You are on page 1of 10

Sound field extrapolation: Inverse problems,

virtual microphone arrays and spatial filters


Chambatte1, and Alain Berry1
Philippe-Aubert Gauthier1, Cedric Camier1 , Yann Pasco1 , Eric
1 GAUS,

Universite de Sherbrooke, Sherbrooke, J1K 2R1, Canada and CIRMMT, McGill University, Montreal, H3A 1E3, Canada

Correspondence should be addressed to Philippe-Aubert Gauthier (philippe aubert gauthier@hotmail.com)


ABSTRACT
Sound field extrapolation is useful for measurement, description and characterization of sound environments
and sound fields that must be reproduced using spatial sound system such as Wave Field Synthesis, Ambisonics, etc. In this paper, two methods are compared: inverse problems and virtual microphone arrays
with filtering in the cylindrical harmonics domain. The goal was to define and identify methods that could
accommodate to various non-uniform sensor arrays (i.e. non array-specific methods) and that are less sensitive to measurement noise. According to the results presented in this paper, it seems that the method based
on inverse problem with Tikhonov regularization is less sensitive to measurement noise.

1.

INTRODUCTION

In recent years, a strong and wider interest for acoustic


transducer arrays and multichannel signal processing is
observed. Applications range from: acoustic imaging,
sound source localization and separation using nearfield
acoustical holography [1], beamforming [2], subspace
methods [2], time-reversal and other algorithms [2] to:
spatial sound reproduction using Wave Field Synthesis
(WFS) [3, 4], Ambisonics [5], multichannel Surround
sound [6], etc. These applications motivate a plethora of
research works both for enhanced spatial sound recording and spatial sound reproduction. This paper deals with
spatial sound recording for subsequent spatial sound reproduction. More precisely, it belongs to a larger project
of sound environment reproduction in aircraft cabins and
cockpits.
1.1.

Microphone array and acoustic imaging

Most of the microphone arrays applications developed


for acoustic imaging aim at the experimental visualization and characterization of noise sources. Redesign or
control can then be undertaken to reduce the noise radiated by the given object or machine. For those applications, a microphone array is used to measure (on a
discrete surface) the sound field radiated by the source.
Post-processing then involves: a) sound field extrapolation outside the measurement grid up to the sound source
[1] or b) direction-dependent mapping of the incoming

sound level on the array [2]. This type of applications is


common for noise abatement purposes.
Applications such as sound source localization, identification and separation are based on algorithms which can
either state: a) angular direction from which the target
sound comes from (parameter estimation [2]) or b) output a signal which is caused by a single target source in
a noisy or reverberating environment (source separation
or waveform estimation [2]). This second type of applications is mostly sought after for human-machine interfaces and communication purposes. Large microphone
array technologies and research are also gaining attention for audio applications [7].
1.2.

Multichannel sound reproduction

On the loudspeaker counterpart of array processing, part


of the applications are related to sound field reproduction. In this paper, we are specifically concerned by microphone array measurements for spatial sound field reproduction. The sound field reproduction applications
can be further differentiated in terms of their targets. The
most straightforward being a sound pressure field: the
target is then sound pressure as function of spatial coordinates. Other methods may involve spatial targets such
as sound intensity, sound diffuseness [8], sound contrast
maximization, sound power minimization [9], random
pressure fields on planar surfaces [10] or even psychoacoustics metrics. For all these sound field reproduction
methods, the target is either measured in situ or synthe-

AES 40TH INTERNATIONAL CONFERENCE, Tokyo, Japan, October 810


1

Gauthier et al.

Sound field extrapolation

2.

sized from a theoretical or virtual definition.


1.3.

