B ( s) =
= Z(s)[1 – R b(s)]A1(s)/s X ( s)
(7)
[ ]
1 H E1,R1( s) 1
= HB(s)A1(s), (3) = 1 + H E2,R1( s) Z( s) [1 − R b ( s)],
2 1 − H loop ( s) s
where H B(s) is the acceleration-to-force transfer
function at the bridge. Now or more compactly, based on the above notation,

A1(s) = HE1,R1(s)XE1,eq(s) + Hloop(s)A1(s), (4) H E,B( s ) = H E ( s )H E1,R1( s )S ( s )H B( s ) , (8)

where which represents the cascaded contribution of each

part in the physical string system.
H loop(s) = R b(s)H R2,E2(s)H E2,E1(s)H E1,R1(s), (5) At this point, we approximate the continuous-time
model of the guitar in the Laplace transform domain
i.e., H loop(s) is the transfer function when the signal is with a discrete-time model in the Z-transform
circulated once around the loop. Thus, the sum terms domain. This approximation is needed to make the
of Equation 4 correspond respectively to the model realizable in discrete-time form. Rewriting
equivalent excitation signal X E1,eq(s) transferred to Equation 8 in the Z-transform domain, we obtain
point R1, and to the signal A 1(s) transferred once
along the loop. This yields H E,B ( z) = H E ( z)H E1,R1 ( z)S( z) H B ( z) , (9a)
1 where
A 1 ( s ) = H E1,R1( s ) X E1,eq ( s )
1 − H loop ( s ) (6)
= H E1,R1( s )S( s )X E1,eq ( s ), H E ( z) =
1
2
[
1 + H E2,E1 ( z) , ] (9b)

where S(s) is the string transfer function that and

represents the recursion around the string loop.
Putting all this together, we can solve for the overall

4 Computer Music Journal

 between. The gain control is a simple multiplier. The
1
S( z ) = , (9c) timbre control may be a first- or second-order
1 − H loop ( z ) recursive filter that can be adjusted to attenuate or
boost high frequencies, for a softer or sharper attack,
and respectively (Jaffe and Smith 1983). Note that the
H B ( z) = Z( z)I ( z)[1 − R b ( z)] . (9d) timbre-control filter must change with fundamental
frequency to give a fixed percept of attack sharpness.
Filter I(z) is a discrete-time approximation of the The pluck-position control is an adjustable comb
time-domain integration operation. We interpret the filter, where the delay corresponds to the time it
result of Equation 9a by depicting a block diagram in takes for the excitation to travel the left-side route
Figure 4. It shows qualitatively the delays and the around from E2 to E1 in Figure 3. In practice, one can
discrete-time approximations of the filter instead use the right-side delay from point E1 to E2
components of Equation 9 inherent in each part of (which is normally shorter), since it will only slightly
the transfer function. Note that the top block in change the attack part of the model’s impulse
Figure 4 contains the minus sign of the reflection response. The comb filtering will create a series of
function included in H E2,E1(z). However, in the zeros in the transfer function at frequencies fm =
bottom block, the minus sign included in Rb(z) m/tD, where tD is the delay (in seconds) of the comb
cancels the minus sign of Equation 9d. filter, and m is an integer index (m = 0, 1, 2, ...).
For practical sound synthesis, the model of The bridge-output integrator can be approximated
Equation 9 and Figure 4 can be approximated by a by a first-order recursive low-pass filter whose cutoff
simplified model depicted in Figure 5, without frequency lies below the lowest fundamental
compromising the sound quality. The following frequency to be synthesized. One can also consider
simplifications have been made in Figure 5. The the pluck-position filter and the output integrator
transfer function H E2,E1(z) in Equation 9b is almost a together, since at zero frequency the zero in the
lossless delay, and we can drop the low-pass filter former filter will be canceled by a pole in the latter.
block in the excitation-position filter (or replace it Figure 6 illustrates the effect of both filters as found
with a constant slightly less than 1). Next, we notice in their combined magnitude transfer function. The
that in normal playing conditions the wave plucking position is expressed as a distance to the
propagation from the excitation point (E1) to the bridge relative to the string length, e.g., 50 percent
bridge position (R1) is also a nearly lossless delay, so refers to the middle point of the string (see Figure
that it can be left out without perceivable effects. The 6a). This effect is equivalent to the traditional
string loop S(z) in Equation 9c cannot be reduced, interpretation of plucked-string behavior (see
because the delay and the low-pass-type loop filter Fletcher and Rossing, Figure 2.5 [page 39] and Figure
are critical to the sound quality. Finally, the term [1 – 2.7 [page 41]).
R b(z)] in the bridge block H B(z) in Equation 9d can The string loop in Figure 5 is similar to the one in
jω
be approximated by the constant 2, since R b(e ) ≈ – Figure 2. The delay block must allow for fine tuning
1. The errors due to the reductions above can be of the time delay to achieve all desired pitch values
compensated for with the timbre-control filter in (Jaffe and Smith 1983). This is accomplished by
Figure 5. fractional delay filtering techniques, such as first-
The sound-synthesis model of Figure 5 includes order all-pass filtering or Lagrange interpolation
several controllable elements. The excitation table (Laakso, Välimäki, Karjalainen, and Laine 1996).
can contain any useful signals, such as a unit impulse Also, the loop filter has to be controllable, so that the
or a complex aggregate excitation (Smith 1993; decay of the harmonic components can be adjusted
Karjalainen, Välimäki, and Jánosy 1993). For properly according to the string length and other
example, this is an efficient way to implement the varying parameters of the string. One popular choice
resonances of the instrument body, since the body is a one-pole digital loop filter with two parameters:
response can be commuted back to the excitation, a DC gain and a cutoff frequency parameter (Jaffe
owing to the LTI assumption. The table can also be a and Smith 1983; Välimäki et al. 1996).
set of excitations to choose from or to interpolate

