2011 DERIVATION OF AMBISONICS SIGNALS AND PLANE WAVE DESCRIPTION OF MEASURED SOUND FIELD USING IRREGULAR MICROPHONE ARRAYS AND INVERSE PROBLEM THEORY

8/11/2019 2011 DERIVATION OF AMBISONICS SIGNALS AND PLANE WAVE DESCRIPTION OF MEASURED SOUND FIELD USING IR

1/17

AMBISONICSSYMPOSIUM2011June 2-3, Lexington, KY

DERIVATION OF AMBISONICS SIGNALS AND PLANE WAVE

DESCRIPTION OF MEASURED SOUND FIELD USING IRREGULARMICROPHONE ARRAYS AND INVERSE PROBLEM THEORY

P.-A. Gauthier1, . Chambatte1, C. Camier1, Y. Pasco1, A. Berry1

1 Groupe dAcoustique de lUniversit de Sherbrooke,

2500 boul. de lUniversit, Sherbrooke, Qubec, J1K 2R1 CANADA

[email protected]

Abstract: Spatial sound field reproduction involves two steps: spatial sound field capture and subsequent spatial sound

field reproduction. This paper deals with sound field capture within the context of aircraft cabin sound environment re-production while complying to practical and geometrical constraints. Higher-order Ambisonics (HOA) are common tools

for spatial sound capture. Spherical microphone arrays are often used for this purpose. However, in some practical

situations, one may favor irregular or non-spherical array geometry. This is the purpose of this paper. Sound field ex-

trapolation (SFE) is aimed at the prediction of a sound field in an extrapolation region using a microphone array in a

measurement region different from the extrapolation region. The reported SFE method is based on an inverse problem

formulation combined with a recently proposed regularization method: a beamforming matrix in the discrete smoothing

norm of the cost function. In post-processing stages, we are interested in the derivation of HOA signals from the inverse

problem solution. The HOA signals are derived from the plane wave amplitudes obtained by the resolution of the inverse

problem. This approach gives the B-Format Ambisonics signals at each point of the extrapolation region, thus providing

a virtually movable Ambisonic microphone. Experimental results which validate the proposed SFE method in an extrapo-

lation region are presented. The paper finally presents the predicted Ambisonics signals from the SFE method applied to

the experimental SFE.

Key words: Ambisonics, higher-order Ambisonics, microphone array, sound field extrapolation, beamforming, inverseproblem, virtual acoustics, virtual Ambisonics microphones

1 INTRODUCTION

Spatial sound field reproduction usually involves two steps:spatial sound field capture and subsequent spatial soundfield reproduction[1]. Metrics derived from the capturedsound field (sound pressure level, sound energy density,source localization, etc.) will depend on the experimen-tal technique employed in the sound field capture. Usually,

sound fields are measured in areas such as outdoor places,concert halls[2], or more confined spaces such as vehicleinteriors. This paper deals with microphone array process-ing within the context of aircraft cabin sound environmentreproduction [3,4].

In many situations, the spatial sound capture phase raisesthe specific problem of choosing the measurement tech-nique which will provide the more extended and accuratesound field measurement while complying to practical con-straints. Ambisonics-based systems and higher-order Am-bisonics (HOA) are common tools for this kind of problem[5]. Through a spherical harmonic decomposition of the

local sound pressure field, they have the ability to providesound pressure levels and directions of propagation with alow number of transmitted channels. The main drawback

is that the sound field information is local and extrapola-tion of the sound field to other locations is difficult, imply-ing the need to multiply the number of measurement posi-tions or to increase the Ambisonics order. Spherical micro-phone arrays are often used for this purpose. However, insome practical situations, one may favor irregular or non-spherical array geometry [1]. This paper considers irregulararray geometry.

For spatial sound field reproduction purpose, one is typi-cally interested by the use of microphone arrays in order toextrapolate the sound field outside the measurement region.In general, sound field extrapolation (SFE) is aimed at theprediction of a sound field in an extrapolation region usinga microphone array in a measurement region different fromthe extrapolation region [6]. SFE finds application in noisesource identification, sound source localization [7], soundsource reconstruction [6,8,9] and sound field measurementfor spatial audio[2,10]. In the context of sound field repro-duction in a vehicle cabin, we are interested in the deriva-tion of first-order (B-format) and HOA signals from an arbi-

trary and irregular microphonearray. The derivation of Am-bisonics signals from an irregular array geometry should bebased on a method that circumvents the fact that the spheri-


2/17

cal harmonics are not necessarily orthogonal functions overthe spatially-sampled region defined by an irregular or ar-bitrary microphone array geometry [1]. To circumvent this,we recently proposed a method based on sound field ex-trapolation that combines inverse problem theory and beam-forming[11].

This method is based on an inverse problem formulationcombined with a recently proposed regularization approach:a beamforming matrix in the discrete smoothing norm ofthe cost function. In our application, the inverse solutionis a set of plane waves that best approximates the soundfield captured by the array. The B-Format Ambisonics first-order signalsW,X,Y andZare finally derived from theplane wave complex amplitudes obtained by the resolutionof the discrete inverse problem. This approach gives the B-Format Ambisonics signals at each point of the extrapola-tion region, thus providing a virtually movable Ambisonicsmicrophone. Furthermore, the derivation of HOA signals is

possible.This paper presents the derivation of the HOA signals fromthe SFE results along with the first experimental validationof the proposed SFE method based on the beamformingreg-ularization matrix.