INVERSE PROBLEM

Sound field measurement and description

For spatial sound and sound field reproduction using


loudspeaker arrays, there is therefore a need for on-site
sound field measurement, characterization and description. These spatial measurements and characterizations
of the sound environment are subsequently used as targets at the reproduction stage. In this paper, we search
for a sound field extrapolation method that allows later
post-processing and sound field characterization for subsequent sound field reproduction using loudspeaker or
vibration source arrays in closed space (i.e. listening
rooms or mock-ups). The main objective of the presented
methods is to achieve the largest extrapolation region so
that any relevant spatial parameters (sound pressure field,
spatial coherence, intensity maps, direct and diffuse energy densities as function of spatial coordinates, etc.) can
be computed and characterized over a large region from
a single measurement array. Another requirement was
the resulting extrapolation description should be general
enough: plane wave spectrum, helical wave spectrum,
etc [1]. Furthermore, we were interested by the definition of a method that could accommodate various microphone array configurations, regular or not, and that
can address measurement noise issues. Two methods are
compared: a) regularized inverse problems and b) cylindrical harmonics filtering at virtual sensor array. The second method is a new method (different from the most
common inverse problem regularization methods [11])
which is compared to the first to evaluate its potential.
For illustration purpose, numerical results are shown for
a specific and simple array. However, the reader should
note that the method can be applied to any array geometry.
1.4. Paper outline
After the introductory Sec. 1, Sec. 2 presents the classical
inverse problem formulation adapted to sound field measurements and extrapolation from a region centered on
the recording area. Typical sound field spectra and cylindrical harmonics spectra are analyzed in Sec. 3 where
the nature of the second method is evoked. In Sec. 4, it is
shown how the inverse problem formulation can function
as a spatial resampling algorithm on the basis of virtual
microphone arrays. In Sec. 4, classical spatial transform
methods are finally applied to virtual sensor arrays. Section 5 presents the comparison between the two methods. Concluding remarks, future research avenues and
upcoming practical developments are presented in Sec. 6.

In this section, the inverse problem method is defined


and illustrated for a microphone array. With this approach, one is looking for an optimal source distribution
that best fits the measured sound field at the microphone
array [11]. This optimal source distribution is then used
to extrapolate the measured sound field outside the area
of the array. The inverse problem, which differs from extrapolation methods based on spatial transform (such as
for classical Nearfield Acoustic Holography (NAH) [1],
or in-plane sound field extrapolation [7]), is especially
well-suited to irregular microphone arrays or particular
configurations. Although inverse problems are used for
non-conformal nearfield holography, these applications
involve sound extrapolation in a direction perpendicular
to the array. In our case, we are concerned by in-plane
extrapolation.
In this paper, an URA (Uniform Rectangular Array) configuration was selected for illustration purposes since it is
the configuration which is the easiest to build. The selection of this configuration was also influenced by the fact
that a similar array is currently built in our laboratory for
practical applications. Moreover, as the URA covers an
entire spatial region, such type of two-dimensional array
might more easily cope with a real 3D sound field or with
the proximity of a sound source or a diffracting object in
a confined space such as a vehicle cabin.
2.1.

Inverse problem formulation

A typical geometrical model of URA is depicted in


Fig. 1. The symbol and geometrical convention holds for
any microphone array, either 2D or 3D. Any field point is
described by a position vector x = [x1 , x2 , x3 ] in rectangular coordinates. Using cylindrical coordinates, this point
is described by [ , r, x3 ]. A real microphone is described
(m)
by a position vector xm . The complete real microphone
(m) (m)
(m)
set is described by a vector set x1 , x2 , , xM for an
array made of M microphones. Using this URA, a complex sound pressure field is measured for an angular frequency [rad/s] and stored in a M-component vector
(m)
p( , xm ) [Pa]. In case of real measurements, the complex pressure field is obtained by a Fourier transform of
the measured microphone signals, assuming stationary
sound fields or short-duration frames. The smallest distance between two microphones is denoted x(m) . The

AES 40TH INTERNATIONAL CONFERENCE, Tokyo, Japan, October 810


Page 2 of 10

Gauthier et al.

Sound field extrapolation

x3

Impinging wave

unit vector pointing in the l-th propagation direction).


j t , with time t [sec],
The selected time convention
is e
and imaginary number j = 1.

x
x2
x3

(m)

, xm ) matches
The inverse problem is posed so that p(
(m)
p( , xm ). From Eq. (1)

x1
(m)

(m)

(l)

(l)

p( , xm ) = ZS( , l , l )

xm

(3)

and the solution is


(l)

Fig. 1: A microphone array. Microphones are marked by


(m)
small filled circles. x is any field point location, xm is
the m-th microphone location. The region covered by the
URA is shown by the rectangle.
following principles hold up to the spatial aliasing frequency falias = c/2x(m) [Hz], where c = 343 m/s is the
sound velocity.

