Professional Documents
Culture Documents
Abel,* Andrei
Vladimirescu, and Julius O. Smith*
Numerical Methods for
*Center for Computer Research in Music and
Acoustics (CCRMA)
Simulation of Guitar
Stanford University
Stanford, California 94305-4088 USA
Distortion Circuits
ccrma.stanford.edu
{dtyeh, abel, jos}@ccrma.stanford.edu
Electric guitarists prefer analog distortion effects are accurately described in the audio frequency
over many digital implementations. This article band by nonlinear ODEs. A circuit simulator such
suggests reasons for this and proposes that detailed as SPICE (Simulation Program with Integrated
study of the electrical physics of guitar distortion Circuit Emphasis; see, e.g., Vladimirescu 1994)
circuits provides insight to design more accurate solves these systems of nonlinear ODEs to accu-
emulations. This work introduces real-time emula- rately predict their behaviors. However, SPICE
tion applied to guitar audio amplifiers in the form of simulation is computationally involved, so real-
a tutorial about relevant numerical methods and a time effects processing requires a simplified ap-
case study. The results here make a compelling case proach. Often, the circuits can be approximately
for simulating musical electronics using numerical partitioned into stages, neglecting loading effects
methods in real time. where possible (Yeh, Abel, and Smith 2007a), or
Analog guitar distortion effect devices known as even incorporating the loading effects as an equiva-
solid-state distortion boxes commonly include a lent circuit. Linear stages can be efficiently imple-
diode clipper circuit with an embedded low-pass mented by infinite impulse response (IIR) digital
filter. These distortion-effect devices can be mod- filters, although the remaining nonlinear ODEs
eled and accurately simulated as Ordinary Differen- may need to be solved by a numerical method or
tial Equations (ODEs). A survey and a comparison other approximation, usually employing a static
of the basic numerical integration methods are nonlinearity.
presented as they apply to simulating circuits for The diode-clipper circuit with an embedded
audio processing, with the widely used diode clipper low-pass filter forms the basis of both diode clipping
presented as an example. distortion and overdrive or tube screamer
A dedicated simulator for the diode clipper has effects pedals (Yeh, Abel, and Smith 2007a), and it is
been developed to compare several numerical found in many other products that implement
integration methods and their real-time feasibility. guitar distortion using solid-state circuitry. This
We found that implicit or semi-implicit solvers are popular circuit block is taken as a case example to
preferred, although the prefilter / static nonlinearity evaluate the performance and feasibility of using
approximation comes surprisingly close to the numerical integration methods to solve nonlinear
actual solution. ODEs in real time for an audio system and how it
compares to a static nonlinearity approximation.
(The terms diode clipper and diode limiter refer
Background to the same circuit and are used interchangeably in
this article.)
Analog guitar effects, whether based upon vacuum The Boss DS-1 circuit (Roland Corporation 1980)
tubes or solid-state devices, consist of circuits that is a distortion pedal, and its schematic can be
approximately divided into blocks as shown in
Computer Music Journal, 32:2, pp. 2342, Summer 2008 Figure 1. In this article, we focus on the saturating
2008 Massachusetts Institute of Technology. nonlinearity block, which is the diode clipper. We
Yeh et al. 23
Figure 1. Partitioning Figure 2. RC low-pass filter Figure 3. Linearized
scheme and block diagram with diode limiter. diode-clipper circuit.
for the Boss DS-1 circuit.
Yeh et al. 25
for highly nonlinear systems, direct simulation by circuit nodes in digital circuits to speed up the
numerical methods is more computationally effi- simulation (Newton and Sangiovanni-Vincentelli
cient (Bilbao 2006). Even with many terms, Volterra 1984). Nonlinear analog circuits for audio pro-
series, which use polynomial models, do not con- cessing, however, are typically highly coupled,
verge sufficiently for efficient implementation of a possibly with global feedback, and still require the
clipping nonlinearity with large signal excursions. use of traditional, or direct, ODE methods (White
and Sangiovanni-Vincentelli 1987).
