PAPERS

Hammerstein Kernels Identification by Means of a

Sine Sweep Technique Applied to Nonlinear Audio
Devices Emulation*

Thomas Schmitz, AND Jean-Jacques Embrechts, AES Member

(T.Schmitz@ulg.ac.be) (jjembrechts@ulg.ac.be)

Department of Electrical Engineering and Computer Science, University of Liege, Belgium

Numerous audio systems are nonlinear. It is thus of a great importance to study them
and understand how they work. Volterra series model and its subclass (cascade Hammerstein
model) are usual ways to model nonlinear systems. In this paper we first describe some diffi-
culties that are encountered during the implementation of these identification methods. Solu-
tions are proposed to solve each of these problems, some are originals, others are inspired by
the literature. The different corrections provided in this paper lead to the Hammerstein Ker-
nels Identifications by Sine Sweep (HKISS) method. To evaluate its efficiency, we emulate a
distortion guitar effect and two tubes guitar amplifiers.

0 INTRODUCTION netic algorithm [12], the extended Kalman filtering [13]

and the particle filtering models [14].
The modeling of nonlinear systems has been a central
topic in many engineering areas, as most real-world de-
vices exhibit nonlinear behaviors. The study of these top-
(.) 1 h1
ics is of great interest as they include a large range of
applications such as industrial processes, control systems,
economic data, life sciences, social systems, physics and
many more. This project focuses on one particularly inter-
x[n] y[n]
esting application: the modeling of a guitar players audio (.) 2 h2
chain. Such a chain is usually composed of compressors,
distortion effects, preamplifiers, amplifiers, loudspeakers
and microphones. Most of these elements exhibit nonlinear
characteristics [1, 2, 3]. The idea is to replace these devices
by a computer emulation that could be suitable for guitar
players in many situations [1, 4]. (.) M hM
Most weakly nonlinear systems can be approximated by
linearization. In that case, the modeling can be achieved
by the convolution technique [5], as for classical Linear
Time Invariant (LTI) systems. However, stronger nonlin-
Fig. 1. The Global Polynomial Hammerstein model
earities need more specific models in order to keep an
accurate approximation of the real system. Such models
include the Volterra [6], the multiple-input-single-output In our research, we focus on a technique based on the
(MISO) [7], the single-input-single-output (SISO) [8], the Volterra series model [6, 15], which recently received an
multiple-input-multiple-output (MIMO) [9], the nonlin- increased interest with the improvement of computers
ear auto-regressive moving average with exogenous input performance. Considering only static (memoryless) non-
(NARMAX) [10], the neural network [11], the hybrid ge- linearities, this model can be simplified into a Hammerstein
model (see Fig. 1) which is a cascade of M parallel (the
systems order) branches where each branch m is com-
* To whom correspondence should be addressed Tel: +32-4366- posed by a static monomial nonlinearity (putting the input
2706; e-mail: T.Schmitz@ulg.ac.be signal to the power m) followed by a linear filter hm [n]

J. Audio Eng. Sco., Vol. 1, No. 1, 2015 October 1

Th.Schmitz AND J.J.Embrechts
which is the mth Hammerstein kernel (also called diagonal
Volterra kernel) [7, 15, 16, 17].
In the last few years, a nonlinear convolution method
[18] has been elaborated with the aim of emulating nonlin-
ear audio systems. Farina et al. [18, 19] indeed proposed
to use the properties of the Exponential Sine Sweep (ESS)
signal in order to identify the Hammerstein kernels. With
this signal at the input of a nonlinear system, they showed
that it is possible to separate from each other the har-
monic responses created by the harmonic distortions. This
method has been quite successfully applied to model non-
linearities in loudspeakers [1, 20], guitar distortion gener-
ators [21, 22] (the famous Tube-Screamer is often taken as
an example) and compressors [23].
Recently, Novak et al. [24] have brought important im-
provements to the ESS method with their use of a syn-
chronized swept-sine (SSS) signal. This new formulation
of the sine sweep signal corrects some inaccuracies that
were inherent to the original implementation, in particular
the problem of the phase residue that will be described in
the next section. In the same paper, these authors have also
used an efficient analytical formula for the deconvolution
of the sweep signal. Also recently, Vanderkooy et al. [25]
have proposed a different method to correct the problem of
the phase residue.
In this paper, we first highlight some difficulties that
must be considered in the implementation of the ESS
method, in order to perform an accurate identification of
nonlinear (NL) systems. The nature of these problems is
described and solutions (either original or inspired by the
literature) are proposed for each of them. In a second part
of the paper, these solutions are validated by the emulations
of real nonlinear audio devices.
The paper is organized as follows : the basic identifica-
tion technique is recalled in section 1. This, theory is illus-
trated with a simple third-order model and then extended
to a given order M. In section 2 , five problems occur-
ring in the implementation of the identification method are
described. We propose an algorithm for testing the recon-
struction of the Hammerstein kernels and use it to illustrate
the methods that can solve these problems. This leads to the
definition of an Hammerstein Kernels Identification by Sine
Sweep (HKISS) method. In section 3, we apply this method
and prove its efficiency on three real nonlinear audio de-
vices: two tubes amplifiers and a distortion guitar effect.
The section 4 offers an overview on the limitations of the
HKISS method. Finally section 5 presents a Matlab toolbox
helping to the implementation of the HKISS method.
1 BASIC IDENTIFICATION METHOD
This section briefly presents the principles of the
Volterra series and the Hammerstein model. In order to
ensure a good understanding of the method, we illustrate it
with a third-order model, before extending it to a general
order M.
2 J. Audio Eng. Sco., Vol. 1, No. 1, 2015 October
PAPERS
1.1 Volterra Series
Considering time invariant nonlinear systems, the
Volterra series expresses the relationship between an input
signal x(t) and its corresponding output signal y(t) as an
infinite sum of multiple convolutions [6, 15]:
Z + Z +
y(t) = ... vm (1 , . . . , m )
m=1
m (1)
x(t i ) d1 . . . m
i=1
Where {vm (1 , . . . , m )}mN+ are the Volterra kernels
characterizing the system.
1.2 Cascade of Hammerstein Models
It can be shown that any continuous nonlinear system
can be represented by a set of M parallel branches com-
posed by a static non linearity Pm (.) surrounded by two
linear filters [26, 27]. The cascade Hammerstein model
is a subclass of this general model where we consider a
static nonlinear monomial function followed by a linear
filter hm (t) (the Hammerstein kernels) as shown in Fig.1.
Considering a nonlinear discrete time causal system with a
finite order M, we can express the input/output relationship
as:
M
y[n] = hm [n] ~ xm [n] (2)
m=1
Where ~ is the symbol of the convolution operator. We can
see from Eqs. (1) and (2) that the Hammerstein model cor-
responds to the Volterra series model if the Volterra kernels
are diagonal (i.e. the Volterra kernels differ from zero only
for the values 1 = 2 = ... = M ). That is why the cascade
of Hammerstein models is also called a diagonal Volterra
model.
1.3 Hammerstein Kernels Identification with the
ESS method
In order to extract the Hammerstein kernels, we need
to excite the nonlinear system with a suitable input sig-
nal. Such a signal should allow us to separate the different
Hammerstein kernels from each other after the deconvolu-
tion by its inverse filter. The next two subsections explain
which suitable input signal is chosen and how it can be
used to compute the Hammerstein kernels.
1.3.1 Exponential Sine Sweep Phase Properties
In the following, our development starts from the origi-
nal formulation of Farina et al. [18, 19]. The equations are
expressed in discrete time for a better correspondence with
the practical implementation. Let ss[n] be the ESS signal:
ss[n] = A sin( [n]) (3)
The phase [n] grows exponentially such as :
R n
[n] = 1 .(e R 1) n [0, N 1] (4)
fs
Where:
PAPERS Hammerstein Kernels Identification by Means of a Sine Sweep Technique