Karjalainen, Välimäki, and Tolonen 5

Figure 4. A diagram char-
acterizing the building
blocks of the plucked
string with force output
at the bridge.

Filtering due to excitation position

H (z)
E

1/2 Delay Low-pass

Input

Wave propagation, from excitation to bridge

H (z)
E1,R1

Delay Low-pass

String loop
S ( z)

Low-pass Delay

Filtering due to bridge coupling

H (z)
B

Low-pass

Output
Integrator

6 Computer Music Journal

Figure 5. A schematic
sound-synthesis model for
plucked string instruments.
The delay in the string loop
must be continuously
variable, which can be
achieved with a fractional
delay filter.

Gain Delay String loop

- +
Timbre
Wavetable control Integrator
+ -
Loop
Excitation processing Pluck position Delay filter

Figure 6. Two examples of

the comb-filter effect caused
by excitation-position and
bridge-integrator filtering,
when the relative distance
from the plucking point to
the bridge is 50 percent (a)
and 22.5 percent (b) of the
string length.

1 1

0.8 0.8
(a) 50 % (b) 22.5 %
Magnitude

Magnitude

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 2 4 6 8 0 2 4 6 8
Harmonic index Harmonic index
(a) (b)

Karjalainen, Välimäki, and Tolonen 7

Figure 7. Dual delay-line
waveguide model for a
plucked string with a
pickup output.

H L1,E1( s) H E1,O1 ( s) H O1,R1 ( s)

L1 E1 O1 R1

Delay line

X 1( s) A 1( s)
1/2
R ( s) P o( s) R b( s)
f X ( s) U ( s)
s
X 2( s)
A 2 ( s)

Delay line

L2 E2 O2 R2

H E2,L2 ( s) H R2,O2( s)

Plucked String with Pickup Output U( s) = Po ( s)V ( s)

P o ( s)
In the next analysis, we derive the transfer function
and the synthesis block diagram for a plucked string
=
s
[A 1(s) + A 2 (s)]
when the output is taken by a velocity-sensitive
[ ]
P ( s)
sensor, such as a magnetic pickup in an electric = o A 1( s) + H O1,R1( s)R b ( s)H O2,E2 ( s) A 1( s) (10)
s
guitar. Figure 7 shows a dual delay-line waveguide
model for this case. The pickup position is denoted
P ( s)
= o
s
[ ]
1 + H O1,O2 ( s) A 1( s)
by (O1, O2).
A comparison of Figures 3 and 7 reveals that the = H P ( s) A 1( s),
excitation transfer function is the same in both cases. where the pickup transfer function P o(s) models the
This implies that Equation 4 is valid also for a string effect of the pickup microphone on guitar sound. For
with a pickup output, provided that the pickup point a magnetic pickup, P o(s) is typically a second-order
(O1, O2) is to the right of the excitation point (E1, low-pass filter. Notice that it is combined with the
E2). integrator block 1/s at the output in Figure 7.
The output signal is now a variable, such as Corresponding to Equation 4, we can now write
voltage from a magnetic pickup, that is proportional
to the transversal string velocity (sum of the left- and A 1 ( s) = H E1,O1 ( s)X E1,eq ( s) + H loop ( s) A 1 ( s) , (11)
right-going components) at the pickup position.
Velocity V( s) is proportional to the time integral of which yields
acceleration, and the voltage output from the pickup 1
U( s) can be expressed as A 1 ( s) = H E1,O1 ( s) X E1,eq ( s)
1 − H loop ( s) (12)
= H E1,O1 ( s)S ( s)X E1,eq ( s).