ODE solvers use numerical integration methods
Simulation Methods to approximate a solution to the differential equa-
tion. SPICE typically offers the choice of Backward
The use of circuit simulation for real-time distor- Euler, trapezoidal rule, and BDF (often referred to as
tion processing was possibly first mentioned by Gear), which are implicit but stable. The popular
Sapp, Becker, and Brour (1999). Serafini and Zam- explicit Runge-Kutta method of order four has great
boni (2002) simulated a high-pass variant of the accuracy and is easy to use, but it is computation-
diode clipper using trapezoidal integration and ally expensive (Press et al. 1992). The extrapolation
compared it to a prefilter / static nonlinearity imple- technique (Stoer and Bulirsch 2002) is an efficient
mentation. Santagata, Sarti, and Tubaro (2007) also way to dramatically boost the accuracy of the
applied a numerical technique on a memoryless solution, but requires more work than the simple
circuit, which amounts to an implementation of methods.
Newtons method with only one iteration per time
step. Huovilainen (2004, 2005) and Vlimki and
Huovilainen (2006) effectively simulated the Moog Wave Digital Filter
filter and other effects circuits (e.g., phaser, flanger,
and chorus effects) using the explicit Forward Euler An alternative formulation to the ODE problem is to
method to generate a computable filter algorithm express the signals and states in terms of wave vari-
from the ODE of the circuit. This article extends ables and apply component-wise, or local, discretiza-
the investigation of implicit ODE methods previ- tion (Fettweis 1986) at a uniform sample rate. This
ously presented in Yeh, Abel, and Smith (2007b). formulation is known as the Wave Digital Principle,
and the resulting ODE solvers are Wave Digital
Filters (WDF). WDFs typically apply trapezoidal-rule
Prior Work in Ordinary Differential Equation integration in the form of the Bilinear Transform
Solvers 2 z 1
s= (3)
Numerical solution of ODEs is a mature topic in T z +1
applied mathematics, and many sophisticated because it preserves stability across continuous- and
algorithms exist for efficiently attaining accurate discrete-time domains, but other passive ODE
solutions (Gear 1971; Press et al. 1992; Shampine methods with greater orders of accuracy have been
1994; Stoer and Bulirsch 2002). The MATLAB developed for WDFs as well (Frnken and Ochs
scientific computing environment features a rich 2001, 2002). A procedure to automatically generate
suite of ODE solvers (Shampine and Reichelt 1997) the minimal WDF from an arbitrary circuit topology
that can be used for experimentation and gaining has been developed (Meerktter and Frnken 1996;
experience with the solution of ODEs. The circuit Frnken, Ochs, and Ochs 2005). The nonlinear WDF
simulator SPICE (McCalla 1987; Vladimirescu has been investigated by Meerktter and Scholz
1994) is essentially a nonlinear ODE solver. (1989) and Sarti and De Poli (1999) among others,
More advanced simulation techniques were and it was used by Karjalainen and Pakarinen (2006)
subsequently developed based upon relaxation to simulate the ODE of a simplified vacuum tube
methods, which take advantage of loosely coupled preamplifier circuit for guitar distortion.
Yeh et al. 27
the only practical order-preserving discretization described in depth in numerical methods textbooks
method that does not introduce artificial damping, (e.g., Press et al. 1992). At time n, the implicit
that is, turning unstable continuous-time poles into method must be solved for current state X = x[n] and
stable discrete-time ones (Smith 2007). The WDF can be rewritten in the general form
implementation of the trapezoidal rule was also
0 = F(X) (11)
tried, but it was found to produce exactly the same
results while requiring more operations; therefore, which represents a nonlinear root-finding problem.