r R = (N 1).(log 2 )1 is the inverse frequency chang- The goal is to obtain M equations giving the M Ham-
1
ing rate. merstein kernels hm [n] in Eq. (2). The method consists in
r N is the length of the sweep in samples. exciting the nonlinear system by the ESS given in Eq. (3)
r 1 , 2 are the initial and final angular frequencies re- and then to deconvolve the result by its inverse filter.
spectively.
r fs is the sample rate. Including the ESS signal (3) as an input of the model
from Eq. (2) gives:
An interesting property of the ESS [1] is that the mth har-
y[n] = A sin( [n]) ~ h1 [n]
monic of the ESS can be derived from the ESS itself de-
layed by m , as: + A2 sin2 ( [n]) ~ h2 [n] (9)
3 3
m. [n] = [n + m ] B(m 1) n N 1 m (5) + A sin ( [n]) ~ h3 [n]
Where: Using the trigonometric formulas [34] to decompose in
B = 1 .R order-one formulation, we have:

fs (6) y[n] = A sin( [n]) ~ h1 [n]
m = R. log(m)

A2
Choosing B 2k, with k N allows to neglect the B [cos(2 [n]) 1] ~ h2 [n] (10)
2
term in Eq. (5) as it just adds a multiple of 2 to the sine A3
phase. However, this constraint is not mandatory as it will + [3 sin( [n]) sin(3 [n])] ~ h3 [n]
4
be seen later. Note also that the values m are real and
non-integer (except for m=1). This will necessitate a phase In order to deconvolve the signal y[n] properly, we want
correction during the extraction of the harmonic impulse to change the cosine term in a sine expression. This can be
responses. This will be explained in subsection 2.4. done using the Hilbert Transform [32],[33] :

H {sin(u(t))} = sin(u(t)) ~ h(t) = cos(u(t)) (11)

The property of the ESS signal in Eq. (5) is interesting
because the contribution of an harmonic m in the impulse Where H {.} is the Hilbert Transform and h(t) = 1
t . In-
response will be delayed by m from the first order im- troducing (11) in (10) and using (5) gives:
pulse response, after its deconvolution by the ESS inverse
filter ss[n]. In the particular case of B 2k, we can write: y[n] = A sin( [n]) ~ h1 [n]
N1 A2
ss( [n]) ~ ss( [n]) =

ss[k].ss[n k] = C. [n n ] + [sin( [n + 2 ] B) ~ h[n] + 1] ~ h2 [n]

0 2
k=0
A3

ss(m. [n]) ~ ss( [n]) = C. [n (n0 m )] + [3 sin( [n]) sin( [n + 3 ] 2B)] ~ h3 [n]
4
(7)
Where C is an amplitude constant. n0 is the position = ss[n] ~ h1 [n]
along the time axis of the first order (linear) impulse re- " #
sponse which depends on the ESS length (n0 = N 1 in A ss[n + 2 ] cos(B) ~ h[n]
+ ~ h2 [n]
[28]). The inverse ESS can be calculated in several ways 2 ss[n + 2 ] sin(B)
[17, 24, 28, 29, 30, 31]. According to [24], we can com-
3ss[n]

pute an analytical version of the Fourier transform of ss[n]: 2
A
+ ss[n + 3 ] cos(2B) ~ h3 [n]

r h    
j L 1ln( ) 1 4
i
4
SS() = 2 e 1 (8) ss[n + 3 ] sin(2B) ~ h[n]
2L (12)
The advantages of the analytical inverse sine sweep
computation are that C=1 in the equation Eq. 7 and the The DC offset of the even powers has been neglected as
bandwidth of the inverse filter can be extended until the it is only related to the mean value of the ESS. Moreover,
Nyquist frequency if needed. We must note at this point in practice, most audio devices stop the DC component.
that if m 6= 1, the second convolution in Eq. (7) is not an ex-
act delayed version of the Dirac pulse with the same ampli- In the Fourier domain, using the trigonometric exponen-
tude as in the case m = 1. This problem will be addressed tial property, the Euler formula and the Hilbert Transform
in section 2.6. property [32], we obtain for positive frequencies:
Y ( j) = SS( j).H1 ( j)
1.3.2 A Simple Third Order Model  
A jB
j2
In order to illustrate the origin of the problems men- SS( j). j.e .e f s .H2 ( j)
tioned in the introduction, we propose to establish the 2
A2
 jw3

Hammerstein kernels equations for a simple Hammerstein j2B
+ 3.SS( j) SS( j).e e s .H3 ( j)
f
model developed up to the third order. An extension of the 4
model up to any order M is presented in subsection 1.3.3. (13)

J. Audio Eng. Sco., Vol. 1, No. 1, 2015 October 3

Th.Schmitz AND J.J.Embrechts PAPERS

With the following Hilbert Transform property: 1.3.4 Deconvolution Process

Following the same steps as in section 1.3.2 but apply-
F {H (s(t))} = j.sign()F {s(t)} (14) ing (18) for sine powers, the deconvolution of the Hammer-
stein model (Eq. (2)) in the frequency domain leads to:
Note that since the Hammerstein kernels are real, we
jn0 / fs

can use the Hermitian property in the Fourier domain [5] Z( j) = Y ( j)SS( j)e

to find the Hammerstein kernels for negative frequencies: MM
Hm ( j) = Hm ( j).
= H1 ( j) + [am H2m+1 ( j) + bm H2m ( j)]
m=1
By deconvolving the signal Y ( j) with the inverse filter (19)
SS( j) and by grouping in the Gm ( j) terms having the Where MM = (M 1)/2. In the time domain,
Eqs. (18,19) show that z[n] is a superposition of sev-
same phase delay e j fs m , we obtain: eral versions of each Hammerstein kernel hm [n] delayed
3A2
by k (e jk / fs in the Fourier domain). We name gm [n]
G1 ( j) 1 0 H1 ( j)
jB
4 the superposition of the hm [n] having the same delay m .
G2 ( j) = 0 jAe2 0 H2 ( j) (15)

Furthermore, we consider that all these harmonic impulse
G3 ( j) 2
A e j2B H3 ( j)
0 0 4 responses start at n=0, which necessitates a delay of m
applied to all of them:
By inverting this relation, we finally obtain the Hammer-
stein kernels: G1 ( j) H1 ( j)
.. ..
= T (20)

. .
1 0 3e j2B

H1 ( j) jB
G1 ( j) GM ( j) HM ( j)
H2 ( j) = 0 2 je

0 G2 ( j)

(16)
A
H3 ( j) 4e j2B
G ( j) We can develop the elements Tu,v with u, v M using
0 0 2 3
A Eq. (19) in the matrix form and grouping in Gm ( j) the
Hm ( j) having the same phase delay.
1.3.3 Extension to the Order M 1u

The generalization up to the order M can be done by (1) 2

vu u+v
Cv 2 v u & mod( )=0
Tu,v = 2v1 2
means of the trigonometric power formulas of sine [34]:
0 else
1n n (21)

2n+1

sin (t) = n
(1)k Ck2n+1 sin([2n + 1 2k]t) Finally the Hammerstein kernels in the positive fre-

4 k=0
quency domain are given by the following relationship:

n N H1 ( j) G1 ( j)
.. ..