8 Computer Music Journal

Figure 8. A pickup output
block for the string model
of Figure 7. The block LP
is a low-pass filter.

HP( z )
From the -1
string loop Delay LP

y (n) P o (z ) I ( z) u ( n)
1

Based on Equations 1, 10, and 12, we can solve for 4. A block diagram of the string-output stage for a
the overall excitation-to-pickup transfer function, pickup is shown in Figure 8. Note that again the
minus sign due to the reflection from the end of the
U( s)
H E,P ( s) = string is explicitly shown with multiplication by – 1
X ( s) (compare with Figure 4).
(13)
[ ] [ ]
1 H E1,O1( s) P o ( s) A sound-synthesis model for a plucked string
= 1 − H E2,E1 ( s) 1 + H O1,O2 ( s) ,
2 1 − H loop ( s) s with a pickup output will be the same as the
diagram shown in Figure 5, except that a pickup-
or more compactly, based on the above notations, position comb filter and a pickup filter, P o(s), must
be cascaded. The comb filter can be similar to the
H E,P ( s) = H E ( s)H E1,O1 ( s)S ( s)H P ( s) . (14) pluck-position filter of Figure 5, and the pickup filter
If the pickup point O is to the left of the excitation can be a second-order low-pass filter.
point (E1), we can derive a similar formulation The synthesis models derived above can be
where the indices 1 and 2 are interchanged, as are further simplified using the commuted aggregate
the string terminations, L and R. excitation techniques (Smith 1993; Karjalainen,
As with the case of the bridge output, we now Välimäki, and Jánosy 1993), whereby the impulse
approximate the continuous-time transfer function in responses of other parts, such as the body, are
the Laplace transform domain with a discrete-time preconvolved with the primary excitation and stored
transfer function in the Z-domain: in a wavetable. If all the components apart from the
string loop are aggregated, we approach a variant of
H E,P ( z) = H E ( z) H E1,O1 ( z)S( z )H P ( z ) . (15) the basic KS algorithm that has an explicit input
signal. All variations are possible, from a fully
The differences between the bridge-output model aggregated excitation containing even the string loop
(Equation 7) and the pickup-output model (Equation (that is, pure wavetable synthesis), to the SDL model
13) are twofold. First, the wave propagation from the of Figure 5, or to the full bidirectional waveguide
excitation point to the output, H E1,R1(s) versus model of Figure 3. This shows the compatibility of
H E1,O1(s), covers different distances along the string. physical modeling with more traditional synthesis
Due to very low losses of wave propagation during techniques, such as source-filter modeling or
normal playing conditions, this difference is sampling. The more simplified versions are
negligible. Second, the transfer functions related to computationally more efficient, naturally, but the
the output couplings are very different. While the price is paid with decreased flexibility of parametric
difference of accelerations at the bridge R has a control.
relatively flat response (Equation 3 and Figure 3), the
summation of the two acceleration waves at the
pickup point O creates a comb-filter effect (Equation 0RUH ([WHQVLRQV WR WKH 6'/ 0RGHOV
10 and Figure 7) similar to the excitation point
filtering. Thus, a string-instrument model with a The SDL models derived above are flexible building
string-velocity pickup has two cascaded comb filters blocks when one is developing model-based sound
to color the response, instead of just the one in Figure synthesis of plucked string instruments. The SDL

Karjalainen, Välimäki, and Tolonen 9

Figure 9. Single delay-loop structure with consolidated
pluck and body wavetables, pluck-position comb filter,
dual-polarity string models, and sympathetic couplings
between strings (Välimäki et al. 1996).