(m)

(l)

(m)

(l)

(l)

(l)

, xm ) = Z( , xm , l , l )S( , l , l )
p(

(1)

where L-component vector S is the plane waves complex


(l)
(l) (l)
(l)
amplitudes, l (from the set 1 , 2 , , L ) the l-th
(l)

plane wave azimuth and l the l-th plane wave elevation (zero for the 2D case). The matrix Z is the transfer
matrix (dimension: M L) from plane wave amplitudes
to resulting pressures at the microphones. The transfer
matrix is defined by its ml-th element
(m)

(l)

(l)

(l)

Zml ( , xm , l , l ) = e jkl
(l)
kl
(l)
(l)
k sin(l ) cos(l )e2

where

(l)

(m)

xm

(2)
(l)

k cos(l ) cos(l )e1

+ k sin(l )e3 , ei being a unit


(l)

vector along xi . k [rad/m] is the wave number and kl is


(l)

(l)

(l)

the wave-number vector (kl = knl , with nl being the

(m)

(4)

where superscript + denotes pseudo-inversion [12].


Pseudo-inversion is used since the matrix Z is generally
(l) (l)
not a square matrix. The solution S( , l , l ) then
gives the plane waves amplitudes which will reproduce
the measured sound field at the array. The sound field
can then be extrapolated for any point x outside or inside
the region covered by the URA using

(m)

To extrapolate the measured sound field p( , xm ) to any


field point x outside or inside the URA region, an inverse
problem is posed. Lets first assume a set of L plane
waves: uniformly covering a 4 steradians solid angle
range for the 3D case and uniformly covering a 2 radians angle range for the 2D (in-plane) case. Note that
selection of the plane waves propagation directions can
be different or adapted to the problem at hand, depending
on a priori information. The pressure produced by this
set of L plane waves at the sensor array is given by

(l)

S( , l , l ) = Z+ p( , xm )

p(
, x) =

Sl ( , l

(l)

(l)

(l)

, l )e jkl

(5)

l=1

with Sl obtained from Eq. (4). This is the basis of sound


field extrapolation using inverse problem. Stated simply,
one should understand that the previous equations and inverse problem aim at computing a set of L virtual planar
sources that optimally (in the least-mean-square sense)
an approximation of the measured sound field
recreate p,
p. A great advantage of inverse problem over spatial
transform method [7] is that the inverse problem approach can be readily applied to any non-uniform or uniform microphone array. Indeed, no particular assump(m)
tions are made about xm to obtain Eqs. (4) and (5).
Moreover, although classical NAH applications using
uniform rectangular array take great advantage of 2D
spatial transform, it is solely for forward and backward
sound field extrapolation along the array perpendicular
axis (x3 in Fig. 1). It is important to note that, although
a set of plane waves was used in the inverse problem
formulation, the operation represented by Eq. (4) does
not correspond to a 2D or 3D Fourier transform of the
sound field as used in NAH [1] or sound field description as reported by Poletti [13]. In our case, we are
mostly interested by in-plane (x1 , x2 ) sound field extrapolation for subsequent in-plane sound field reproduction.
This is a simple situation which prevents the use of the
2D Fourier transform as this spatial transform assumes a

AES 40TH INTERNATIONAL CONFERENCE, Tokyo, Japan, October 810


Page 3 of 10

Gauthier et al.

Sound field extrapolation

Fig. 2: (a) and (b): Real and imaginary parts of the impinging sound field. Microphones are shown as small filled
circles. (c) and (d): Real and imaginary parts of the extrapolated sound field by inverse problem. The isocontour lines
of the resulting local quadratic extrapolation error are shown for 0.1% (white line) and 5% (black line).
repetitive spatial signal along the in-plane axes (x1 , x2 ).
This is not the case for a real sound field impinging
on an array. Finally, the selection of a plane wave set
as a solution basis for the inverse problem is not arbitrary. Indeed, other general orthogonal solutions of the
Helmholtz equation such as cylindrical or spherical harmonics are not bounded at the origin of the coordinates
system which corresponds to the array center [1, 7] . In
that case, the inverse problem based on these other orthogonal solutions of the wave equation is extremely illconditioned and p may even not approach p. This has
been verified by the authors.
An example of sound field extrapolation using the inverse problem approach is shown in Fig. 2 for an URA
of 64 microphones covering a square with a side length
of 0.77 m. The microphone separation distance is 11 cm.
Theoretically, this URA could then measure a sound field
without spatial aliasing up to 1559 Hz. The harmonic
sound field measured at 500 Hz is shown in Fig. 2(a)
and (b). The sound field impinging on the URA is a
plane wave with a propagation angle of 45 deg. The
corresponding extrapolated sound field obtained by the
inverse problem solution is shown in Fig. 2(c) and (d).
The inverse problem included a set of L = 128 plane
waves covering 2 radians. The inverse problem method
clearly performs sound field extrapolation in the measurement plane. The effective extrapolation region covers (defined by a local quadratic error less than 0.1%)
a circular region of roughly 2 m radius which is larger
than the URA. It should be kept in mind that the size of
effective extrapolation region varies with frequency.