1
= (T) (22)

. .
HM ( j) GM ( j)
1n n1 Cn2n

sin2n (t) = 2n1 (1)k Ck2n cos([2n 2k]t) + 2n

2 2

k=0 1.4 Hammerstein Model And Input Level

Dependency
n N0 The cascade of Hammerstein models takes the amplitude
(17) of the input signal into account, for example, Fig. 2 com-
By choosing B 2k in Eq. (5) and using it in Eq. (17), pares the result of a power series when the input signal is a
we derive power formulas of the ESS signal for the positive sinus of amplitude 0.5 at 1000Hz with its emulation using
frequency domain defining SSn as the Fourier transform of the Hammerstein kernels calculated with an ESS of ampli-
ssn : tude 1. The signal is reconstructed correctly. However, for

1n n a real device, the model of the cascade of Hammerstein

SS2n+1 =SS n (1)k Ck2n+1 e j2n+12k / fs could no longer match perfectly and therefore the Ham-

4 k=0



merstein kernels can be slightly different, according to the


amplitude of the input signal [22, 35]. For example, Fig.
=an .SS n N




3 presents the comparison between an ESS (focus around
(18) the frequency f=1000Hz) with = 0.5 passing through the
distortion part of the amplifier TSA15h [36] and its emula-
1n n1


SS2n = j.SS. 2n1 (1)k Ck2n e j2n2k / f s


tion using Hammerstein kernels calculated with an ESS of




2 k=0 amplitude = 1 ( Hm1 ( j) ). We can see that the signal



coming from the distortion effect (y1 ) is different from the

=bn .SS n N0 emulated signal (y2 ).
From now on Hm ( j) are the Hammerstein kernels ob-
Cn2n tained using an ESS with an amplitude :
The term 22n
is the DC component and has been ne-
glected. ss[n] = sin( [n]) (23)

4 J. Audio Eng. Sco., Vol. 1, No. 1, 2015 October

PAPERS Hammerstein Kernels Identification by Means of a Sine Sweep Technique

For the rest of the paper, the symbol over the Hammer- the Nyquist band in such a way that the created model is
stein kernels will be omitted if = 1. equivalent to a power series.
Using the ESS as input signal, the output y1 is the result
1 of the power series (see section 2.7 for aliasing consider-
0.8 y1 ation). Deconvolving y1 with the ESS inverse filter ss[n]
y2 allows to calculate the Hammerstein kernels of this nonlin-
0.6 ear system. Then, using Eq. (2) we can compute a signal
y2 based on the Hammerstein kernels and the ESS powers.
Amplitude

0.4
If the Hammerstein kernels are correctly reconstructed, y1
0.2 and y2 have to be identical.
0
ss[n] (.) M
0.2
ssm [n]
0.4
0 20 40 60 80 100 y1
Sample [n]

Fig. 2. Comparison between a sine of amplitude 0.5 passing ss[n]

ssm [n]
through a series of power (y1 ) and its emulation (y2 ) with Ham- z[n]
merstein kernels measured for an ESS of amplitude 1. hm [n]
Comput e Volt erra Nonlinear
y2
kernels Convolut ion

1
y1
Fig. 4. Test algorithm: y1 and y2 have to be identical
y2

0.5

2.2 The Best Case

Amplitude

Following our top down approach, we first present our

0
best match between the power series y1 (up to the order
M=6) and its emulation by means of the Hammerstein ker-
0.5 nels y2 (see Fig. 5). For clarity, we choose to zoom on a
part of the sweep signals, (arbitrarily) around the frequency
f = 1000Hz. The power series is correctly reconstructed,
1
0 50 100 150 200 as it can be seen in Fig. 5. This best match has been ob-
Sample [n] tained by following all the precautions and corrections de-
scribed in the next subsections.
Fig. 3. Comparison between an ESS of amplitude 0.5 passing
through the distortion effect TSA15h (y1 ) and its simulation (y2 ) 1.8
by means of Hammerstein kernels measured for an ESS of ampli- y1
tude 1. 1.6 y2
1.4

1.2
Amplitude

1
2 OPTIMIZATION OF THE HAMMERSTEIN
KERNELS IDENTIFICATION 0.8

0.6
Eq. (22) can be used to calculate the Hammerstein ker-
0.4
nels. However several difficulties must be taken into ac-
count to make an accurate emulation of a nonlinear sys- 0.2
tem. In order to illustrate them, we choose a top down ap- 0
proach, starting with the optimized case (subsection 2.2)
0 50 100 150
where all the precautions are taken and then removing them Sample [n]
one at a time to prove their necessities (subsections 2.3,
2.4, 2.5, 2.6). Note that other problems such as artifacts in Fig. 5. Comparison between y1 [n] and y2 [n] at 1000Hz (best case)
the impulse responses have been identified and analyzed in
[37, 38, 39].
2.3 Hammerstein Kernels Phase Errors
2.1 Test Algorithm The basic model (16) shows that the Hammerstein ker-
We propose a test algorithm (see Fig. 4) based on the nels depend on a constant factor B introduced in the
Hammerstein model described by Eq. (2). The kernels phase of the ESS. During the general formulation given
hm [n] are chosen to be simple all-pass filters restricted to at Eq. (22) we have made the simplifying assumption that

J. Audio Eng. Sco., Vol. 1, No. 1, 2015 October 5

Th.Schmitz AND J.J.Embrechts PAPERS

B 2k, with k N (see section 1.3.1 ). Forcing B close y

enough from 2k would imply to tweak the ESS parame- 1.2 1
y
2
ters. This constraint and other conditions discussed in the 1
next subsection would to a difficult optimization problem 0.8
that could have no exact solution, leading to compromises 0.6

Amplitude
on the ESS parameters values.As mentioned in the intro- 0.4
duction, Novak et al. [24] have overcome this difficulty by
0.2
their use of a synchronized swept-sine signal. In this pa-
0
per based on the original formulation of the sweep signal
[18, 19], we rather propose to directly include a correction 0.2

by multiplying term by term the (T)1 matrix with CorrB , 0.4

a corrective M by M matrix, taking into account the B fac- 0.6
tor: 0 50 100 150 200
Sample [n]
j0
e . . . e j(v1)B

CorrB (u, v) = ... . . . .. Fig. 7. Comparison between y1 [n] and y2 [n] at 1000Hz when the
(24)

. B correction is not applied
e j0 . . . e j(v1)B
This correction is equivalent to the one given in [25].In phase mismatch between the kernels gm [n] that impacts the
this last paper, the authors have introduced a generalized computation of the Hammerstein kernels (see Eq. (22)).
form of the Hilbert Transform which results in shifting all
the phases by the same amount, for all frequencies. In-
deed, after application of the CorrB matrix in Eq. 24, all
the phases of the Hammerstein kernels are equal to zero Ceil(4 ) Ceil(3 ) Ceil(2 )
g1 [n]
(see Fig.6), which is the condition imposed in [25]. g2 [n]

Fig. 7 shows the comparison between y1 and y2 when g4 [n]

g3 [n] 2

the matrix CorrB is not applied (with mod(B, 2) = 0.7). 3

4 [n]
The emulated signal is very sensitive to the Hammerstein
kernels B factor as it no longer corresponds to the power
4 3 2 N-1
series signal.