1.1. Paper structure

Section2recalls the recently developed SFE method basedon inverse problem theory and beamforming regularizationmatrix. The derivation of the HOA signals from the SFEsolution is presented in Sec.3.1. Numerical results to ver-ify the HOA signals derivation method are presented in

Sec.3.2. The experiments and corresponding results aredescribed in Sec.4. To facilitate the understanding of thepotential applications of the proposed approach, Sec.5 il-lustrates a possible implementation of the proposed method.A short conclusion closes the paper. Full-page figures arepresented at the end of the paper.

2 INVERSE PROBLEMS: THEORY

The generic microphone array and coordinate system areshown in Fig. 1. The array includes Mmicrophones. Fora given frequency, a complex sound pressure field measure-

ment is stored in p(xm) M

. In this paper, bold smallletters represent vectors and capital bold letters representmatrices.

2.1. Inverse problem: Tikhonov regularization

The general discrete direct sound radiation problem in ma-trix form is:

p(xm) = G(xm,yl)q(yl), (1)

withp M,G ML andq L, (2)

where qis the monopolar source strength [12]vector,Gis

the transfer matrix that represents sound radiation and p isthe resulting sound pressure vector at the microphone loca-tions xm. The l-th source is located in yl. In this paper,

Figure 1:Illustration of the spherical and rectangular coor-dinate systems. Microphones are located in xm. Any fieldpoint is denoted byx.

a simpler model of the direct problem is used: qare planewave amplitudes. Then: Gml = ejklxm withklbeing thewavenumber vector for thel-th plane wave. We adhere totheejt time convention. The direct problem includes L

sources andMpressure sensors, hence explaining the ma-trix and vectors dimensions reported in Eq. (2).

The goal of inverse problem is to predict the measuredsound field at the microphone array: p pusing a knownsystem modelG. A typical approach to that problem is tocast it as a minimization problem with Tikhonov regulariza-tion [13]

q = argmin||pGq||2

2+ 2(q)2

. (3)

Accordingly, the quadratic sum of the prediction errore =pp at the microphone array is attenuated. The inverseproblem solution qshould approach the real sound source

distribution or should, at least, be able to achieve SFE. Thefact that the discrete direct model G is representative ofacoustical wave propagation suggests that if the predictionis good at the array (i.e. e 0), the recreated sound fieldin the vicinity of the array should also approach the exactmeasured sound field. In Eq. (3),|| ||2represents the vec-tor 2-norm, is the penalization parameter and () is adiscrete smoothing norm. For classical Tikhonov regular-ization, the discrete smoothing norm is the solution vector2-norm[13]:(q) =||q||2. The optimal solutionof Eq. (3)the becomes [8,13,14]

q= GHp

GHG + 2

I. (4)

Once the inverse problem solution q is obtained fromEq. (4), the extrapolated sound pressure field [Pa] at anylocationx is then computed using a linear combination ofplane waves

p(x) =Ll=1

eiklxql, (5)

where the complex plane wave distributionqlis centeredaround the coordinate system origin x = 0. Indeed, onenotes that the sound pressure at the origin is the direct linearcombination of the plane wave complex amplitudes

p(0) =Ll=1

ql. (6)

Page 2 of17


3/17

However, one should keep in mind that a plane wave sourcemodel could be replaced with any source model. Therefore,the selection of the plane wave sources does not limit thegenerality of this method.

For any field point x that excludes the array origin, it wouldbe interesting to obtain an expression similar to Eq. (6)where a new complex plane wave distribution qwould becentered around the field point x. This is expressed as fol-lows

p(x) =Ll=1

ql(x), (7)

withql(x) = e

jklxql. (8)

This simple expression will allow the direct computation ofAmbisonics signals in xfromq. According to Eq. (5), it isnow assumed that the inverse problem solution qlis a planewave description of the measured sound field. It will nowbe refereed as the full-spherical plane wave description ofthe sound field. This operation somewhat corresponds to athree-dimensional spatial Fourier (or plane wave) transform[6]to obtain an angular spectrum. However, it was achievedhere withoutanyassumption about the array geometry. Spa-tial Fourier transform methods are more constraining: theyneed a regular microphone array geometry that fits the spa-tial transform coordinates system [6].

2.2. Review of beamforming

It is possible to write the simple delay-and-sum beamform-ing spatial responses QBF L using [15]

QBF = GHp, (9)

or for thel-th listening direction (or point)

QBFl =gHl p, (10)

where the columns glof matrix G (as in Eq. (1)) exactlycorresponds to a classical non-normalized steering vectorused for non-focused (plane waves) beamforming (it is alsopossible to use focused beamforming if point sources wereused in the initial direct problem Eq. (1)). Indeed, the steer-ing vector is the evaluation of the Green function for one

listening direction (or point) at the microphone array. Thiscorresponds to the Gdefinition.

2.3. Inverse problem: beamforming regularization

In our recent theoretical work[3], we finally reached theconclusion that inverse problem sensitivity to measurementnoise is best controlled using Tikhonov-like regularizationmethod. To obtain an even better spatial resolution, we in-vestigated the possibility to use a beamforming regulariza-tion matrixLin the discrete smoothing norm: BF(q) =||Lq||2. The original idea was to take into account a pri-ori information obtained by delay-and-sum beamforming towisely regularize the inverse problem. The diagonal matrix

Lis given by

L=

diag|GHp|/||GHp||

1

LL. (11)

where | |denotes elementwise absolute value of the argu-ment and|| || is the vector infinite norm [16]. As onecan note, the effect ofLis to put a stronger penalization onthe sources inq for which the delay-and-sum beamforminggives a low output. Solution of the minimization problemEq. (3)is

qBF = GHp

GHG + 2LHL, (12)

or

qBF = GHp

GHG + 2

diag (|QBF|/||QBF||)21

,

(13)wherediag()2 represents the elementwise squared valuesof a vector placed on the main diagonal of a square ma-trix. Preliminary theoretical studies provided promising re-sults. Further theoretical developments and explanations of

this method is the topic of manuscript under review for theJournal of Sound and Vibration[11].