it is not presented here. Newtons method requires the Jacobian, denoted
Another numerical integration method, the second- JF(X), of Equation 11 with respect to X and evaluated
order Backward Difference Formula (BDF2), is com- at X. In the case of the diode clipper (Equation 2),
monly used in circuit simulation, and it deserves which has only one state variable, JF(X) is simply the
mention here. It is a multi-step implicit method derivative of Equation 11 with respect to X. For
that only requires a single function evaluation of most implicit ODE methods, this can be easily
the time derivative of Equation 4 per iteration: computed from the Jacobian of f(t,x,u) (Equation 4),
4 1 2T ! denoted as Jf(X).
x[n] = x[n 1] x[n 2] + x[n] (9) Newtons method is in general given by
3 3 3
Finally, a popular higher-accuracy-order one-step !X = J F1(X)F(X)
(12)
method is the explicit fourth-order Runge-Kutta X := X + !X
Formula (RK4):
and iterating until X is below some acceptable
k1 = Tg(n 1, x[n 1]), value. It converges rapidly if the iteration is started
k2 = Tg(n 1 2 , x[n 1] + k1 2), with an initial guess for X that is close to the final
solution. Typically, this guess is the previous state
k3 = Tg(n 1 2 , x[n 1] + k2 2), (10) x[n 1], which works well when the system is
k4 = Tg(n, x[n 1] + k3), oversampled and successive samples are close in
value to each other.
k1 k2 k3 k4
x[n] = x[n 1] + + + +
6 3 3 6
Semi-Implicit Methods
where g(m,x) = f(t[m],x,u[m]) and f(t,x,u) is as defined
in Equation 4. Note that RK4 requires function eval- Given the observation that guitar-distortion sys-
uations every half sample, and therefore it requires tems are highly oversampled to suppress aliasing,
input at twice the sampling rate of the output, also the implicit methods can be modified to evaluate
noted by Huovilainen (2004). On the contrary, the only one step of the Newton method iteration. This
expanded bandwidth of the distorted output signal is known as the semi-implicit method (Press et al.
requires a sampling rate higher at the output than at 1992), which has constant cost. Effectively, this
the input to reduce aliasing. Implicit Runge-Kutta converts implicit integration methods into explicit
constructions also have been developed extensively form by removing constraints on the final X,
(Gear 1971; Butcher 1987; Frnken and Ochs 2001). which also relinquishes control over the error. The
expressions to evaluate X for the two most well-
known implicit methods are given below; semi-
Newtons Method for Solving Nonlinear Equations implicit BDF2 can be similarly derived. In the
following analysis, it is observed that X = x[n 1],
The nonlinear equation produced by the application and f(n,X) is understood to be shorthand for
of implicit numerical integration formulas can be f(t[n],X,u[n]). Extension to systems with multiple
solved by Newton-Raphson iteration, which is states is straightforward.
The Semi-Implicit Backward Euler (BE s-i) is The traditional measure of accuracy is Local Trun-
given by cation Error (LTE), which is the lowest-order differ-
x[n 1] X + Tf(n, X) Tf(n, X) ence between the full Taylor Series expansion of the
!X = = (13) solution and the result of the method. For example,
1 TJ f (X) 1 TJ f (X)
x(t) x(t) x(t T) is first-order-accurate because the
and the Semi-Implicit Trapezoidal Rule (TR s-i) is error is proportional to T as T0. The trapezoidal
given by rule, on the other hand, exhibits an error propor-
tional to T 2 as T0, so it is second-order-accurate.
x[n 1] X + 0.5T ( f(n, X) + f(n 1, x[n 1]))
!X = Manifestations of this error are aliasing and fre-
1 0.5TJ f (X) quency warping. Oversampling reduces error by the
0.5T ( f(n, X) + f(n 1, x[n 1])) accuracy order of the method. For example, it is
= (14)
1 0.5TJ f (X) known that the trapezoidal rule has the smallest
truncation error of any method of order two (Mc-
Calla 1987). Specifically, its truncation error de-
Approximation of an ODE by Static Nonlinearity creases as one-twelfth the square of the sampling
and Digital Filters interval (second-order convergence).