Hv1 Fig. 8. Deconvolution of an ESS passing through a power series,

150 Hv2 phase mismatch m
Hv3
100 Hv4
Hv5
Hv6 Choosing sweep parameters in order to minimize the m
50
could be very restrictive and difficult when the order M in-
Phi

0 creases. Novak proposes in [24] to overcome this problem

by shifting, the signal z[n] by (m Ceil(m )) before ex-
50
tracting each gm [n]:
100
j fs (Ceil(m )m )
Zm ( j) = Z( j)e
1 (25)
150
This phase mismatch mostly introduces some recon-
0.5 1 1.5 2
Hz 4 struction errors at high frequencies. Fig. 9 shows the com-
x 10
parison between y1 and y2 if we dont apply the phase mis-
Fig. 6. Hammerstein kernels phase diagrams of a power series match correction. A significant difference between the two
after application of the matrix CorrB signals starts at 3000Hz and increases with frequency.

2.5 ESS Fade In/Out

2.4 Phase Mismatch
The harmonic impulse responses (gm [n]) appear in the
deconvolved signal z[n] at specific positions on the time
axis. They can be extracted, starting at n0 m for the
first sample. However, no matter how we adjust the ESS
parameters, it is impossible to satisfy the condition m
N, m [1, M] (except for the trivial case M = 1). There-
fore, if we extract gm [n] from n0 Ceil(m ) an error m is
created on the starting position as it could be seen at Fig.
8. As this error m depends on the index m, this leads to a
6 J. Audio Eng. Sco., Vol. 1, No. 1, 2015 October
In order to avoid spectral leakage due to abrupt termi-
nation of the ESS, appropriate fade-in and fade-out proce-
dures have been applied, according to [24]. If the loss of
bandwith is expressed as a fraction of f1 and f2 , for the
fadeIn and fadeout respectively, then:
length f adeIn = R log(1 + loss)
1 loss (26)
length f adeOut = R log( f2 )+1N
f1
PAPERS Hammerstein Kernels Identification by Means of a Sine Sweep Technique

this offset. Thus,we propose to cut the kernels gm [n] at

y1
1.2 n0 m 1000, compute Hammerstein kernels and then
y2
1
delete the first thousand samples of the Hammerstein ker-
nels.
0.8 Fig. 11 shows that this method gives excellent results.
Amplitude

In the case where the bandwidth is short (measurement of

0.6
a tweeter for example) the Dirac impulses become larger
0.4 and larger. Simply cutting the first thousand samples after
the computation of the Hammerstein kernels could lead to
0.2 a wrong representation. In this case, a compromise has to
0
be found between the accuracy of the Hammerstein ker-
nels and their possible use in real-time emulations. In the
0 5 10 15 20 25
Toolbox described in section 5, the choice of the offset be-
Sample [n] fore the beginning of the Hammerstein kernels is therefore
left open to the user, depending on the bandwidth and the
Fig. 9. Comparison between y1 [n] and y2 [n] at 6000Hz when the real-time constraint of the users application.
phase correction from Eq. (25) is not applied (y2 [n] perfectly
matches y1 [n] if the correction is applied)
One can notice that the sampling rate used here is al-
ways 44100Hz as we try to ensure compatibility with basic
sound cards. Nevertheless, the use of higher sampling rate
2.6 Spreading of the Dirac Impulse allows a better compactness of the impulse response as the
We have assumed in Eq. (7) that the deconvolution of harmonic m extends its bandwidth to the frequency m. f2 or
ss(m. [n]) is a delayed Dirac impulse [n (n0 m )]. until the Nyquist frequency of the sampling rate is reached.
As the signal is band-limited, the Dirac impulse is in fact
a Sync function m0 . It follows that each measured kernel 1.2
y1
gm [n] is in fact: 1 y2

gm,measured [n] = gm [n] ~ m0 (27)

0.8
Moreover, as the signal ss(m. [n]) starts at frequency
Amplitude

m. f1 instead of f1 , m0 is more and more band limited as 0.6

the order m increases. These limitations spread the kernels

0.4
gm [n] around the positions n0 m .
0.2
Extracting the kernels gm [n] from z[n], by starting at
the exact positions m can therefore be problematic. We 0

prefer cutting the kernels gm [n] a few samples before that

0.2
position to have a better reconstruction of the Hammerstein 0 50 100
Sample [n]
150 200
kernels. In practice, 1000 samples have been chosen and
there is no need to take a larger number of samples (at Fig. 10. Comparison between emulated signals cutting gm [n] at
fs = 44100Hz). For example, Fig. 10 shows the compari- n0 m 1000 (see y01 ) and at n0 m (see y2 ).
son between the emulated signal using kernels gm [n] cut at
n0 m 1000 (signal y01 ) and the emulated signal using
1.8
kernels gm [n] cut at n0 m (signal y2 ). The signal y1 is y1
better aligned to the signal y1 from Fig. 5. Table 1 shows 1.6 y2
the mean error between y1 and y2 over 2 periods around 1.4
the frequency f=1000Hz for different cutting offsets. 1.2
Amplitude

1
Table 1. Mean error around 1000Hz between y01 and y2 for 0.8
different offsets
0.6

0.4
cutting offset -1000 -500 -100 -10 0
0.2

Mean error 0.005 0.007 0.014 0.062 0.155 0

Note. Bandwidth = [5,22000]Hz with fs = 44100Hz 0 50 100 150

Sample [n]

For real time emulations, introducing such a delay in Fig. 11. Comparison between the output power series y1 and the
the impulse responses could affect the output signal in the emulated signal y2 calculated by cutting gm [n] at n0 m 1000
same way since the impulse responses will be convolved and deleting the first thousand samples of hm [n]
with the signal to emulate, this signal will be delayed by

J. Audio Eng. Sco., Vol. 1, No. 1, 2015 October 7

Th.Schmitz AND J.J.Embrechts PAPERS

2.7 Computing the Powers of the Input Signal compares the signal y1 coming from the distortion device
Eq. (2) shows that implementing the Hammerstein with the signal y2 computed with the HKISS method using
model requires the powers of the input signal. Some 6 Hammerstein kernels. The two curves are very close. The
precautions must be taken to avoid aliasing. In order to mean error is :
raise the input signal x[n] to the power m, the following
procedure is used: y1 [n] y2 [n]
N1 | max(|y |) max(|y2 |) |
1
< 1% (28)
r Resample the input signal at m. fs by zeros padding n=0 N
where fs is the original sample frequency. The spectral contents of a pure tone (500Hz) passing
r Filter the signal in the initial Nyquist band fny = fs . through the distortion device or through the simulator (us-
2
r Raise the obtained signal to the power m. ing 9 Hammerstein kernels) are compared in Table 2. A
r Filter the result in the Nyquist band fny . deviation smaller than 1.5db up to the 9th harmonic can be
r Downsample the filtered signal at the fs frequency. noticed.
0.4
y
3 MEASUREMENTS AND EMULATIONS 1
0.3 y2
In the precedent section, a set of precautions necessary
0.2
to a better computation of the Hammerstein kernels has
been presented. Taking these precautions (on phase error, 0.1
phase mismatch, last sample, spreading of the Dirac im-
0
pulse) leads to a Hammerstein Kernels Identification by
Sine Sweep (HKISS) method. This section presents several 0.1
tests made on real nonlinear audio devices in order to prove
0.2
the efficiency of this method.
0.3
3.1 Evaluation Algorithm
0.4
We propose two methods to measure the accuracy of the 0 20 40 60 80 100 120 140 160