3 DERIVATION OF AMBISONICS SIGNALSFROM PLANE WAVE DESCRIPTION

In this section, the derivation of the HOA signals from theinverse problem solution (qor qBF) is first presented. Toillustrate the validity of the proposed derivation method, atheoretical test case is reported for the direct comparison ofthe exact HOA signals with the predicted HOA signals.

3.1. Theory

Within the field of spatial audio, Ambisonics is an estab-lished method both for sound field reproduction and soundfield capture [17,18]. As it will be shown in the next para-graphs, the proposed SFE method based on irregular mi-crophone array and inverse problem formulation can eas-ily be used to derive the Ambisonics signals from the in-verse problem solution q or qBF , as long as the spatialaliasing criterion is respected. Moreover, one of the greatinterest of Ambisonics signals derivation from SFE or in-verse problem theory is that virtual HOA microphones canbe placed anywhere in the effective SFE region. This can-not be done using a single sound-field microphone[17] in

a given point. This possibility and corresponding nomen-clature are depicted in Fig.2. This opens onto many poten-tial applications, including the post-processing translationof the sound field capturepoint that feeds the HOA decodersat the reproduction stage. This possible dependency of theHOA signals (W,X,Y ,Z,R,S,T,UandV) on spatialcoordinatesx is a new view of the HOA signals. To high-light this different paradigm, the dependency of the HOAsignals will always be identified as in W(x)wherexis thetranslated coordinate system origin as introduced in Eq. (7)and shown in Fig.2.

One can directly derive the B-Format Ambisonics first-

order signalsW(x), X(x), Y(x)and Z(x)from the zeroand first order spherical harmonics [6] coefficients that cor-responds to a plane wave distribution q (q or qBF) [5].

Page 3 of17


4/17

Figure 2: Schematic representation of the coordinate sys-tem and virtual Ambisonics first-order microphone in x.

However, in this paper, we will simply rely on the deriva-tion of the Ambisonics B-Format from virtual microphoneswith appropriate directivity (W, X , etc.) in agreementwith the Furse-Malham higher-order format [19]. There-fore, one writes the first-order Ambisonics signals for any

field point x from the translated plane wave distributionq(x,q)(Eq. (8))

W(x) =Ll=1

W(

l,

l)ql = 1/

2

Ll=1

ql, (14)

X(x) =Ll=1

X(

l,

l)ql =Ll=1

cos(l)cos(

l)ql, (15)

Y(x) =

Ll=1

Y(l, l)ql =

Ll=1

sin(l)cos(l)ql, (16)

Z(x) =Ll=1

Z(

l,

l)ql =Ll=1

sin(l)ql, (17)

where the angles l = l + and

l =l are the lis-tening directions in x. They are introduced to distinguishthe plane wave propagation directions l, l from the cor-responding listening directions l,

l. In the case of full-sphere second-order Ambisonics, the R(x), S(x), T(x),U(x), andV(x)components are easily derived as follows

R(x) =

Ll=1

R(

l,

l)ql=

Ll=1

(1.5 sin(l)2 0.5)ql (18)

S(x) =Ll=1

S(

l,

l)ql =Ll=1

cos(l) sin(2

l)ql (19)

T(x) =

Ll=1

T(

l,

l)ql =

Ll=1

sin(l) sin(2

l)ql (20)

U(x) =Ll=1

U(

l,

l)ql =Ll=1

cos(2l)cos(

l)2ql (21)

V(x) =Ll=1

V(

l,

l)ql =Ll=1

sin(2l) cos(

l)2ql (22)

x2

x1

x3

Figure 4:Microphone array geometry used for the reportedtheoretical and experimental results. The microphone cap-sules are marked by black dots. To increase the visibilityof the array, vertical black lines are shown between the mi-crophone acoustical centers and thex1x2plane. The greylines create an horizontal grid that corresponds to the x1, x2

positions of the microphones.

The typical directivity patterns (W, X , etc.) associatedwith these Ambisonics signals are illustrated in Fig.3(page5).

3.2. Verification

To verify the efficiency of the proposed HOA signals deriva-tion from SFE and corresponding plane wave distributionas reported in Eqs.(14)to (22), a numerical simulation isproposed to compare the exact HOA signals (W(x), X(x),etc.) to the derived HOA signals (W(x), X(x), etc.) over

an extended SFE region.

To evaluate the exact HOA signals inx from numerical sim-ulation, several assumptions are required. First, it is as-sumed that the HOA microphones used for that purpose arepoint-like with ideal directivity patterns (W,X ,Y, etc.)[20]. Accordingly, the exact HOA signals are then the exactlocal sound pressurep(x)weighted by the directivity valuethat correspond to exact local sound intensity opposite di-rection, [20]. Therefore, we write

W(x) = W(,)p(x), (23)

X(x) = X(,)p(x), (24)

Y(x) = Y(,)p(x), (25)

Z(x) = Z(,)p(x), (26)

R(x) = R(,)p(x), (27)

S(x) = S(,)p(x), (28)

T(x) = T(,)p(x), (29)

U(x) = U(,)p(x), (30)

V(x) = V(,)p(x). (31)

The verification test case involves an array of 96 micro-phones. The microphone array configuration is shown in

Page 4 of17


5/17

Figure 3:Second-order full-sphere Ambisonics directivity patterns according to Eqs. (14) to(22).