Yeh et al. 29
Figure 5. (a) Regions of (b) regions of stability for
stability for explicit implicit methods Back-
methods Forward Euler ward Euler (BE), BDF2,
(FE) and Runge Kutta 4 and trapezoidal rule (TR)
(RK4) are inside boundary; are outside boundary.
(a) (b)
poles to discrete-time poles. If all resulting discrete- frequency poles, possibly causing some unstable
system poles lie within the unit circle, the discreti- poles to be mapped to stable poles in the digital
zation yields a stable numerical solution. domain. The trapezoidal rule is the bilinear trans-
form, mapping the complex frequency axis onto the
unit circle and introducing no additional damping.
Explicit Methods
An A-stable method will always converge to a stable
The plot of the region of stability on the complex result as long as the nonlinear solver converges.
T-s plane (the s plane normalized times T) forms a
bounded region where the method is stable. Explicit
Stiff Stability
methods, such as Forward Euler and explicit RK4,
result in polynomial stability conditions (Stoer and For the ODEs found in analog circuits, it has been
Burlisch 2002), which trace out an external boundary found in practice that implicit methods drastically
of the stability region in the s-plane (see Figure 5a). reduce the sampling-rate requirement relative to
Consequently, this places a limit on the largest explicit methods and are ultimately more efficient
magnitude negative eigenvalue the system may (McCalla 1987). Circuit simulation problems often
have to assure bounded behavior. have eigenvalues that are highly separated in value, a
property known as stiffness in the ODE literature,
requiring a long time scale to compute the solution,
Implicit Methods
and a small time step for stability when using ex-
For implicit methods, the stability region extends to plicit methods. Stiffly stable solvers place no re-
infinity in the negative half-plane (see Figure 5b), quirement on the minimum sampling rate needed to
thereby placing no limit on the maximum magnitude ensure a bounded solution. Instead, considerations
of an eigenvalue of a system (if it is not complex) for accuracy such as aliasing govern the choice of
and allowing a low sampling rate. For trapezoidal, step size. None of the explicit methods can be stiffly
Backward Euler, and BDF2, the regions encompass stable (Stoer and Bulirsch 2002), because they require
the entire left half-plane, so all stable continuous- a minimum sampling rate to operate properly.
time systems will map to stable discrete-time The left-half plane eigenvalue, or pole, of the
systems (A-stability). Backward Euler and BDF2 diode clipper can be found from the small-signal lin-
will introduce artificial damping to higher- earization of the circuit shown in Figure 3. When Vo
Comparative Results
Considerations for Application to Audio Distortion
Circuits The basic methods presented in the previous sec-
tion were applied to the ODE of the diode clipper
and compared for various input signals. In all cases
Error
the methods were run with 8 oversampling. The
The typical implementation of an ODE solver tar- iterations terminate when the correction given by
gets applications with different error requirements Newtons method for the previous iterate is less
than real-time audio. Error for audio is best defined than 5 mV. This is acceptable, because, in an actual
spectrally and perceptually in the short-time circuit, noise from the components is greater than
frequency domain using masking information as in this. In some cases, a 32 oversampled trapezoidal-
perceptual audio coding (Bosi and Goldberg 2003). rule result also serves as a highly accurate reference
In this article, the audio band is defined to be for comparison. In the Boss DS-1 (Yeh, Abel, and
020 kHz, where a match to the accurate solution Smith 2007a), the signal is hard-clipped to 4.5 V by
of the ODE is desired. Subsonic frequencies are an operational amplifier gain-stage before being
included here, because they can mix through the smoothed by the diode clipper. Therefore, the test
nonlinearity and cause perceptible amplitude inputs are normalized to 4.5 V.
modulation of the output. High frequencies are
assumed to be sufficiently low due to the spectral
roll-off typical of guitar signals such that mixing Two-Tone Sine
products are negligible.