emulated signal.
1. The first method, presented in Fig. 12, compares the
output signal y1 coming from a nonlinear device driven
by an ESS with its emulated signal y2 obtained using the
Hammerstein kernels calculated from y1 .
2. The second method compares the spectral content of the
signal coming from the nonlinear device and from the
simulator when they are driven by a pure tone.
NL device
ss[n]
under t est
ss[n]
Fig. 13. Comparison between the sweep signals y1 and y2 around
1000Hz for the preamplifier distortion TSA15h
Table 2. Harmonic amplitudes at the output of the TSA15h
distortion for a pure tone input signal at 500Hz
Harmonic order Distortion effect(db) Simulation(db)
1 92.2 92.2
3 65.3 64.9
5 56.4 56.1
y1
7 51 50.8
9 46.5 45

Comput e Volt erra Note. Only the odd harmonics are presented here as the distortion
(.) M
kernels z[n] effect clips the signal in a symmetric way.

hm [n]

Nonlinear
y2
Convolut ion
ssm [n]
Fig. 12. Comparison between the signal coming from the nonlin-
ear device under test y1 and its emulation y2
3.2 A Guitar Distortion Effect
The first nonlinear audio device under test is the distor-
tion effect coming from the guitar Tube Screamer Ampli-
fier TSA15h [36] with its knob Gain at its maximum. Fig 13
3.3 Tubes Amplifier TSA15h
One of the most popular application in the field of audio
nonlinear modeling is the emulation of tubes amplifier so
dear to guitarists. The emulation of the tubes amplifier
TSA15h with all the knobs Gain at their maximum is pre-
sented in Figs. 14 ,15 & 16 around the frequencies 200,
1000 and 5000Hz respectively. The simulations result y2
(using 6 Hammerstein kernels) is very close to the mea-
sured signal y1 , especially at middle and high frequencies
The comparison of the spectral content for a pure tone of
500Hz passing through the amplifier or through the simu-

8 J. Audio Eng. Sco., Vol. 1, No. 1, 2015 October

PAPERS Hammerstein Kernels Identification by Means of a Sine Sweep Technique

lator is slightly worse than for the distortion effect but still 1
y1
acceptable (see Table 3). Several reasons can explain these
y2
differences :
0.5
r the nonlinearities created by the tube amplifier are more

Amplitude
pronounced than those generated by the distortion effect
(see Tables 2&3); 0
r the tube amplifier has an output transformer whose hys-
teresis behavior can discard from a simple Hammerstein
system; 0.5
r the tubes power compression (the sagging effect [41]) is
not a time invariant effect.
1
0 5 10 15 20 25
Sample [n]
Finally, we have compared the signals coming from the
amplifier and from the simulator using different numbers
Fig. 16. Comparison between the sweep signals y1 and y2 for the
of Hammerstein kernels during the emulation (see Fig. 17). tubes amplifier TSA15h around 5000Hz
A good agreement can already be obtained with only few
Hammerstein kernels (around five or six) which is interest-
ing for real time implementations.
1.5
y1
0.8 y2(order = 1)
y1
1 y2(order = 5)
0.6 y2
y2(order = 14)
0.4
Amplitude
0.5
0.2
Amplitude

0 0

0.2
0.5
0.4

0.6 1
0 20 40 60 80 100
Sample [n]
0.8
0 100 200 300 400 500
Sample [n]
Fig. 17. Comparison between sinusoidal signals y1 and y2 for
the tube amplifier TSA15h at 500Hz with different order of
Fig. 14. Comparison between the sweep signals y1 and y2 for the
simulation
tubes amplifier TSA15h around 200Hz

0.8
y1 between y1 and y2 around 200Hz and 1000Hz respectively.
0.6 y2
Note that the system is strongly nonlinear. In this case, the
0.4 emulation has been made with 14 Hammerstein kernels.
A little overshoot can be observed on the emulation signal
0.2
at 1000Hz. The mean error as defined by (28) is less than
Amplitude

0 2.3%.
0.2

0.4
Table 3. Harmonic amplitude at the output of the
0.6 amplifier TSA15h for a pure tone input signal at 500Hz
0.8
0 20 40 60 80 100 120
Sample [n] Harmonic order Tube amplifier(db) Simulation(db

Fig. 15. Comparison between the sweep signals y1 and y2 for the 1 99.5 99.5
tubes amplifier TSA15h around 1000Hz
2 89 88.9
3 86.1 84.5
3.4 Tubes Amplifier: Engl Retro Tube 50 E762 4 79.3 78.3
Another tubes amplifier has been tested, the Engl retro 5 75.4 76.7
tube 50 E762 [42]. Figs. 18 & 19 present the comparison

J. Audio Eng. Sco., Vol. 1, No. 1, 2015 October 9

Th.Schmitz AND J.J.Embrechts PAPERS

0.8 1.4
y y
1 1
0.6 y2 y
1.2 2

0.4
1
0.2

Amplitude
Amplitude

0.8
0
0.6
0.2
0.4
0.4
0.2
0.6

0
0.8
0 100 200 300 400 500 600 0 50 100 150 200
Sample [n] Sample [n]

Fig. 18. Comparison between y1 and y2 for the tubes amplifier Fig. 20. Comparison between the sweep y1 and y2 for a power
Engl at 200Hz. series up to the 20th order around 1000Hz

0.6 0.5
y1 y1
0.4 y2 y2
0.4

0.2 0.3
Amplitude

Amplitude
0 0.2

0.2 0.1

0.4
0

0.6
0.1

0.8
0 50 100 150 200 0.2
Sample [n] 0 5 10 15 20 25 30 35 40
Sample [n]

Fig. 19. Comparison between y1 and y2 for the tubes amplifier Fig. 21. Comparison between the sweep y1 and y2 for a power
Engl at 1000Hz.
series up to the 20th order around 16000Hz

4 DISCUSSION AROUND THE LIMITATIONS OF (the upper bound could be 2 if we initially chose 2 equal
THE METHOD to the Nyquist frequency). It means that the deconvolution
The HKISS method is pretty stable. Applying again the by the inverse filter is also band limited in [m.1 , m.2 ]
test algorithm of section 2 (power series model) up to the which gives a band limited Dirac impulse m .
order 20 causes no problem, as it can be seen in Fig. 20. Eq. (7) can be rewritten as:
The model is also accurate for high frequencies (see Fig. ss (m. [n]) ~ ss( [n]) = D.m [n (n0 m )] (29)
21). The remaining limitations are discussed in this section.
In the frequency domain, the impulse m delayed by m
4.1 Order Limitation is a step function I( j) on the domain [m.1 , m2 ] (see
As the amplitude of the kernels gm [n] decreases when the Fig. 22). Therefore :
order m increases, it finally reaches the background noise. SSm ( j).SS( j) = I[m.w1 ,m.w2 ] ( j) (30)
This creates a maximum order limit. Improving the signal
to noise ratio (SNR) is thus important if high order em- In order to understand the impact on the Hammerstein
ulation is needed. It could be convenient to take several kernels computations, lets take Eq. (13) again and con-
measurements of a device and averaging the results in or- volve this output signal y(t) with the ESS inverse filter us-
der to improve the SNR. However, we must be sure that the ing Eq. (30). We obtain:
system is perfectly time invariant. The measurement of a 3A2

G1 ( j) = I
[w1 ,w2 ] .H1 ( j) + H3 ( j).I[w1 ,w2 ]

loudspeaker at high power for example could be difficult

4
as the heating of the coil changes its characteristics.