Fig.4.The microphone array is arranged in a double-layer(12.25 cm vertical space) rectangular array and aligned ona horizontal grid with a spacing of 12.25 cm. The sourcedistribution used in the direct and inverse problems is madeof a spherical distribution of impinging plane waves. Thedirection cosines of the propagation direction of the planewave distribution is shown in Fig.5, 642 plane waves areused for the theoretical verification.

The true sound field and HOA signals at 600 Hz for a singlemonopole inx1 2.5m, x2 = x3 = 0m in free fieldare shown in Fig.14(page12). One can note the polarityreversal of the HOA signals as x changes. This is clearin the case ofY(x)shown in Fig.14(d)where a movementalong x2involves a polarity reversal when one crosses x2 =0. One also observes that Z(x), S(x)and T(x)are null.This was expected since the source is located in the x1x2plane that corresponds to the illustration plane.

To achieve SFE using inverse problem theory and the beam-forming regularization matrix, the inverse problem solutionqBF(l, l)is obtained from Eq. (9) with = 0.00001.

The inverse problem solution qBF(l, l) is shown inFig.6 (page6). From this figure, it is clear that the SFEmethod based on inverse problem with the beamformingregularization matrix is able to localize the sound sourcefrom the negative x1.

The extrapolated sound field p(x)and deduced HOA sig-nals (W(x), X(x), Y(x), etc.) are shown in Fig. 15(page13). To identify the effective size of the SFE region,contour lines of the local quadratic error e2(x) (e2(x) =|p(x)p(x)|2, e2(x) =| W(x)W(x)|2, etc.) are su-perimposed on the field values. Important conclusions canbe drawn from this theoretical verification test case: 1) The

SFE extrapolation method is able to predict the sound fieldin a region which is larger that the microphone array and 2)The inverse problem solution and SFE methods are able to

0.50

0.51

0.5

0

0.5

11

0.5

0

0.5

1

cos(l)cos(

l)sin(l

)cos(l)

sin(l)

Figure 5:Direction cosines of the spherical plane wave dis-tribution withL = 642plane waves.

predict the HOA signals in an effective area which is alsolarger that the microphone array. This validates the effec-tiveness of the proposed method for the derivation of HOAsignals from an inverse problem approach to a measuredsound field using an irregular microphone array as reportedin Secs.2and3.1.

We recall that the effective size of the SFE area (and cor-responding HOA signals predictions) are, like for classicalAmbisonics using single sound-field microphone that pro-vides first-order directivity pattern shown in Fig.3,depen-

dent on frequency. Larger effective extrapolation area areexpected for the longer wavelengths while smaller effectiveSFE area are expected for smaller wavelengths. This is il-

Page 5 of17


6/17

Figure 6: Absolute value (normalized to unity) of the in-verse problem solution|qBF(l, l)| (linear (radius) anddB ref 1 (color) scales) for the verification case.

lustrated in Fig.16(page14) where SFE results and localquadratic errors are reported for several frequencies for thesame reported configuration. This should be kept in mindfor practical applications: virtual HOA sound field captureshould be done in relative close vicinity of the exact micro-phone array.

4 EXPERIMENTAL RESULTS

To first validate the efficiency of the proposed method

for SFE and sound source localization, a sound field wasmeasured using a microphone array in a hemi-anechoicroom. The sound field was created using an omnidirectionalsource. The array was moved in the vicinity of the originalsource position and the sound field was again measured inthat extrapolation region. Subsequent multichannel signalprocessing as presented in Sec. 2was applied to the firstmeasurement data to predict the sound field in the extrap-olation region. To achieve the validation, we present: 1)the corresponding inverse problem solution (full-sphericalplane wave description of the sound field) for comparisonwith the real source known position and 2) the compar-

ison between the measured sound field in the extrapola-tion region and the extrapolated sound field in that region.Once this experimental validation of the SFE methods isachieved, the predicted HOA signals are presented for theexperiments in Sec.4.3.

4.1. Experimental setup

The setup is shown in Figs.4, 7,8(a)and 8(b)(page7).The microphone array is made of 96 custom-made micro-phones arranged in a double-layer (12.25 cm vertical sep-aration) rectangular array and aligned on a horizontal gridwith a spacing of 12.25 cm. The microphones are madeof electret 6 mm capsules, each microphones sensitivity

is calibrated at 1 kHz. These pressure sensors are con-nected to eight-channel custom-made preamplifiers usinghigh-quality analog components. These custom-built hard-

(a) Electret microphones with casing

(b) 8-channel preamplifiers circuits

(c) 24-channel preamplifier rack

Figure 7:Pictures of the custom-built electret microphonesand 24-channel preamplifier racks.

ware parts are shown in Fig.7. They are designed for in-flight measurements of aircraft cabin sound environment

evaluation and characterization. The preamplifier outputsare digitized using four MOTUTM 24IO sound cards con-nected to a computer through AudiowireTM cables. The ar-ray was installed in a hemi-anechoic chamber and the orig-inal sound field was created using a 600 Hz sinus of 1-slength on a 12-loudspeaker B&K omnidirectionalsource lo-cated in the plane of the antenna. The experimental setup isshown in Fig.8(a).

4.2. Validation of the sound field extrapolation

The inverse problemsolutions using classical Tikhonov reg-ularization q(Eq. (3)) or beamforming regularization ma-trix qBF(Eq.(12)) are shown in Fig.9for the microphone

array located in position b) (see Fig.8(b)). For these twocases the regularization parameter was set to 1. Also notethat a much denser plane wave distribution ofL = 1442

Page 6 of17


7/17

(a) Photography of the experimental setup in the hemi-anechoic room.