2.2.

Inverse problem condition number

Although the example shown in Fig. 2 seems promising


for sound field extrapolation, the condition number of
Z is very large: 3.5 1016 The condition number is the
ratio of the largest and smallest (greater than zero) singular values: = max /min where the singular values are
obtained by singular value decomposition [12].
As long as measurement noise is absent, the pseudoinverse solution (Eq. (4)) to the inverse problem gives a
consistent result. However, any noise or input variability
has the potential to make the solution S excessively large.
Indeed, any variation p on the measured sound field (including noise) will imply a variation S on the solution
S that is bounded by the condition number [11, 14]
||S||/||S|| (Z)||p||/||p||

(6)

where ||...|| denotes the Euclidean norm. This highlights the amplification potential of the condition number. However, even if S becomes very large, the corresponding extrapolated sound field may be able to approach the measured sound field at the sensor array
(m)
(m)
, xm ) p( , xm )). However this solution might
(p(
not provide a convenient sound field extrapolation outside the array region. An example is shown in Fig. 3 for
a case similar to the one shown above with an additional
0.1% of random and spatially incoherent noise. Fig. 3 illustrates that the extrapolated sound field is only valid in
the array region. At first glance, this suggests that inverse
problem method for sound field extrapolation might not

AES 40TH INTERNATIONAL CONFERENCE, Tokyo, Japan, October 810


Page 4 of 10

Gauthier et al.

Sound field extrapolation

Fig. 3: Real and imaginary parts of the extrapolated


sound field by inverse problem with 0.1% of noise (60
dB SNR). Sound pressure is clipped at values larger than
100.
be appropriate, except, possibly, as a spatial resampling
algorithm for the region covered by the array.
Note that for the two cases reported in Secs. 2.1 and 2.2,
the numerical pseudo-inversion only involves the discarding of singular values smaller than 9 1012. This
type of regularization (truncated singular value decomposition) [15] simply prevents instability caused by numerical machine limited precision.
2.3.

Inverse problem regularization

To circumvent this noise sensitivity problem, regularization is introduced. We first recast the problem as an error
minimization task using a quadratic cost function
J = eH e + SH S

(7)

where e is the error column vector defined by e = p p .


In Eq. (7), is the regularization parameter. Combining this error vector definition with Eq. (3), it is straightforward to show that the cost function is a Hermitian
quadratic function of S. The optimal S(opt) that will minimize this cost function is [16, 17]
(opt)

SL1 =

ZH
ZH Z + I

pM1

(8)

LM

where I is the L L identity matrix. The purpose of the


regularization parameter is simple: it populates the main
diagonal of the matrix to be inverted. This filters any too
small singular values which have the potential to amplify
noise. This is Tikhonov regularization [15].
The choice of the regularization parameter is not trivial.
Methods such as ordinary cross-validation and generalized cross-validation are defined for this unique purpose

in acoustics [18]. Methods for choosing the regularization parameter are, by themselves, a research topic. On
the one hands, while the selection of a small penalization parameter may stabilize the inversion (p may approach p very closely in the microphone array region),
it is again not guaranteed that optimal solution will give
any significant results outside the sensor array. It might
again diverge at large distance from the array. On the
other hands, selecting a large penalization parameter may
strongly stabilize the inversion but also reduce the spatial
resolution of the extrapolated sound field (including the
array region). Examples of sound field extrapolation using inverse problem method with varying regularization
parameter are shown in Fig. 8(a) to (f). Sec. 5 explains
these results.
One way to select the optimal regularization parameter
is based on the evaluation of the L-curve [15] which is
a representation of the quadratic error eH e as function
of the solution quadratic amplitude S(opt)H S(opt) for a
wide range of . However, this selection implies a serious amount of computation time as the problem must
be solved for an entire distribution of . It would be interesting to benefit from a more physical approach that
would simplify the regularization process. In the following section, the typical cylindrical harmonics spectra of
extrapolated sound field with and without noise will provide some hints on how could that be done. Moreover,
this original method will be illustrated.
3.