A

G2 ( j) = j.e jB .H2 ( j).I[2.w1 ,2.w2 ]
2
4.2 Low Frequency Limitation

2

A
e j2B .H3 ( j).I[3.w1 ,3.w2 ]

G3 ( j) =

As the ESS is band limited in the frequency interval 4
[1 , 2 ], its mth harmonic is band limited in [m.1 , m.2 ] (31)

10 J. Audio Eng. Sco., Vol. 1, No. 1, 2015 October

PAPERS
100
95
90
Amplitude db
f1 3.f 6.f
1 1
85
80
75
70 1 2
10
Frequency in Hz
Fig. 22. FFT of m with m {1, 3, 6}
By inverting the relation to have the Hammerstein ker-
nels, we have:
4 transforms are compared in Fig. 24. It involves that we can
H3 ( j).I[3.w1 ,3.w2 ] = 2 e j2B .G3 ( j)
A
2 jB
H2 ( j).I[2.w1 ,2.w2 ] = j.e .G2 ( j)
A
3A2
[w1 ,w2 ] = G1 ( j)
H ( j).I H3 ( j).I[w1 ,w2 ]
1
4 on the measure at full amplitude, we can emulate the sig-
(32)
This last expression indicates that to calculate
H1 ( j).I[w1 ,w2 ] , H3 ( j).I[w1 ,w2 ] must be known but we
only have H3 ( j).I[3.w1 ,3.w2 ] . However the formula is
still correct if only the domain [31 , 32 ] is considered.
Therefore, we can calculate the Hammerstein kernels for a
nonlinear system of order M only if:
valid M.1
To illustrate the problem, the Fig. 23 presents the com-
parison in frequency domain between the first Hammer-
stein kernel of a power series (up to the 6th order) and the
first Hammerstein kernels reconstructed with the HKISS
method. The reconstructed one does not correspond to the
initial one until it reaches the frequency f = 6 f1 .A way
to get around this problem is to choose the ESS parameter
1 according the desired bandwidth. For example, if we
would like to emulate a 10 order nonlinear system on the
bandwidth [100, 20000]Hz, we must take f1 = 10Hz.
4.3 Hammerstein Kernels Input Level
Dependency
As discussed in section 1.4, the accuracy of the emu-
lated signal depends on the match between the input sig-
nal amplitude and the amplitude of the sine sweep used
to measure and compute the Hammerstein kernels. During
the emulation test of the amplifier TSA15h, we have no-
ticed that the Hammerstein kernels Hm ( j) have the same
shape for a given order m but have different relative ampli-
tudes (especially for higher-order kernels). For example,
the first and the fifth Hammerstein kernels of the device
TSA15h have been measured using an ESS with full ampli-
tude ( = 1) and half amplitude ( = 0.5). Their Fourier
J. Audio Eng. Sco., Vol. 1, No. 1, 2015 October
Hammerstein Kernels Identification by Means of a Sine Sweep Technique
1
3
6
10
Fig. 23. Comparison between the first Hammerstein kernel com-
ing from a power series up to the order 6 and its reconstruction
by the HKISS method.
calculate the Hammerstein kernels Hm ( j) from Hm1 ( j)
by changing the amplitude factor A appearing in Eq. (16).
The Fig. 25 shows an example of an adapted Hm1 ( j) ob-
tained by changing the A factor, in order to fit the Hm0.5 ( j)
kernels. With these adapted Hammerstein kernels based
nal at half amplitude (see Fig. 26) with less than 4.7% of
mean error. The parameter A has been manually adjusted
for our example. To avoid having a large set of impulse re-
sponses Hm ( j) corresponding to each level in the simu-
lator (space memory) and to avoid some difficulties with
a variable convolution kernels, the Hammerstein model
equation can be weighted as follows:
(33)
M
1
y[n] = h1m [n] ~ xm [n] (34)
m=1 [A()]m1
A way to determine the function A() is to find for each
level the A factor that minimize the areas between the
Fourier Transform of the Hammerstein kernels measured
at the level and the A corrected Hammerstein kernels
measured at the level = 1. Unfortunately, some devices
could have their Hammerstein kernels Hm too different
from each other to use this method (in general strong non-
linear devices). The Hammerstein kernels can then be ap-
proximated for a continuous range of levels by means of an
interpolation procedure as proposed by Tronchin and Coli
in [35] where each Hammerstein kernel hm [n] needs its own
Am () function.
5 A MATLAB TOOLBOX
The practical implementation of Hammerstein identifi-
cation technique is complex and may be subject to a lot of
possible errors. Hence, we have proposed a Matlab Tool-
box (see Fig. 27) for the Hammerstein identification base
on the methods presented in the previous sections. The
toolbox is composed of three parts: the ESS generation,
where the signal to send to the nonlinear device under test
11
Th.Schmitz AND J.J.Embrechts PAPERS

20 DUT and the emulation of the nonlinear DUT. The toolbox

h1
Full can be downloaded in [43] and more information about it
15 h5Full
could be found in [44].
h10.5
10 h50.5
Amplitude db

10

15 2 3 4
10 10 10
Frequency in Hz

Fig. 24. FFT of the first and fifth Hammerstein kernels of the
distortion TSA15h measured with an ESS with full and half am-
plitudes

20
h1Full

15 h5Full
h10.5
10 h50.5
Amplitude db

Fig. 27. The Hammerstein Identification Matlab Toolbox

5

0
5.1 Power series example
5
A short example using the toolbox to compute the Ham-
10 merstein kernels of a power series up to the order six is
given here.
15 2 3 4
10 10 10
Frequency in Hz 5.1.1 Sine sweep generation
The tab Sweep generation is the first step to measure a
Fig. 25. FFT of the first and fifth Hammerstein kernels of the
nonlinear device. It allows the generation of the ESS where
distortion TSA15h measured with an ESS with full amplitude and
adapted to match the kernels of half amplitude changing the A the parameters are : f1 the initial frequency of the ESS, f2
factor. the final frequency of the ESS, N the length of the ESS in
samples, fade in (or out) is the percentage of the value of
1 f1 (or f2 ) until a fade in (or out) is allowed. For example if
y1
y2
f1 = 10Hz and fade in = 0.1, it means that a fade in will be
apply between the frequency f1 = 10Hz and ten percents
0.5 more so 10+10*0.1 = 11Hz. For the fade out it will be apply
between the frequency f2 and f2 minus ten percents. These
Amplitude

values are shown in the frame Bandwidth. The generate

0 button exports the ESS in a .wav format.