Sympathetic
Horizontal polarization couplings
to other strings
Pluck Pluck
shaping position Sh(z)
filter comb filter mo out
Pluck and Body
E (z) P (z)
Wavetable n
1- m o
Sympathetic S v(z)
couplings from
other strings Vertical polarization

models show improved efficiency compared to the “synthesizer-like” character. An example of the
bidirectional digital waveguides, because the basic effect of mistuning the two polarization models is
DSP building blocks are maximally consolidated and shown in Figure 10. In Figure10a, the model
simplified. For example, a single fractional delay parameters are equal, and exponential decay is
filter for fine-tuning the pitch is sufficient. However, resulted. In Figure10b, the fundamental frequencies
if the string behavior contains essential nonlinearities of the models are equal, but the loop-filter
or time-varying characteristics, bidirectional parameters are different, and a two-stage decay is
waveguide formulations are needed (see, for produced. In Figure10c, the loop-filter parameters
example, Karjalainen, Backman, and Pölkki 1993). are equal, but the frequencies are mistuned to obtain
It is possible to add further details to the SDL a beating effect.
models to improve the naturalness and other sound- Several principles to simulate the sympathetic
quality features. Figure 9 illustrates an extended coupling between strings have been proposed (Jaffe
model in which additional properties are and Smith 1983; Smith 1993; Välimäki et al. 1996). A
implemented (Välimäki et al. 1996). The string physically correct method is a bridge-coupling filter
model’s excitation is realized with wavetables that presented by Smith (1993). A simple feedback
store consolidated pluck excitations and body coupling is added in Figure 9 to simulate the
responses, for easy and extremely efficient (but non- sympathetic coupling between strings. This
physical) modeling of the body (Smith 1993; approach is potentially unstable, because there is a
Karjalainen, Välimäki, and Jánosy 1993). Alternative feedback from the output of all strings to their
wavetables can be applied for different pluck styles inputs. When the feedback signals are attenuated
and qualities. A pluck-shaping filter E(z) can be used using small gain coefficients between all outputs and
to fine-tune the timbre of a single pluck-table inputs, it is possible to reach a stable simulation.
excitation. The comb-notch filter effect caused by Nevertheless, it would be safer to use a sympathetic
pluck position and pickup position are easy to add. coupling model that is inherently stable. Jaffe and
The pluck-position filter P(z) is shown in Figure 9, Smith (1983) proposed using a separate bank of
but the pickup-position filter is not included. sympathetic strings that get their input signal from
The beat effects caused by dual polarization of the output of the main strings that are plucked. This
string vibration (horizontal and vertical with respect approach is always stable. However, it would be
to the top plate) can be realized by mixing the desirable to use the existing strings of the instrument
outputs of two string models (Jaffe and Smith 1983). model for generating sympathetic vibrations, and
When the two models are slightly mistuned, a not implement separate models. In the following
natural sounding beat effect results that reduces the section, we introduce such a configuration.

10 Computer Music Journal

Figure 10. An example of
the effect of mistuning the
polarization models: equal
parameter values (a),
mistuned decay rates (b),
and mistuned fundamental
frequencies (c).

0.5
Amplitude

−0.5

−1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time (sec)
(a)
1

0.5
Amplitude

−0.5

−1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time (sec)
(b)
1

0.5
Amplitude

−0.5

−1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time (sec)
(c)

Karjalainen, Välimäki, and Tolonen 11

Figure 11. A modified
plucked-string synthesis
model with a new method
to include sympathetic
coupling between strings.

C Sympathetic couplings
Horizontal polarization to other strings
Sh(z)
mp
mo out
Pluck and Body
E (z) P (z)
Wavetable n gc
1- m p
1- m o
Sympathetic
couplings from S v(z)
other strings
Vertical polarization