The error criterion for general solvers adaptively A dual-tone excitation (110 and 155 Hz, 4.5-V peak)
adjusts the variable step size to an excessively small was applied at the input of the diode-clipper using
value. When a stiffly stable method is used, an each of the stable integration methods. The implicit
audio-band error criterion greatly improves effi- and semi-implicit methods generate almost identi-
ciency by allowing a larger step size, as explained in cal time-domain responses. Figure 6a shows only
the following. the TR and static approximation. Figure 6b and 6c
plot the error of the 8-oversampled methods
relative to the 32-oversampled reference. All of the
Oversampling
numerical methods exhibit similar profiles with
It is well known that a nonlinearity expands the low error. The second-order-accurate methods TR
bandwidth of a signal. In the digital domain, this and BDF2 have almost identical error. Semi-implicit
extra bandwidth folds over at the Nyquist fre- and implicit versions of the same method have
quency, resulting in aliasing. To achieve high- almost identical error for these low frequencies.
quality distortion effects with a base sampling rate The static approximation shows noticeably larger
of 48 kHz, eight-times (8) oversampling is typically error than the numerical solvers, but it is typically
Yeh et al. 31
Figure 6. Time-domain and BDF2 are almost iden- Figure 7. Peaks in spectra
results for 110 Hz + 155 Hz tical, both being second- of responses to 110 Hz +
input. (a) Waveforms for order. (c) Error for BE, TR, 155 Hz input, connected
trapezoidal (TR) and static and semi-implicit versions by solid lines, for (a) semi-
approximation, 8 over- BE s-i, TR s-i. Only BE and implicit trapezoidal (TR
sampling. They are indis- TR can be distinguished s-i) and (b) static nonlin-
tinguishable in the figure. here, because semi- earity approximation. The
(b) Error for Backward implicit is practically other methods are practi-
Euler (BE), TR, BDF2, and identical to fully implicit cally identical to TR s-i.
static approximation. TR for low-frequency input.
(a) (a)
(b)
(b)
Figure 8
(a) (b)
Figure 9
Single High-Frequency Sine and Verification The same input was provided to the SPICE
with SPICE simulator LTspice (Linear Technology 2007), which
has WAV-file import capability, using trapezoidal-
A high-level, high-frequency sine-wave excitation rule integration. This comparison verifies the
(4.5 V, 15,001 Hz) reveals inadequacies in the semi- methods and approximations used in this work.
implicit methods, which exhibit overshoot in the An exact time-domain match to SPICE cannot be
time-domain plots (see Figure 9) and spurious tones expected because of differences in convergence
in the frequency domain. criteria and numerical handling. SPICE uses an
Yeh et al. 33
Figure 10. Magnitude each at 8 oversampling,
response of (a) 15,001 Hz, and (e) using LTspice,
4.5 V input, 32 oversam- which linearly interpolates
pled reference using (b) TR, output to the 8 over-
(c) TR s-i, and (d) static, sampling grid.
Sine Sweep
(a)
Figure 12
(a)
(b)
(b)
(c)
Figure 13
Yeh et al. 35
Figure 14. Log spectro-
grams of sine sweep from
20 Hz20 kHz: (a) DS-1
and (b) simulated model.
number of iterations n averaged over the 32-sample cal model. Huovilainen (2004) successfully applied
frame is assumed to be 1.8, as suggested herein. an explicit method, because the Moog filter is
typically weakly nonlinear. When systems become
strongly nonlinear, they may operate in device
Discussion regions that result in high-frequency poles, which
cause stability problems with explicit methods as
shown here. Convergence by Newtons method for
Choice of Method circuits with strongly nonlinear regions is also
difficult in general.