CYLINDRICAL HARMONICS SPECTRA

To analyze typical sound field spectra, cylindrical coordinates [ , r, x3 ] are used. The definition of cylindrical
harmonics (orthogonal functions of the Helmholtz equation expressed in cylindrical coordinates) are recalled
for the 2D case. The sound pressure field of diverging
(1)
(2)
(Pk ) and converging (Pk ) cylindrical harmonics (CH)
are given by multiplication of Hankel and complex exponential functions
Pk (r, , ) = Hk (kr)e jk ,

(9)

Pk (r, , ) = Hk (kr)e jk .

(10)

(1)

(2)

(1)

(2)

Any two-dimensional sound field can be represented by

AES 40TH INTERNATIONAL CONFERENCE, Tokyo, Japan, October 810


Page 5 of 10

Gauthier et al.

Sound field extrapolation

(i)

function of order k , Hk is the spatial derivative of the


Hankel function and is the air density [kg/m3 ] (1.18
kg/m3 ). P(R, k , ) and Vn (R, k , ) are the sound pressure and particle normal velocity in the circumferential
wave index and frequency domain.
For a circular continuous aperture, the sound pressure and particle normal velocity spectra are defined as
Fourier series expansion coefficients

Fig. 4: Magnitude of the pressure and velocity angular


spectra for a circular sensor array centered at the origin
with radius R = 0.3 m on which impinges a broadband
plane wave with the propagation direction = 45 deg.
a linear combination of CH [1, 7]:
p(r, , ) =

k =

M (1) (k , )Hk (kr)e jk +

k =

M (2) (k , )Hk (kr)e jk . (11)


(2)

(2)
The terms
, ) and M (k , ) are the complex
CH coefficients, creating a CH spectrum.

The M (1) and M (2) are directly computed from sound


pressure and normal sound pressure gradient Fourier
transforms (denoted P(R, k , ) and Vn (R, k , ), respectively) over a continuous circular aperture of radius
R [7]

M (2) =

(2)
(2)
Hk (kR)P Hk (kR) j cVn
,
(1)
(2)
(2)
(1)
Hk (kR)Hk (kR) Hk (kR)Hk (kR)
(1)

1
2

Vn (R, k , ) =

1
2

2
0
2
0

p(R, , )e jk d

(14)

vn (R, , )e jk d

(15)

where p(R, , ) and vn (R, , ) are the complex-valued


sound pressure and particle normal velocity along a circular sensing aperture (radius R). The discrete versions of these two transforms simply involve replacing
the integration by a proper summation. For the method
presented in this paper, it is assumed that the circular
aperture is a virtual microphone array. The pressure
p(R, , ) and velocity vn (R, , ) at this virtual array
are computed from Eq. (5) once the inverse problem is
solved and used as a non-uniform resampling algorithm
(as evoked in Sec. 2).

(1)

M (1) (k

M (1) =

P(R, k , ) =

For practical cases, Eq. (11) should not be used for extrapolation since Hankel functions are not bounded at
r = 0 and the summation over k may be numerically
unstable [7]. To circumvent this problem, one must first
transform the CH coefficients in another set of diverging
and converging plane waves amplitudes [7]
(1)

(2)

(l)

Sl ( , l ) =
(l)

(l)

(1/2 ) jk M (1) (k , )e jk l ,

k =

Sl ( , l ) =

(16)
(1/2 ) jk M (2) (k , )e

(l)
jk l

k =

(12)

(17)
The extrapolated sound field is then computed without
numerical difficulty using

(1)

Hk (kR)P Hk (kR) j cVn

(2)
(1)
(1)
(2)
Hk (kR)Hk (kR) Hk (kR)Hk (kR)

(13)

where is the angular frequency [rad/s], k the wavenumber (k = /c) [rad/m], k the circumferential wave in(i)
dex (k = ... 2, 1, 0, 1, ...), Hk is the i-type Hankel

p( , x) =

(1)

(l)

Sl e jkl

(2)

(l)

+ Sl e jkl

(l)

(18)

l=1

where (l) is the angular separation [rad] of the plane


waves propagation directions.

AES 40TH INTERNATIONAL CONFERENCE, Tokyo, Japan, October 810


Page 6 of 10

Gauthier et al.