Coupling Phenomenon gains of the output signal to be sent from the kth
horizontal string to its parallel vertical string, and
Figure 11 illustrates a modified version of the
coefficients cmk are the gains of the kth horizontal
plucked-string synthesizer for a single dual-
string output to be sent to the mth vertical string.
polarization string. This time, the excitation for
There is a physical motivation to use real numbers
sympathetic vibrations is taken from one of the
less than 1 for all the elements of matrix C. However,
parallel strings that model the two polarizations (in
the structure’s stability does not depend on these
Figure 11, the horizontal one is used). To avoid
values, since there is no feedback.
feedback, the input from other strings is added to the
The model shown in Figure 11 also divides the
input of only those parallel strings that do not have
excitation signal that is sent to the two polarizations,
sympathetic coupling output. This implies that input
by using a mixing coefficient mp that is chosen to
must be fed into the vertical string model in Figure
have a value between 0 and 1. A non-zero value of
11, because the output is taken from the horizontal
parameter gck enables coupling of the two
model. This kind of signal coupling is
polarizations. If at the same time mp = 1, the model
unconditionally stable, and produces realistic
for the vertical polarization becomes a resonance
sympathetic coupling phenomena. Furthermore,
string that receives input only from the upper string
separate sympathetic string models need not be
model in Figure 11.
implemented, since sympathetic vibrations are now
An example of sympathetic coupling is pictured in
produced in a natural manner by all the strings
Figure 12. The primary vibration (the waveform
included in the synthesis model.
displayed in the upper part of Figure 12) excites
In the general form of this algorithm, there is a
another string whose output signal is shown in the
matrix C of coupling coefficients that determine the
lower part of Figure 12. In this example, the
proportion of the output signal to be sent to a
fundamental frequency of the secondary string is an
particular parallel string. This matrix can be written
octave higher than that of the primary string. Notice
as
the slow attack of the vibration in the coupled string
 g c1 c12 c13  c1 N  (the lower part of Figure 12).
c g c2 c 23 
 21 
C =  c 31 c 32 gc3   (16)
 
6WLOO 0RUH ([WHQVLRQV DQG )XWXUH 'LUHFWLRQV
 
 
c N 1
  g cN  Topics for future work in physical modeling of
plucked strings includes pluck simulation, different
where N is the number of dual-polarization strings, kinds of nonlinearities, the interaction between the
the coefficients gck (for k = 1, 2, 3, ..., N) denote the string and the body, modeling of the body response

12 Computer Music Journal

Figure 12. An example of
the simulation of sympa–
thetic coupling using the
model of Figure 11. The
primary vibration (top)
excites vibration in another
string (bottom).

0.5
Amplitude

−0.5

−1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (s)

0.5
Amplitude

−0.5

−1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (s)

using a digital filter, and calibration of the model by Karjalainen and Smith (1996). In the
parameters. Some of these problems have been computationally efficient synthesis models, it has
tackled in recent literature. Rank and Kubin (1997) been advantageous to avoid the use of a body model
proposed a nonlinear model for cases where the and instead use the principle of commuted
amplitude of string vibration is limited by contact waveguide synthesis. However, if one wishes to
with frets, such as in slap-bass playing techniques. A simulate the two-way interaction between the strings
passive nonlinear filter structure was devised by and the body, an explicit model for the body can be
Pierce and Van Duyne (1997). They presented an developed, although this is not required in the LTI
example where the nonlinear generation of missing case.
harmonics in string vibrations was successfully Calibration of the parameter values of a plucked-
simulated by their simple digital model. Digital filter string model was tackled by Välimäki et al. (1996).
approximations of the body response were discussed The proposed technique was based on short-time