5 4 3 2 1 0 1 2

2

1.5

1

0.5

0

0.5

1

1.5

2

x1[m]

x2

[m]

a) b)

(b) Top view of the sound source and microphone array ar-rangement in hemi-anechoic room.

Figure 8: Photography of the experiments (a) and experi-mental configuration (b). The sound source is shown as alarge black circle. The acoustic center of the sound sourceis shown as a small white dot. Two microphone array posi-tions are labeled a) and b). For the theoretical test case, themicrophone array position was b). The spatial regions thatcorresponds to the SFE evaluation area is circumscribed bya dash-line square.

plane wave was used. Clearly, the inverse problem solutionwith Tikhonov regularization is not accurate: the solution isnot able to localize the sound source nor the floor reflection.However, this does not mean that SFE based on classicalTikhonov regularization does not work, it only means thatfor this specific regularization parameter, the regularizationwas not strong enough. For the inverse problem solutionusing the beamforming regularization matrix with a simi-lar regularization parameter , the solution gives somethingthat is meaningful. Indeed, one can distinguish the directsound source (plane waves that propagates along positivex1) from the floor reflection (plane waves that goes up-ward). Therefore, this illustrates the fact that the inverse

problem method using the beamforming matrix requiresmuch less regularization to provide a meaningful result.The inverse problem solution with classical Tikhonov reg-

(a) Inverse problem solution |q|

(b) Inverse problem solution |qBF|

Figure 9: Spherical mapping of the absolute value (nor-malized to unity) of the inverse problem solutions: (a) withclassical Tikhonov regularization|q(l, l)|and (b) withbeamforming regularization matrix|qBF(l, l)| (linear(radius) and dB ref 1 (color) scales) (= 1) for the experi-ments with the microphone array position a) (see Fig.8(b)).

ularization and beamforming regularization with a stronger

penalization parameter = 25are shown in Fig.10. Thistime, the Tikhonov solution is able to localize the soundsource and the ground reflection. However, the spatial res-olution is much less compared to the inverse problem withthe beamforming regularization matrix shown in Fig.9(b).

These four inverse problem solutions evaluated from a mi-crophone array measurement in position b) (see Fig.8(b))were then used to predict (Eq.(5)) the sound field in arrayposition a) (see Fig.8(b)) for direct comparison of the SFEpredictions with the true measured sound pressure field.This comparison is presented in Fig.17(page15). Besidethe direct comparison of the original and extrapolatedsound

field, the local absolute value of the prediction error is alsoshown. Note that the color scale used for the prediction er-ror is not the same than for the sound fields. As expected

Page 7 of17


8/17

from the previous remarks, SFE using classical Tikhonovregularization (Figs.17(b)and17(c)) with = 1does notprovide an efficient SFE. SFE using the beamforming reg-ularization (Fig.17(d)) with = 1does however providea smooth extrapolation that approaches the original soundfield. Moreover, one notes that for this case the prediction

error is lower in the close vicinity of the measurement ar-ray (x1 closer to 0) and higher as the extrapolation goesaway from the measurement array (toward negative x1):the extrapolation error increases as one moves away fromthe microphone array that provided the original measure-ment p(xm). Interestingly, SFE using classical Tikhonovregularization (Figs.17(b)and 17(c)) with = 25 doesprovide a relatively efficient SFE. Sound field extrapolationusing the beamforming regularization matrix or classicalTikhonov regularization with = 1and = 25, respec-tively, are able to provide and efficient SFE, see Figs.17(d),17(e),17(h)and17(i).This illustrates a specific and advan-

tageous feature of the beamforming regularization matrix incomparison with the classical Tikhonov regularization: theformer method is much sensitive to the variation and selec-tion of the regularization parameter. Moreover, one shouldkeep in mind that the beamforming regularization matrixmethod provided the most efficient extrapolation of all thereported test cases.

To simplify the comparison of the prediction errors shownin Fig. 17, the normalized quadratic sum of the predic-tion error in the extrapolation area was computed: E =1/M

Mm=1 |e(xm)|2 [Pa2] for each of the prediction error

fielde(x)shown in Fig.17. For the Tikhonov regulariza-

tion with = 1,E = 0.6446Pa2

. For the Tikhonov reg-ularization with = 25,E = 0.2319Pa2. For the beam-forming regularization approach with = 1,E = 0.2070Pa2 and with = 25,E = 0.2194Pa2. For the four testcases, the beamforming regularization approach systemati-cally gives a lower SFE error in the extrapolation area. Thisvalidates the aforementioned feature. The general behav-ior of the SFE efficiency for the two methods as functionof the regularization parameter is summarized in Fig.11.One clearly notes that for the beamforming regularizationmethod the prediction error goes through a larger minimumplateau: this highlights the fact that the beamforming regu-larization is much less sensitive to the selection of the regu-

larization parameter .

4.3. Extrapolation of HOA signals

In this section, another experimental test case is reported.This time, a set ofL = 642plane waves as shown in Fig.5is used. This test case corresponds to the source and micro-phone array positions used for the theoretical test case. Theinverse problem solution based on beamforming regulariza-tion matrix (Eq. (12)) is shown in Fig.12with = 0.5. Bycomparisonwith the theoretical test case, one notes the floorreflection that is superimposed to the direct sound.