5.1.2 Hammerstein kernels calculation

0.5
Once the ESS signal has been sent to the DUT, the y
signal at the output of the DUT has to be convolved with
1
the inverse filter of the ESS. In this example, the output
0 50 100 150 200
Sample [n] signal of a power series until the 6th order named yPow-
erSeries is provided. It can be loaded in the toolbox by
Fig. 26. Comparison between an ESS of amplitude 0.5 passing pushing the Load y response button. After having pressed
through the distortion effect TSA15h and it simulation (y2 ) by the Deconvolve button, the z impulse is obtained as has de-
means of Hammerstein kernels measured for an ESS of amplitude picted at Fig.28. Before cutting z into gm kernels, we must
1 but corrected with the A factor (see text).
specify the following parameters : Nb Kernels is the re-
quired number of Hammerstein kernels for the emulation
(DUT) is generated, the calculation of the Hammerstein and the Length Kernels is the desired size of the impulse
kernels through the signal recovered at the output of the responses. To compute the Hammerstein kernels, we also

12 J. Audio Eng. Sco., Vol. 1, No. 1, 2015 October

PAPERS
have to specify the Decaycut which is the number of sam-
ples kept before the theoretical beginning of the Hammer-
stein kernels. This number has to be comprised in the in-
terval [-1000 0].
5.1.3 Emulation of the DUT
The third tab allows the emulation of the DUT by con-
volving the Hammerstein kernels with any chosen sig-
nal in a .wav format. For this example, we can load the
sweep.wav signal and convolve it with the Hammerstein
kernels to emulate the power series giving back a signal
similar to yPowerSeries.
Fig. 28. The Hammerstein kernels identification tabs
6 CONCLUSION
Many researches have been done on Volterra kernels
identification in the last few years. As the impulse re-
sponse characterizes a linear system very precisely, we
expect to extend the method to nonlinear systems with
an equivalent success. As we have shown, the cascade
of Hammerstein models can lead to a great sensitivity to
implementation inaccuracies, especially when the order
M increases. In this paper, we have highlighted different
kinds of errors and we have shown how they can affect
the emulation of a nonlinear system. Moreover we have
proposed a way to get around each problem leading us
to an accurate emulation of three nonlinear systems: the
Tubes-Screamer like overdrive effect, the TSA15h tube am-
plifier and the Engl Retro Tube 50 E762 amplifier. We have
also described the limitations of the model (i.e. maximum
order, low frequency limitation, input level dependency).
The method seems well suited for real time emulation of
audio nonlinear systems. Finally, we have proposed a so-
lution in Eq (34) for the input amplitude level dependency
which seems to be valid for some nonlinear devices. How-
ever, for strongly nonlinear systems, further research is still
needed to solve this practical problem.
J. Audio Eng. Sco., Vol. 1, No. 1, 2015 October
Hammerstein Kernels Identification by Means of a Sine Sweep Technique
7 REFERENCES
[1] T. Schmitz, J.J. Embrechts, Nonlinear Guitar Loud-
speaker Simulation, presented at the 134th Convention of
the Audio Engineering Society (2013 May),convention e-
Brief 96.
[2] W. Klippel, Loudspeaker NonlinearitiesCauses, Pa-
rameters, Symptoms, presented at the 119th Convention of
the Audio Engineering Society (2005 October).
[3] I. Cohen, T. Helie, Simulation of a Guitar Ampli-
fier Stage for Several Triode Models: Examination of Some
Relevant Phenomena and Choice of Adapted Numerical
Schemes, presented at the 127th Convention of the Audio
Engineering Society (2009 October)
[4] M.J. Kemp, Analysis and Simulation of Non-
Linear Audio Processes Using Finite Impulse Responses
Derived at Multiple Impulse Amplitudes, presented at the
106th Convention of the Audio Engineering Society (1999
May)
[5] A.V. Oppenheim, A.S. Willsky, Signals and Systems
(Prentice Hall, 1996).
[6] M.Schetzen, The Volterra & Wiener Theory of Non-
linear Systems (John Wiley & Sons, 1980).
[7] J.S. Bendat, Nonlinear System Techniques and Ap-
plications, New York: Wiley, 1998.
[8] M. Schoukens, K. Tiels, M. Ishteva, & J. Schoukens,
Identification of Parallel Wiener-Hammerstein Systems
with a Decoupled Static Nonlinearity, presented at the
19th World Congress of the International Federation of
Automatic Control, Cape Town, South Africa, pp. 24
29, (2014 August). http://dx.doi.org/10.3182/
20140824-6-ZA-1003.00496.
[9] M. Schoukens, G. Vandersteen, & Y. Rolain, An
Identification Algorithm for Parallel Wiener-Hammerstein
Systems, presented at the 52th Annual Conference on
Decision and Control (CDC), Firenze,(2013 Decem-
ber). http://dx.doi.org/10.1109/CDC.2013.
6760659.
[10] F. Thouverez, & L. Jezequel, Identification of
NARMAX Models on a Modal Base, Journal of Sound
and Vibration, Vol 189, No. 2, pp. 193213, (1996).
http://dx.doi.org/10.1006/jsvi.1996.
0015
[11] O. Nelles, Nonlinear System Identification: From
Classical Approaches to Neural Networks and Fuzzy Mod-
els. Berlin, Germany: Springer-Verlag, (2001). http://
dx.doi.org/10.1007/978-3-662-04323-3.
[12] Y.W. Chen, S. Narieda, & K. YamashitaBlind
Nonlinear System Identification Based on a Constrained
Hybrid Genetic Algorithm, IEEE Transaction on Instru-
mentation and Measurement, Vol 52, No. 3, pp. 898902,
(2003).
[13] H. Sorenson, Kalman Filtering: Theory and Appli-
cation, Montvale, NJ: IEEE Press, (1985).
[14] O. Cappe, S.J. Godsill, & E. Moulines, An
Overview of Existing Methods and Recent Advances
in Sequential Monte Carlo, Proceeding of IEEE, Vol 95,
No. 5, pp 899924, (2007).
13
Th.Schmitz AND J.J.Embrechts
[15] T.Ogunfunmi, Adaptive Nonlinear System Iden-
tification: The Volterra and Wiener Model Approaches
Springer, (2007). http://dx.doi.org/10.1007/
978-0-387-68630-1.
[16] A. Novak,L. Simon, P. Lotton, & F. Kadlec, A
New Method for Identification of Nonlinear Systems Us-
ing Miso Model with Swept-Sine Technique: Application
to Loudspeaker Analysis, presented at the 124th Conven-
tion of the Audio Engineering Society (2008 May), conven-
tion paper 7441.
[17] M. Rebillat, R. Hennequin, E. Corteel, & B. Katz,
Identification of Cascade of Hammerstein Models for
the Description of Nonlinearities in Vibrating Devices,
Journal of Sound and Vibration, Vol. 330, Issue 5, pp.
10181038 (2011 February). http://dx.doi.org/
10.1016/j.jsv.2010.09.012.
[18] A. Farina, A. Bellini, & E. Armelloni, Nonlinear