Sound field extrapolation

without noise. For the case without noise, all the spectra
components are relevant sources of information about the
original sound field. Indeed, even the part of the spectra which belongs to the evanescent-like region brings
information about the original sound field. Therefore,
that information should be conserved except if noise is
present. The spectra for the same impinging plane wave
with additional noise at the original microphone array
(URA shown in Fig. 2) used for sound field extrapolation to the virtual circular sensor array are also shown
in Fig. 5. (For that case, the spectra are obtained using Eqs. (14) and (15) where sound pressure and particle
normal velocity are computed using the inverse problem
extrapolated sound field evaluated at the virtual circular aperture.) Clearly, outside the passband region, the
signal-to-noise ratio is drastically reduced when noise is
present. Once the sound field spectra are used to compute the CH spectra using Eqs. (12) and (13), the noise
outside the passband region can be drastically amplified
by the Hankel functions included in both the numerator
and denominator of Eqs. (12) and (13).

Magnitude of pressure angular spectra P(k ) at 500 Hz

|P(k)|

10

10

10

10

No noise
0.1% noise
0.001% noise
20

20

10

10

20

Magnitude of velocity angular spectra Vn(k) at 500 Hz

10

10

20

20

k=kR

10

k=kR

|Vn(k)|

10

10
0
10
k = M/2...,2,1,0,1,...M/2

20

Fig. 5: Magnitude of sound pressure and velocity angular spectra at 500 Hz without noise and with 0.1% of
noise. Additional spectra with noise (103%) are also
shown as dotted lines. The arrows indicate the decrease
of these curves for a decreasing noise level.
3.1.

CH spectra: Numerical examples

The hypothesis which underlies the method described in


the remainder of this paper relies on a simple observation
of the typical sound field spectra and corresponding CH
spectra. It will be shown how a very specific region of
the sound field spectra is affected by the inverse problem
sensitivity to noise. This observation helps in the definition of the second method based on CH filtering.
The noise-free sound-field spectra are shown in Fig. 4 for
a virtual circular aperture of radius R = 0.3 m on which
impinges a broadband plane wave with the propagation
direction = 45 deg. The virtual sensor array is made of
100 microphones arranged in a double circle (inner and
outer) for the discrete evaluation of pressure and velocity. Note that most of the spectras significant energy is
contained in the |k | kR region, here called the passband region. This spectra region corresponds to propagative waves. For |k | > kR, the corresponding waves
are cylindrical evanescent-like waves [1].
A slice of these spectra are shown in Fig. 5, with and

Resulting CH spectra are shown in Fig. 6. One


can clearly see how noise affects the computed CH
spectra and consequently the extrapolated sound field.
Two important features of these CH spectra should be
noted: a) the passband region is not affected by noise
and b) the peaks in the CH spectra with noise all occur
at |k | = 14 (for the shown frequency) (thin dashed line),
i.e. the peaks positions do not depend on noise level. The
region between these peaks and |k | = kR will be defined
as the transition band. The peaks in the CH spectra are
caused by the amplifying Hankel functions in Eqs. (12)
and (13). An example of the terms of Eq. (12) for the reported case is shown in Fig. 7 where one can see the noise
amplification potential of |H (2) |, |H (2) | and Eq. (12)s
denominator. The Hankel functions H (2) and H (2) in
the numerator boost the angular spectra with increasing
k while the denominator becomes suddenly large for a
given k which seems to define the end of the transition
band. Any noise present in the transition band will be
amplified (see Fig. 6).
4.

VIRTUAL MICROPHONE ARRAYS

To benefit from the observations mentioned above, the


inverse problem will be used as a non-uniform spatial resampling method. This resampling allows the use of virtual microphone arrays of any size and of varying geom-

AES 40TH INTERNATIONAL CONFERENCE, Tokyo, Japan, October 810


Page 7 of 10

Gauthier et al.