Before actually taking a look at the extrapolated sound field

and HOA signals, the comparison of the actual measure-ment pand predictionp at the measuring microphone ar-

(a) Inverse problem solution |q|

(b) Inverse problem solution |qBF|

Figure 10: Spherical mapping of the absolute value (nor-malized to unity) of the inverse problem solutions: (a) withclassical Tikhonov regularization|q(l, l)|and (b) withbeamforming regularization matrix|qBF(l, l)| (linear(radius) and dB ref 1 (color) scales) (= 25) for the experi-ments with the microphone array position a) (see Fig.8(b)).

ray using the inverse problem solution qBF is shown in

Fig.13. Clearly, the predicted sound field (p = GqBF)approaches the measured sound field (p). This is sup-ported by a low prediction error at the microphone array(Fig.13(c)). Therefore, it is assumed that the inverse prob-lem solution achieves its goal, i.e. ensures a prediction errorclose to zero e0 at the measuring microphone array (seeSec. (3.1)).

The SFE and predicted HOA signals in a SFE area that ex-tends beyond the microphone array are presented in Fig.18(page16). It is important to keep in mind that the size ofthe region surrounded by the walls of the anechoic spaceis relatively small (see Fig.8(a)). Indeed, the SFE region

shown in Fig. 18(page16) that covers2 x1 2,2 x2 2extends somewhat beyond the surroundingboundary. Therefore, the SFE results are not expected to be

Page 8 of17


9/17

102

101

100

101

102

101

100

101

102

103

Normalizedquadraticsumo

fthe

predictionerror(E)

Classical Tikhonov regularization

Beamforming regularization

Figure 11:Normalized quadratic sum of the prediction er-ror in the extrapolation area using classical Tikhonov reg-ularization and beamforming regularization as function ofthe penalization parameter . The cases with = 1 and= 25are marked by thin vertical dashed lines.

valid in that region. As expected from the equivalent the-oretical test case, the strongest HOA signals areW,X,Y,R, Uand V. However, by marked contrast with the theoret-ical test case, the concrete floor reflection introduces somesignificant signals forZandS. This is expected since theground reflection involves a propagative component along

x3. Interestingly, one notes that theZHOA signal involvesa trace wavelength in the x1x2 plane that is longer thanfor the other HOA signals. This suggests that the ZHOAsignal is able to tackle the ground reflection which involvesa longer trace wavelength in thex1x2plane according tothe non-null wavenumber vector component along x3, i.e.kx3= 0for the ground reflection. This completes the ex-perimental prediction of HOA signals from an irregular mi-crophone array measurement.

5 APPLICATION EXAMPLE

To facilitate the understanding and illustrate the potential ofthe proposed method, an application example is presentedschematically in Fig. 19 (page 17). For a given sound sourceand sound environment (ambient sounds, room acoustics,etc.), a microphone array measurement is done and the rawdata is stored. Using this raw data, virtual HOA signals arederived for any given spatial location x. These signals arefirst obtained by the transformation of the sound pressuresat the microphone array in a plane wave description of thesound field using inverse problem in the frequency domain(Sec.2). Next, the full-spherical plane wave description ofthe measured sound field is directly used to compute the

predicted HOA signals using virtual point-like microphonewith appropriate directivity patterns (Sec.3.1). These HOApredicted signals are then converted back in the time do-

Figure 12: Absolute value (normalized to unity) of the in-verse problem solution|qBF(l, l)| (linear (radius) anddB ref 1 (color) scales) for the experiments with the micro-phone array position b) (see Fig.8(b)).

main to be used with any classical HOA decoder and 2D or3D loudspeaker array.

6 CONCLUSION

The aims of this paper were twofold: provide the exper-imental validation of a recently proposed sound field ex-trapolation (SFE) method [11]and provide the derivationof higher-order Ambisonics (HOA) signals from a plane

wave description of a measured sound field obtained fromthe SFE method.

In a first instance, the SFE method combining inverse prob-lem theory and the beamforming regularization matrix wasrecalled. The obtained inverse problem solution was usedto directly derive the HOA signals in an extended extrap-olation area. Generally speaking, the objective of the in-verse problem approach is to find a potential cause that cre-ated a measured effect. In applied acoustics, the cause isan a prioriunknown sound source and the effect is an ob-served sound pressure field. To illustrate the validity of theproposed SFE method and HOA signals derivation, a sim-

ple theoretical test case was presented. It was shown thatboth SFE and HOA signals predictions were efficient in aextended SFE area. As for classical finite-order Ambison-ics description or reproduction of a measured sound field,the effective size of the SFE region tends to get smaller forsmaller wavelengths.

In a second section, we described experiments to evaluatethe efficiency of the new regularization method of the in-verse problem using the beamforming regularization ma-trix. The experiments were achieved in known and con-trolled conditions (an hemi-anechoic room) for validationpurposes. By comparison with classical Tikhonov regular-

ization method it was shown that this technique can offer amore precise sound source localization. Moreover, the ex-trapolated sound field seemed even closer to the monitored

Page 9 of17


10/17

(a) Measured sound fieldp(xm)

(b) Predicted sound fieldp(xm)

(c) Prediction errore = p p

Figure 13:Real and imaginary parts (left and right, respec-tively) of the measurement (p(xm)), prediction (p(xm))and prediction error (e = pp) of the experimental testcase at 600 Hz at the microphone array.

sound field than for the classical regularization method. Atthe light of the results presented in this paper, we believein the potential future developments, refinements, applica-tions of the beamforming regularization method in inverseproblem.