Karjalainen, Välimäki, and Tolonen 13

Fourier analysis of recorded plucked string tones. sound quality was presented, including a new way
The excitation signal was obtained by inverse- to simulate sympathetic vibrations.
filtering a recorded sound using the inverse transfer
function of the string model. This enabled high-
quality resynthesis of tones that were well behaved $FNQRZOHGJPHQWV
in that their harmonics decayed nearly
exponentially. Problematic cases are tones where This work was originated by the first author when he
beats occur in harmonics or the decay rate of was a visiting scholar at CCRMA, Stanford
harmonics changes over time in some other way. A University, during the academic year 1994–1995. The
recent improvement to this problem is a method authors would like to thank Professor Julius O.
where the excitation signal is obtained by subtracting Smith and his students for many fruitful discussions.
a sinusoidal model of the harmonics from an
analyzed tone (Tolonen and Välimäki 1997; Välimäki
and Tolonen 1998; Tolonen 1998). This method References
accounts for the time-varying decay rate of
Elmore, W. C., and M. A. Heald. 1969. Physics of Waves.
harmonics, and yields a clean excitation signal. A
New York: Dover.
remaining problem in parameter calibration is to
Fletcher, N. H., and T. D. Rossing. 1991. The Physics of
reliably estimate the pluck position from a recording
Musical Instruments. New York: Springer-Verlag.
of a plucked-string tone (Välimäki et al. 1996). It
Hiller, L., and P. Ruiz. 1971a. “Synthesizing Musical
would be useful to extract this information and
Sounds by Solving the Wave Equation for Vibrating
cancel its effect in the excitation signal. Thereafter,
Objects: Part 1.” Journal of the Audio Engineering Society
the excitation position would be a free parameter in
19(6):462–470.
the synthesis stage.
Hiller, L., and P. Ruiz. 1971b. “Synthesizing Musical
The realism of plucked-string synthesis could be
Sounds by Solving the Wave Equation for Vibrating
further improved by including different kinds of side
Objects: Part 2.” Journal of the Audio Engineering Society
effects, such as those generated when the player
19(7):542–550.
slides his or her finger along a string (friction noise)
Jaffe, D. A., and J. O. Smith. 1983. “Extensions of the
or touches the body of a guitar (Jánosy, Karjalainen,
Karplus–Strong Plucked-String Algorithm.” Computer
and Välimäki 1994). Such sound effects could be
Music Journal 7(2):76–87. Reprinted in C. Roads, ed.
incorporated by using samples that are triggered at
1989. The Music Machine. Cambridge, Massachusetts:
the right time according to certain rules or, to obtain
MIT Press.
more expression and variation, by developing
Jánosy, Z., M. Karjalainen, and V. Välimäki. 1994.
physical models for these effects. Furthermore, a
“Intelligent Synthesis Control with Applications to a
virtual plucked string instrument should allow the
Physical Model of the Acoustic Guitar.” Proceedings of
player to use the complete range of playing styles,
the 1994 International Computer Music Conference. San
including the left-hand techniques such as pull-offs,
Francisco, California: International Computer Music
hammer-ons, trills, slurs, and the use of a slide, as
Association, pp. 402–406.
well as right-hand techniques such as fret-tapping.
Karjalainen, M., J. Backman, and J. Pölkki. 1993. “Analysis,
Jaffe and Smith (1983) discussed the simulation of
Modeling and Real-Time Synthesis of the Kantele, a
some of these styles. Nevertheless, realistic and
Traditional Finnish String Instrument.” Proceedings of
computationally efficient physical modeling of these
the 1993 IEEE International Conference on Acoustics,
playing techniques remains a future challenge.
Speech, and Signal Processing, vol. 1. Piscataway, New
Jersey: IEEE, pp. 229–232.

&RQFOXVLRQV Karjalainen, M., and U. K. Laine. 1991. “A Model for Real-

Time Sound Synthesis of Guitar on a Floating-Point
Signal Processor.” Proceedings of the 1991 IEEE
In this article, we focused on algorithms for the
International Conference on Acoustics, Speech, and Signal
synthesis of plucked-string tones. The relation and
Processing, vol. 5. Piscataway, New Jersey: IEEE, pp.
equivalence of the bidirectional digital waveguide
3653–3656.
and single delay-loop (SDL) models of plucked
Karjalainen, M., and J. O. Smith. 1996. “Body Modeling
string instruments were discussed. The derivation of
Techniques for String Instrument Synthesis.”
SDL models based on the digital waveguide
Proceedings of the 1996 International Computer Music
approach was given for the cases of a bridge output
Conference. San Francisco, California: International
and a pickup output. A further extension of the SDL
Computer Music Association, pp. 232–239.
model with several additional features for improved