Sound field extrapolation

Terms gains for the evaluation of M(1) at 500Hz

CH spectra at 500 Hz

10

(2)

|Hk |

|H(2)|

|M(1)| (noise)

|M(2)| (noise)

|M(1)|

500

|M(2)|

Fig. 6: CH spectra at 500 Hz for the case without


and with 0.1% of noise. Additional spectra with noise
(103%) are also shown as dotted lines. Arrows indicate
a decreasing noise level. Spectras peaks are circled
etry as function of frequency. In this paper, the inverseproblem extrapolated sound field described by Eqs. (4)
and (5) will be used to compute the sound field over a circular virtual array of microphones described by location
(v)
(v)
(v)
vectors xv (from a set x1 , , xV ). If it is assumed
that the virtual array is located inside the region covered
by the real array, the resampling by inverse problem can
be based on pseudo-inversion without Tikhonov regularization. Indeed, for this method based on virtual arrays,
we are not concerned by the inverse problem reproduced
sound field outside the URA. The virtual array is made
circular so that the CH transforms mentioned in Sec. 3
can then be applied.
Using a virtual array that fits the circular array as used
in Fig. 4, the CH coefficients are computed (Eqs. (12)
and (13)), then transformed in plane waves amplitudes
(Eqs. (16) and (17)) and finally converted to the extrapolated sound field (Eq. (18)). Note that for the case without noise, conversion from inverse problem plane waves
solution to the CH spectra using spatial resampling and
transform gives an extrapolated sound field very similar
to the one obtained by the inverse problem method. If
noise is present, the sound field extrapolation by CH is
less efficient and it is expected that a low-pass CH filter
might remove the noise peaks in the CH spectra. Accordingly, a CH filter is designed so that conversion from the
CH coefficients M (1) and M (2) involve a zeroing of the
CH coefficients which corresponds to a k above kR but

20

|H(2)/Den|

k=14

20

10
0
10
k=M/2...,2,1,0,1,...M/2

20

|H(2)/Den|
k=kR

250

20

k =kR

|1/Den|
k =14

10

k=kR

10

10

Gains [dB ref 1]

250

k=kR

|M(1)(k)| and |M(2)(k)|

500

10
0
10
k =M/2,...,1,0,1,...M/2

20

Fig. 7: Gains at 500 Hz of the various terms in Eq. (12).


The maximums, which cause the peaks if Fig. 6, are
marked by small circles.
within the transition region. The zeroing starts from a
cut-off k . Examples of sound field extrapolation using virtual sensor array and CH filtering are shown in
Fig. 8(g) to (l). These results are explained in Sec. 5.
5.

COMPARISON OF THE TWO METHODS

Figure 8 shows sound field extrapolation examples for


the two methods where both the regularization parameter and the cut-off k are varied within an range of interest. This range was selected to reach almost diverging
and spatially-limited sound field extrapolation. Although
the virtual array with CH filtering method seems to reduce the diverging part of the extrapolated sound field
(see the diverging case in Fig. 3), the size of the effective
extrapolation region diminishes greatly as the cut-off k
is reduced. Interestingly, the effect of the regularization
parameter in the regularized inverse problem is different.
Indeed, the effective extrapolation region is much less reduced as the regularization parameter is increased (and
as the diverging part of the extrapolated sound field is decreased). In fact, the size of the effective extrapolation
seems to be less affected by the regularization parameter than the CH cut-off k . This suggests that sound
field extrapolation on the basis of inverse problem with
Tikhonov regularization might be more efficient than the
second method based on virtual sensor array and CH filtering. Moreover, the inverse problem method involves
less processing than the virtual sensor method which in-

AES 40TH INTERNATIONAL CONFERENCE, Tokyo, Japan, October 810


Page 8 of 10

Gauthier et al.

Sound field extrapolation

Fig. 8: Comparisons of the extrapolated sound fields by the inverse regularized problem and CH filtering with virtual
arrays with 0.1% noise at the URA. (a) and (b): Inverse problem with = 1e 9. (c) and (d): Inverse problem with
= 1e 7. (e) and (f): Inverse problem with = 1e 5. (g) and (h): Virtual array and a cut-off at k = 9. (i) and (j):
Virtual array and a cut-off at k = 7. (k) and (l): Virtual array and a cut-off at k = 5.
volves two additional transforms. Finally, the inverse
problem method directly outputs a set of plane waves as
a sound field descriptor. This is also a great advantage,
beside its simplicity, of sound field extrapolation method
based on regularized inverse problem.
6.

CONCLUSION

This paper has investigated the properties of two sound


field extrapolation methods that may be applied to uniform rectangular array or non-uniform array. The first
method based on inverse problem theory is sensitive to
noise. The introduction of Tikhonov regularization in

the inverse problem can however circumvent this noise


sensitivity and provide efficient sound field extrapolation. The second method based on virtual sensor array
and CH filtering was inspired by the fact that measurement noise manifests itself in a very specific region of
the CH spectra. The performance of this second method
is however less interesting. Indeed, the reduction of the
diverging part of the extrapolated sound field by CH filtering also involved a strong reduction of the effective
region of the extrapolation. On the basis of these observations, it seems that inverse problem extrapolation is a
better method for in-plane sound field extrapolation.