Convolution: A New Approach for the Auralization of Dis-
torting Systems, presented at the 110th Convention of the
Audio Engineering Society (2001 May), convention paper
5359.
[19] A. Farina, Simultaneous Measurement of Impulse
Response and Distortion with a Swept-Sine Technique,
presented at the 108th Convention of the Audio Engineer-
ing Society (2000 February), convention paper 5093.
[20] T. Schmitz, J.J. Embrechts, Improvement in Non-
linear Guitar Loudspeaker Sound Reproduction, presented
at the 21th International Conference on Systems, Signals
and Image Processing (2014 May), international confer-
ence on IEEE.
[21] L. Tronchin, The Emulation of Nonlinear Time-
Invariant Audio Systems with Memory by Means of
Volterra Series, J. Audio Eng. Soc., Vol. 60, No. 12,
(2012 December)
[22] A. Novak, L. Simon, & P. Lotton, Analysis,
Synthesis, and Classification of Nonlinear Systems Us-
ing Synchronized Swept-Sine Method for Audio Effects,
Eurasip Journal on Advances in Signal Processing, Vol.
2010:793816, (2010 July). http://dx.doi.org/10.
1155/2010/793816.
[23] A. Novak, L. Simon, Nonlinear System Iden-
tification Using Exponential Swept-Sine Signal, pre-
sented at the IEEE Transactions on Instrumentation and
Measurement, Vol. 59, No 8, pp. 22202229, (2009
August). http://dx.doi.org/10.1109/TIM.
2009.2031836.
[24] A. Novak, P. Lotton, & L. Simon, Synchronized
Swept-Sine: Theory, Application and Implementation, J.
Audio Eng. Soc., Vol. 63, No. 10, (2015 October). http:
//dx.doi.org/10.17743/jaes.2015.0071
[25] J. Vanderkooy, & S. P. I. Thomson, Harmonic Dis-
tortion Measurement for Nonlinear System Identification,
presented at the 140th Convention of Audio Engineering
Society (2016 June), convention paper 9497.
[26] G. Palm, On Representation and Approximation
of Nonlinear Systems, Journal Biological Cybernetics,
Vol. 34, Issue 1, pp. 4952 (1979 September). http:
//dx.doi.org/10.1007/BF00336857.
14
PAPERS
[27] M. Rebillat, R. Hennequin, E. Corteel, B. Katz,
Identification of Cascade of Hammerstein Models for the
Description of Nonlinearities in Vibrating Devices, Jour-
nal of Sound and Vibration, Vol. 330, no. 5 (2011). http:
//dx.doi.org/10.1016/j.jsv.2010.09.012.
[28] M. Holters, T. Corbach, & U. Zlzer, Impulse Re-
sponse Measurement Techniques and their Applicability in
the Real World, presented at the 12th Int. Conference on
Digital Audio Effects (2009 September).
[29] Q. Meng, D. Sen, S. Wang, & L. Hayes, Im-
pulse Response Measurement with Sine Sweeps and Am-
plitude Modulation Schemes, 2nd International Confer-
ence on Signal Processing and Communication Systems,
(2008 December). http://dx.doi.org/10.1109/
ICSPCS.2008.4813749.
[30] S. Norcross, G. Soulodre, & M. Lavoie, Evalu-
ation of Inverse Filtering Techniques for Room/Speaker
Equalization, presented at the 113th Convention of the Au-
dio Engineering Society (2002 October),convention paper
5662.
[31] O. Kirkeby, & P. Nelson, Digital Filter Design for
Inversion Problems in Sound Reproduction, J. Audio Eng.
Soc., Vol. 47, No. 7/8, (1999 July/August).
[32] D. Zwillinger, Standard Mathematical Tables and
Formulae, 32nd ed. CRC Press, (2011).
[33] S. QIN, Y. QIN, Y. MAO, Accurate Instantaneous
Frequency Estimation with Iterated Hilbert Transform and
Its Application, presented at the Proceedings of the 7th
WSEAS International Conference on SIGNAL PROCESS-
ING, ROBOTICS and AUTOMATION (ISPRA08),(2008
February)
[34] M.R. Spiegel, S. Lipschutz, & J. Liu, Mathemati-
cal Handbook of Formulas and Tables, 3rd ed. Schaums
Outlines Series, McGraw-Hill eBook, (2009).
[35] L. Tronchin, V.L. Coli, Further Investigations in
the Emulation of Nonlinear Systems with Volterra Series,
J. Audio Eng. Soc., Vol. 63, No. 9, (2015 September).
[36] Tube Screamer Amplifier TSA15H:
http://www.ibanez.com/products/amp page15.php?year
=2015&area id=3&cat id=5&series id=141&ampl id=4
[37] D. Ciric, M. Markovic, M Mijic, & S. Sumarac-
Pavlovic, On the Effects of Nonlinearities in Room Im-
pulse Response Measurements with Exponential Sweeps,
Journal Applied Acoustics, Vol. 74, pp. 375382, (2013).
[38] A. Farina, Advancements in Impulse Response
Measurements by Sine Sweeps, presented at the 122th
Convention of Audio Engineering Society (2007 May),
convention paper 7121.
[39] A. Torras-Rossel, F. Jacobsen, A New Interpre-
tation of Distortion Artifacts in Sweep Measurements, J.
Audio Eng. Soc., Vol. 59, No. 5, (2011 May).
[40] Scip Solver: http://scip.zib.de/
[41] J. Macak, & J. Schimmel, Real-Time Guitar Tube
Amplifier Simulation Using an Approximation of Differen-
tial Equations, presented at the proceeding of the 13th Int.
Conference on Digital Audio Effects, Graz, Austria, (2010
September), DAFx-10.
J. Audio Eng. Sco., Vol. 1, No. 1, 2015 October
PAPERS
[42] Engl Retro Tube E762 Amplifier:
http://www.englamps.de/index.php?id=36&tx ddfproducts
pi1[uid]=119&iPhoneBypass=1
[43] Matlab Toolbox for Identification of the Hammer-
stein Kernels: https://github.com/TSchmitzULG/HKISS
THE AUTHORS
Thomas Schmitz
Thomas Schmitz received the degree in Electrical Engi-
neering (2012) from from the University of Liege (ULg).
His final project focused on the emulation of a electrody-
namics loudspeaker including its nonlinear behavior. He
is presently a Ph.D. student in Laboratory for Signal and
Image Exploitation (INTELSIG) research unit of the Elec-
trical Engineering and Computer Science (EECS) depart-
ment , University of Liege, Belgium. His research interests
are on signal processing, nonlinear modeling and real time
emulation of guitar audio systems.
r
J. Audio Eng. Sco., Vol. 1, No. 1, 2015 October
Hammerstein Kernels Identification by Means of a Sine Sweep Technique
[44] T. Schmitz, & J.-J. Embrechts, A new toolbox for
the identification of diagonal Volterra kernels allowing the
emulation of nonlinear audio devices, presented at the
proceeding of the 22nd International Congress on Acous-
tics, Buenos-Aires, Argentina, (2016 September), ICA.
J-J. Embrechts
Pr. J-J. Embrechts received the degree in Electrical En-
gineering (1981) and the Ph.D. degree (1987) from the
University of Liege (ULg). Since 1999, he is a professor
at the University of Liege, in the Department of Electri-
cal Engineering and Computer Science, teaching acoustics,
electroacoustics, audio and video engineering and lighting
techniques. He is the author and co-author of more than
120 journal and conference papers in the field of audio sig-
nal processing, room acoustics, auralization and lighting
techniques.
15