The evaluation of cost confirms prior findings in For audio-frequency input, the differences be-
the circuit-simulation literature that implicit tween the methods are negligible in the audio band
methods are preferred over explicit ones for simula- because the process is well oversampled to reduce
tion of general circuits. The explicit methods, aliasing, especially when dealing with clipping-type
although simple, do not produce reliably accurate distortions. The oversampling causes the errors of
results for the diode clipper ODE, as measured in the various accuracy-order methods to be very low
the frequency domain, unless they are impractically in the audio band and makes the effect of frequency
highly oversampled. This need for excessive over- warping insignificant. Thus, complicated higher-
sampling (38 for FE and 30 for explicit RK4) was accuracy-order methods such as extrapolation tech-
found to be necessary to avoid numerical instability niques or implicit Runge-Kutta are unnecessary. It
associated with a high-frequency pole in the physi- would seem then that a stable method of low order
Yeh et al. 37
Figure 16. Power chord.
Left: number of iterations
per sample; right: moving-
average of iterations using
a frame size of 256 samples.
Top to bottom: BE, TR,
BDF2.
would be sufficient while guaranteeing bounded place within the iterative loop to evaluate the de-
output if a convergent nonlinear solver is used. The tailed nonlinear models of circuit devices (McCalla
time-domain outputs of the semi-implicit methods 1987). This was also demonstrated by Karjalainen
show significant ringing for high-frequency inputs, and Pakarinen (2006), who showed that tabulation
but this is an extreme case because high amplitudes of the device models significantly speeds up the
at these frequencies are rarely encountered in iterative solution in modeling a tube preamplifier
practical guitar signals as demonstrated by the with WDF. The computational complexity in the
sound examples. WDF approach is similar to that here, which sug-
gests that full implicit solution of circuits should
be feasible.
Real-Time Considerations Scaling to circuits with more nodes should be
possible as long as device models remain simple to
A prototype implementation, with no particular compute or if they are tabulated. This is true
efforts to write efficient code, demonstrates that the especially because the goal of guitar distortion is to
iterated implicit methods run in real time up to 8 model amplifiers with vacuum tubes whose charac-
oversampling on a contemporary CPU (Intel Core 2 teristics should be measured and tabulated for
Duo, 1.6 GHz). greatest realism. Even simplified models, though,
Experience in circuit-simulator development has may prove sufficient to capture the dynamics or
shown that most of the computational effort takes character of the circuit.
Yeh et al. 39
that in the near future, full simulations can be done Audio Coding and Standards. Norwell, Massachusetts:
in real time. If successful, this could harness the Kluwer.
existing skill of circuit designers and hobbyists, Brown, J. C. 1991. Calculation of a Constant-Q Spec-
who can then experiment with their ideas and tral Transform. Journal of the Acoustical Society of
implement them on a real-time digital platform. America 89(1):425434.
Butcher, J.C. 1987. The Numerical Analysis of Ordinary
Because component models are not limited by
Differential Equations. Hoboken, New Jersey: Wiley.
physical reality and availability, they can experi- Doidic, M., et al. 1998. Tube Modeling Programmable
ment with various parameters of the device elec- Digital Guitar Amplification System. U.S. Patent
tronics (tubes, transistors, etc.) and even invent 5,789,689.
fictional ones to discover if their conjectures about Fernndez-Cid, P., J. Quirs, and P. Aguilar. 1999. MWD:
the nature of various phenomena in musical elec- Multiband Waveshaping Distortion. Proceedings of
tronics are true. Continued development in real- the 2nd COST-G6 Workshop on Digital Audio Effects
time circuit simulation (e.g., real-time SPICE) of (DAFx99). Trondeim, Norway: NoTAM. Available
musical electronics has the potential to spark a online at citeseer.ist.psu.edu / 393243.html.
paradigm shift in digital audio effects. Fettweis, A. 1986. Wave Digital Filters: Theory and Prac-
tice. Proceedings of the IEEE 74(2):270327.
Frnken, D., and K. Ochs. 2001. Synthesis and Design
of Passive Runge-Kutta Methods. AE Interna-
Acknowledgments tional Journal of Electronics and Communications
55(6):417425.