AES 40TH INTERNATIONAL CONFERENCE, Tokyo, Japan, October 810


Page 9 of 10

Gauthier et al.

Sound field extrapolation

Future research could be done to evaluate the performance of the inverse problem method for the 3D cases,
i.e. 3D sound field measured using a 2D array or a 2D
array with variations in the vertical direction. A more
in-depth study of inverse problem properties for sound
field extrapolation should be done. Recent works by the
authors in that direction is devoted to a new promising
regularization method for the inverse problem, this will
be the topic of a future paper.
Once sound field extrapolation is achieved, it is possible
to move to the next step: characterization of the sound
field or sound environment for subsequent reproduction.
Beside the definition and uses of classical targets, many
other types of target could come from a multidisciplinary
description of the sound environment that must be spatially reproduced.
7.

ACKNOWLEDGMENT

This work is part of a project involving: Consortium for


Research and Innovation in Aerospace in Quebec, Bombardier Aeronautique, CAE and McGill University, supported by a Natural Sciences and Engineering Research
Council of Canada grant.
8.

convention, Amsterdam, The Netherlands, 2003


March 2225.
[6] F. Rumsey, Spatial audio, Focal Press, 2001.
[7] E. Hulsebos, D. de Vries and E. Bourdillat, Improved Microphone Array Configurations for Auralization of Sound Fields by Wave-Field Synthesis, Journal of the Audio Engineering Society 50
(2002), no. 10, 779790.
[8] J. Merimaa and V. Pulkki,Spatial impulse response rendering I: Analysis and synthesis, Journal of the Audio Engineering Society 53 (2005),
no. 12, 11151127.
[9] M. Jones and S.J. Elliott, Personal audio with multiple dark zones, Journal of the Acoustical Society
of America 124 (2008), no. 6, 34973506.
[10] T. Bravo and C. Maury, The experimental synthesis of random pressure fields: Methodology,
Journal of the Acoustical Society of America 120
(2006), no. 5, 2702-2711.
[11] P.A. Nelson, A review of some inverse problems in
acoustics, International Journal of Acoustics and
Vibration 6 (2001), no. 3, 118134.
[12] G.H. Golub and C.F. van Loan, Matrix Computations, John Hopkins University Press, 1996.

REFERENCES

[1] E.G. Williams, Fourier Acoustics Sound Radiation and Nearfield Acoustical Holography, Academic Press, 1999.
[2] H. Teutsch, Modal Array Signal Processing: Principles and Application of Wavefield Decomposition,
Springer, 2007.
[3] A.J. Berkhout, D. de vries and P. Vogel, Acoustic control by wave field synthesis, Journal of the
Acoustical Society of America 93 (1993), no. 5,
27642778.
[4] P.-A. Gauthier and A. Berry, Adaptive wave field
synthesis with independent radiation mode control for active sound field reproduction: Theory,
Journal of the Acoustical Society of America 119
(2006), no. 5, 27212737.
[5] J. Daniel, R. Nicol and S. Moreau, Further Investigations of High Order Ambisonics and Wavefield
Synthesis for Holophonic Sound Imaging, Convention paper 5788, presented at the AES 114th

[13] M. Poletti, A unified theory of horizontal holographic sound systems, Journal of the Audio Engineering Society 48 (2000), no. 12, 11551182.
[14] P.A. Nelson and S.H. Yoon, Estimation of acoustics source strength by inverse method: Part I, conditioning of the inverse problem, Journal of Sound
and Vibration 233 (2000), no. 4, 643668.
[15] P.C. Hansen, Rank-Deficient and Discrete Ill-Posed
Problems, SIAM, 1998.
[16] P.A. Nelson and S.J. Elliot, Active Control of
Sound, Academic Press, 1992.
[17] S. Elliott, Signal Processing for Active Control,
Academic Press, 2001.
[18] S.H. Yoon and P.A. Nelson, Estimation of acoustics source strength by inverse method: Part II, experimental investigation of methods for choosing
regularization parameters Journal of Sound and Vibration, 233 (2000), no. 4, 669705.

AES 40TH INTERNATIONAL CONFERENCE, Tokyo, Japan, October 810


Page 10 of 10

You might also like