To illustrate the potential application of the HOA signalsderivation for an extended region from the full-sphericalplane wave description of the sound field, the HOA signalswere derived from and illustrated for the experimental SFEresults. Further works could be devoted to the objective

and subjective comparisons of the predicted and measuredHOA signals. Indeed, a direct comparison of the predictedHOA signals at a given point in space with actually mea-sured HOA signals in that same position using a Ambison-ics microphone would allow for an even more rigorous val-idation of the proposed HOA signals derivation.

The proposed derivation of HOA signals from a genericplane wave distribution that describes a given measuredsound field opens many new possibilities for subsequentHOA reproduction. Indeed, from a single microphone arraymeasurement, one can derive at a latter processing stage theHOA signals for any points in the SFE effective area. This

is specially interesting for virtual acoustics or sound envi-ronment reproduction since it would be possible to exposea listener to an HOA reproduction of a given sound field

in any points of the SFE effective area. Moreover, the pro-posed method is not limited to the studied microphone arrayconfiguration. Indeed, it can be applied to any microphonearray geometry. This could be subject of further investiga-tions and verifications.

Beside the evaluation of the HOA signals from the full-spherical plane wave description of the sound field (or an-gular spectrum [6]) obtained from the inverse problem ap-proach, it would also be interesting to apply the deriva-tion of HOA signals from an angular spectrum obtainedfrom other methods such as near-field acoustical hologra-phy (NAH) that typically involves a large microphone arrayin close vicinity of the sound source under study. That isa virtual Ambisonics avenue that could be very interesting.Indeed, once a sound source is characterized using NAHin anechoic conditions, it would be possible to derive anyHOA signals for any spatial coordinate using SFE as writ-ten in the spatial transform domain for NAH. It would then

be possible to listen to the measured sound source at anyrelative position from the sound source. This is especiallyinteresting for virtual acoustic applications, listening tests,sound quality and comfort studies.

7 ACKNOWLEDGMENTS

This work has been supported by CRIAQ (Consortium deRecherche et Innovation en Arospatiale Qubec), NSERC(National Sciences and Engineering Research Council ofCanada) Bombardier Aerospace and CAE.

REFERENCES

[1] M.A. Poletti, "Three-dimensionalsurroundsound sys-tems based on spherical harmonics", Journal of theAudio Engineering Society, 53(11), 2005, 1004-1025.

[2] E. Hulsebos, D. de Vries, E. Bourdillat, "ImprovedMicrophone Array Configurations for Auralization ofSound Fields by Wave-Field Synthesis", Journalof theAudio Engineering Society, 50(10), 2002, 779-790.

[3] P.-A. Gauthier, C. Camier, Y. Pasco, . Chambatte,A. Berry, "Sound field extrapolation: Inverse prob-lems, virtual microphone arrays and spatial filters",

AES 40th

International Conference, Tokyo, Japan,October 2010.[4] C. Camier, P.-A. Gauthier, Y. Pasco, A. Berry, "Sound

field reproduction applied to flight vehicles soundenvironments", AES 40th International Conference,Tokyo, Japan, October 2010.

[5] J. Daniel, R. Nicol, S. Moreau, "Further investigationsof high order ambisonics and wavefield synthesis forholophonic sound imaging", Audio Engineering Soci-ety 114th Convention, Amsterdam, March 2003, Con-vention paper 5788.

[6] E.G. Williams, Fourier acoustics Sound radiationand nearfield acoustical holography, Academic Press,

San Diego, 1999.[7] H. Teutsch, W. Kellermann, "Acoustic source detec-

tion and localization based on wavefield decomposi-

Page 10 of17


11/17

tion using circular microphone arrays", Journal of theAcoustical Society of America, 120(5), 2006, 2724-2736.

[8] P.A. Nelson, S.H. Yoon, "Estimation of acousticssource strength by inverse method: Part I, condition-ing of the inverse problem", Journal of Sound and Vi-

bration, 233(4), 2000, 643-668.[9] S.H. Yoon, P.A. Nelson, "Estimation of acoustics

source strength by inverse method: Part II, experi-mental investigation of methods for choosing regular-ization parameters", Journal of Sound and Vibration,233(4), 2000, 669-705.

[10] J. Merimaa, "Applications of a 3-D microphone ar-ray", Audio Engineering Society 112th Convention,Munich, May 2002, Convention paper 5501.

[11] P.-A. Gauthier, C. Camier, Y. Pasco, A. Berry,. Chambatte, "Beamforming regularization matrixand inverse problems applied to sound field measure-ment and extrapolation using microphone array", un-der review Journal of Sound and Vibration, 2011.

[12] A.D. Pierce, Acoustics: An introduction to its physi-cal principles and applications, Acoustical Society ofAmerica, Woddbury, 1991.

[13] C. Hansen, Rank-deficient and discrete ill-posed prob-lems, SIAM, Philadelphia, 1998.

[14] P.A. Nelson, "A review of some inverse problems in

acoustics", International journal of acoustics and vi-bration, 6(3), 2001, 118-134.

[15] D.N. Ward, R.A. Kennedy, R.C. Williamson, "Con-stant directivity beamforming", in: M. Brandstein,D. Ward (Eds.), Microphone arrays: Signal processingtechniques and applications, Springer, Berlin, 2001, 3-

17.[16] G.H. Golub, C.F. van Loan, Matrix computations,

John Hopkins University Press, Baltimore, 1996.[17] F. Rumsey, Spatial audio, Focal Press, Burlington,

2001.[18] J. Daniel, J.-B. Rault, J.-D. Polack, "Ambisonics en-

coding of other audio formats for multiple listeningconditions", Audio Engineering Society 105th Con-vention, San Francisco, September 1998, Conventionpaper 4795.