Thanks to Dr. Stefan Bilbao for eye-opening discus- Frnken, D., and K. Ochs. 2002. Improving Wave Digital
sions regarding numerical methods. Many thanks to Simulation by Extrapolation Techniques. AE Inter-
Dr. Bill Gear for providing example code to plot stabil- national Journal of Electronics and Communications
ity regions. Thanks to the Acoustics Laboratory at the 56(5):327336.
Helsinki University of Technology for being terrific Frnken, D., J. Ochs, and K. Ochs. 2005. Generation
hosts during the completion of this article. David of Wave Digital Structures for Networks Containing
Yeh was supported by the Stanford, NDSEG, and Multiport Elements. IEEE Transactions on Circuits
National Science Foundation Graduate Fellowships. and SystemsI 52(3):586596.
Gear, W. C. 1971. Numerical Initial Value Problems in
Ordinary Differential Equations. Englewood Cliffs,
New Jersey: Prentice-Hall.
References Goetze, T. 2007. CAPS, the C Audio Plugin Suite.
Available online at quitte.de / dsp / caps.html, accessed
Abel, J. S., and D. P. Berners. 2006. A Technique for 2 November 2007.
Nonlinear System Measurement. Proceedings of the Gustafsson, F., et al. 2004. System and Method for Simu-
Audio Engineering Society 121st Convention, Paper no. lation of Non-Linear Audio Equipment. U.S. Patent
6951. New York: Audio Engineering Society. Application Publication US2004 / 0258250 A1.
Arfib, D. 1979. Digital Synthesis of Complex Spectra Hlie, T. 2006. On the Use of Volterra Series for Real-
by Means of Multiplication of Nonlinear Distorted Time Simulations of Weakly Nonlinear Analog Audio
Sine Waves. Journal of the Audio Engineering Society Devices: Application to the Moog Ladder Filter.
27(10):757768. Proceedings of the 2006 International Conference on
Barbati, S., and T. Serafini. 2002. A Perceptual Approach Digital Audio Effects (DAFx-06). Montreal, Canada:
on Equalization. Available online at www.simulanalog McGill University, pp. 712.
.org / eq.pdf, accessed 2 November 2007. Huovilainen, A. 2004. Nonlinear Digital Implementa-
Berners, D. P., and J. S. Abel. 2004. Ask the Doctors! tion of the Moog Ladder Filter. Proceedings of the
Universal Audio Webzine 2(6). Available online at 2004 International Conference on Digital Audio Ef-
www.uaudio.com / webzine / 2004 / july / text / content2 fects (DAFx-04), Naples, Italy: Federico II University of
.html, accessed 2 November 2007. Naples, pp. 6164.
Bilbao, S. 2006. Personal communication. Huovilainen, A. 2005. Enhanced Digital Models for
Bosi, M., and R. Goldberg. 2003. Introduction to Digital Analog Modulation Effects. Proceedings of the 2005
Yeh et al. 41
Yeh , D. T., J. Abel, and J. O. Smith. 2007a. Simplified, Proceedings of the 2007 International Conference on
Physically-Informed Models of Distortion and Over- Digital Audio Effects (DAFx-07). Bordeaux, France:
drive Guitar Effects Pedals. Proceedings of the 2007 University of Bordeaux, pp. 197204.
International Conference on Digital Audio Effects Yeh, D. T., and J. O. Smith. 2006. Discretization of the
(DAFx-07). Bordeaux, France: University of Bordeaux, 59 Fender Bassman Tone Stack. Proceedings of the
pp. 189196. 2006 International Conference on Digital Audio Ef-
Yeh , D. T., J. Abel, and J. O. Smith. 2007b. Simulation of fects (DAFx-06). Montreal, Quebec, Canada, pp. 16.
the Diode Limiter in Guitar Distortion Circuits by Nu- Zlzer, U., ed. 2002. DAFX: Digital Audio Effects. Hobo-
merical Solution of Ordinary Differential Equations. ken, New Jersey: Wiley.