14 Computer Music Journal

Karjalainen, M., V. Välimäki, and Z. Jánosy. 1993. Smith, J. O. 1992. “Physical Modeling Using Digital
“Towards High-Quality Sound Synthesis of the Guitar Waveguides.” Computer Music Journal 16(4):74–91.
and String Instruments.” Proceedings of the 1993 Smith, J. O. 1993. “Efficient Synthesis of Stringed Musical
International Computer Music Conference. San Francisco, Instruments.” Proceedings of the 1993 International
California: International Computer Music Association, Computer Music Conference. San Francisco, California:
pp. 56–63. International Computer Music Association, pp. 64–71.
Karplus, K., and A. Strong. 1983. “Digital Synthesis of Smith, J. O. 1996. “Physical Modeling Synthesis Update.”
Plucked-String and Drum Timbres”. Computer Music Computer Music Journal 20(2):44–56.
Journal 7(2):43–55. Reprinted in C. Roads, ed. 1989. The Smith, J. O. 1997. “Acoustic Modeling Using Digital
Music Machine. Cambridge, Massachusetts: MIT Press. Waveguides.” In C. Roads, S. T. Pope, A. Piccialli, and
Laakso, T. I., V. Välimäki, M. Karjalainen, and U. K. Laine. G. De Poli, eds. Musical Signal Processing. Lisse, the
1996. “Splitting the Unit Delay—Tools for Fractional Netherlands: Swets & Zeitlinger, chapter 7, pp. 221–
Delay Filter Design.” IEEE Signal Processing Magazine 264.
13(1):30–60. Sullivan, C. S. 1990. “Extending the Karplus–Strong
McIntyre, M. E., R. T. Schumacher, and J. Woodhouse. Algorithm to Synthesize Electric Guitar Timbres with
1983. “On the Oscillations of Musical Instruments.” Distortion and Feedback.” Computer Music Journal
Journal of the Acoustical Society of America 74(5):1325– 14(3):26–37.
1345. Tolonen, T. 1998. “Model-Based Analysis and Resynthesis
McIntyre, M. E., and J. Woodhouse. 1979. “On the of Acoustic Guitar Tones.” Master’s thesis, Report 46,
Fundamentals of Bowed String Dynamics.” Acustica Laboratory of Acoustics and Audio Signal Processing,
43(10):83–108. Department of Electrical and Communications
Morse, P. M. 1976 (reprint). Vibration and Sound. American Engineering, Helsinki University of Technology. Espoo,
Institute of Physics for the Acoustical Society of Finland. A PostScript version is available at URL:
America (first edition 1936; second edition 1948). http://www.acoustics.hut.fi/~ttolonen/thesis.html
Oppenheim, A. V., A. S. Willsky, and I. T. Young. 1983. Tolonen, T., and V. Välimäki. 1997. “Automated
Signals and Systems. New York: Prentice-Hall. Parameter Extraction for Plucked String Synthesis.”
Pierce, J. R., and S. A. Van Duyne. 1997. “A Passive Proceedings of the 1997 International Symposium on
Nonlinear Digital Filter Design which Facilitates Musical Acoustics, book 1. St. Albans, UK: Institute of
Physics-Based Sound Synthesis of Highly Nonlinear Acoustics, pp. 245–250.
Musical Instruments.” Journal of the Acoustical Society of Välimäki, V., J. Huopaniemi, M. Karjalainen, and Z.
America 101(2):1120–1126. Jánosy. 1996. “Physical Modeling of Plucked String
Rank, E., and G. Kubin. 1997. “A Waveguide Model for Instruments with Application to Real-Time Sound
Slapbass Synthesis.” Proceedings of the 1997 IEEE Synthesis.” Journal of the Audio Engineering Society
International Conference on Acoustics, Speech, and Signal 44(5):331–353.
Processing, vol. 1. Los Alamitos, California: IEEE Välimäki, V., and T. Tolonen. 1998. “Development and
Computer Society Press, pp. 443–446. Calibration of a Guitar Synthesizer.” Journal of the Audio
Savioja, L., T. J. Rinne, and T. Takala. 1994. “Simulation of Engineering Society 46(9):766–778.
Room Acoustics with a 3-D Finite Difference Mesh.” Van Duyne, S. A., J. R. Pierce, and J. O. Smith. 1994.
Proceedings of the 1994 International Computer Music “Traveling Wave Implementation of a Lossless Mode-
Conference. San Francisco, California: International Coupling Filter and the Wave Digital Hammer.”
Computer Music Association, pp. 463–466. Proceedings of the 1994 International Computer Music
Smith, J. O. 1983. “Techniques for Digital Filter Design and Conference. San Francisco, California: International
System Identification with Application to the Violin.” Computer Music Association, pp. 411–418.
PhD thesis, Stanford University. Stanford, California. Van Duyne, S. A., and J. O. Smith. 1993. “Physical
Smith, J. O. 1987. “Music Applications of Digital Modeling with the 2-D Digital Waveguide Mesh.”
Waveguides.” Technical Report STAN-M-39, CCRMA, Proceedings of the 1993 International Computer Music
Department of Music, Stanford University. Stanford, Conference. San Francisco, California: International
California. Computer Music Association, pp. 40–47.

Karjalainen, Välimäki, and Tolonen 15