[19] A. Southern, D. Murphy, "A second order differentialmicrophone technique for spatially encoding virtualroom acoustics", Audio Engineering Society 124th

Convention, Amsterdam, May 2008, Convention pa-per 7332.

[20] H. Hacihabiboglu, B. Gnel, Z. Cvetkovic, "Simula-tion of directional microphones in digital waveguidemesh-based models of room acoustics", IEEE Trans-actions on Audio, Speech, and Language Processing,18 (2010) 213-223.

Page 11 of17


12/17

(a) Exact sound field p(x) (b) Exact W(x)

(c) Exact X(x) (d) ExactY(x)

(e) Exact Z(x) (f) Exact R(x)

(g) Exact S(x) (h) Exact T(x)

(i) Exact U(x) (j) ExactV(x)

Figure 14:Real and imaginary parts (left and right plots of the figures, respectively) of the exact sound pressure fieldp(x)

and Ambisonics signal fields (W(x),

X(x), etc.) for the theoretical verification case at 600 Hz with a single monopolesource in free field.

Page 12 of17


13/17

(a) Extrapolated sound fieldp(x) (b) Extrapolated W(x)

(c) Extrapolated X(x) (d) Extrapolated Y(x)

(e) Extrapolated Z(x) (f) Extrapolated R(x)

(g) Extrapolated S(x) (h) Extrapolated T(x)

(i) Extrapolated U(x) (j) Extrapolated V(x)

Figure 15:Real and imaginary parts (left and right plots of the figures, respectively) of the extrapolated sound pressurefieldp(x)and HOA signals (W(x), X(x), etc.) for the theoretical verification case at 600 Hz with a single monopole

source in free field. The local quadratic errors (e2(x) =|p(x) p(x)|2, e2(x) =|W(x) W(x)|2, etc.) are identified ascontour lines ate2 = 0.001(white lines),e2 = 0.01(black dashed lines) ande2 = 0.1(black lines).

Page 13 of17


14/17

(a) Extrapolated sound fieldp(x)at 80 Hz (b) Extrapolated sound fieldp(x)at 160 Hz

(c) Extrapolated sound fieldp(x)at 240 Hz (d) Extrapolated sound fieldp(x)at 320 Hz

(e) Extrapolated sound fieldp(x)at 400 Hz (f) Extrapolated sound fieldp(x)at 480 Hz

(g) Extrapolated sound fieldp(x)at 560 Hz (h) Extrapolated sound fieldp(x)at 640 Hz

(i) Extrapolated sound fieldp(x)at 720 Hz (j) Extrapolated sound fieldp(x)at 800 Hz

Figure 16:Real and imaginary parts (left and right plots of the figures, respectively) of the extrapolated sound pressurefieldp(x)for the verification case at different frequencies with a single monopole source in free field. The local quadratic

errors (e2

(x) =|p(x) p(x)|2

) are identified as contour lines ate2

= 0.001(white lines), e2

= 0.01(black dashed lines)ande2 = 0.1(black lines).

Page 14 of17


15/17

(a) Measured sound field at the monitoring array p(x)

(b) Extrapolated sound fieldp(x)(Tikhonov reg., = 1) (c) Prediction error |e(x)|=|p(x) p(x)|(Tikhonov reg., = 1)

(d) Extrapolated sound fieldp(x)(Beamforming reg. matrix,= 1)

(e) Prediction error |e(x)|=|p(x) p(x)|(Beamforming reg. matrix, = 1)

(f) Extrapolated sound fieldp(x)(Tikhonov reg., = 25) (g) Prediction error |e(x)|=|p(x) p(x)|(Tikhonov reg., = 25)

(h) Extrapolated sound fieldp(x)(Beamforming reg. matrix,= 25)

(i) Prediction error|e(x)| =|p(x) p(x)|(Beamforming reg. matrix, = 25)

Figure 17:Real and imaginary parts (left and right columns of (a), (b), (d), (f) and (h)) of the measured sound field p(x),

extrapolated sound field p(x)and absolute value of the prediction error|e(x)| =|p(x) p(x)|at the monitoring array(in an extrapolation area different from the measurement area) for different inverse problem solutions for the experimentsin hemi-anechoic chamber at 600 Hz. The measurement points are shown as black circles. The color scale is different forthe prediction error.

Page 15 of17


16/17

(a) Extrapolated sound fieldp(x) (b) Extrapolated W(x)

(c) Extrapolated X(x) (d) Extrapolated Y(x)

(e) Extrapolated Z(x) (f) Extrapolated R(x)

(g) Extrapolated S(x) (h) Extrapolated T(x)

(i) Extrapolated U(x) (j) Extrapolated V(x)

Figure 18:Real and imaginary parts (left and right plots of the figures, respectively) of the extrapolated sound pressure

fieldp(x)and HOA signals (W(x),X(x), etc.) for the experimental case at 600 Hz. Approximate room boundary areshown as dashed black lines.

Page 16 of17


17/17

Figure 19:Illustrated application of the proposed method for two virtual Ambisonics microphones signals deduced froma single microphone array. Once the microphone array data is stored, it is possible to predict the HOA signals for anyposition located in the effective SFE area. These predicted HOA signals are then sent to a conventional HOA decoder andloudspeaker array for subsequent sound field reconstruction at the center of the array where the listener stands.

Page 17 of17

Documents

2011 DERIVATION OF AMBISONICS SIGNALS AND PLANE WAVE DESCRIPTION OF MEASURED SOUND FIELD USING IRREGULAR MICROPHONE ARRAYS AND INVERSE PROBLEM THEORY