Sei sulla pagina 1di 6

Use of random noise for on-line transducer modeling in an adaptive

active attenuation system a)


L.J. Eriksson and M.C. Allie

Corporate Research Department, Nelson Industries, P.O. Box600,$toughton, Inc., Wisconsin 53589-0600

(Received January 9 1987; accepted publication October for 26 1988) Activesound attenuation systems bedescribed may using system a identification framework in whichan adaptive filter is usedto modelthe performance an unknownacoustical of plant.An errorsignal maybeobtained froma location following acoustical an summing junction where the undesired noise combined is with theoutputof a secondary sound source. themodel For outputto properly converge a valuethat will minimize errorsignal, is frequently to the it necessary determine transfer to the function thesecondary of sound source thepathto the and errorsignal measurement. Since these transfer functions unknown continuously are and changing a real system, is desirable perform in it to continuous on-linemodeling theoutput of transducer errorpath.In thisarticle, use an auxiliary and the of random noise generator for thismodeling described. is Based a Galoissequence, technique easy implement, on this is to provides continuous on-line modeling, hasminimal and effect thefinalvalue theerror on of signal.
PACS numbers:43.60.Gk, 43.50.Ki

Although this systemidentificationproblemhas been intensively studied thecontrolandsignal-processing in literActivesound attenuation a relativelyold ideathat has is ature, the activeattenuation application complicated is by received considerable attentionin recentyears.This is prithe presence acousticfeedbackfrom the loudspeaker of to marily due to the development improved of signal-processthe input microphone. the past,a variety of solutions In to ing theoryand hardwarethat enablemoresophisticated apthis problemhavebeenproposed that utilize eitherdirecproaches thisproblem.Many of the traditionalproblems to arraysor incorporate compensating a fixed with this technology can now bc treated more effectively tionaltransducer with propersignalprocessing rather than with the direct acoustical approaches the past. of INPUT ERROR This articledescribes complete a activeattenuation sysMICROPHONE MICROPHONE temthat functions correctly the presence acoustic in of feed<") PLANT ERROR back as well as nonidealinput microphone, error micro-

INTRODUCTION

phone, loudspeaker, error pathtransfer and functions. is It completely adaptive andresponds automatically changes to in input signal,acoustic plant, error plant, microphone, and loudspeaker characteristics.
I. SYSTEM IDENTIFICATION

ACOUSTIC

FEEDBACK '

SPEAKER

Activesound attenuation systems maybe described usinga system identification framework whichan adaptive in filterisused modeltheperformance anunknown to of acous-

tical plant, shown Fig.1(a). 1-s input as in An microphone is


used measure undesired to the noise upstream theacoustiof

03) +

calplant. Thissignal used theinputto anadaptive is as filter that generates outputto a loudspeaker an whichis usedto
producea secondary soundthat is acoustically combined with the undesired noise.An error signalmeasured downstream from the acousticalsummingjunction is used to adaptthe coefficients the adaptive of filter to minimizethe residualnoise.When fully adapted,the adaptivefilter response series in with the response the inputmicrophone of and loudspeaker matchesthe response the acoustical of plant.
' Anearlier version this of article presented the112th was at Meeting the of
Acoustical Society America,8-12 December of 1986in Anaheim,CA [J. Acoust.Soc.Am. Suppl.I 80, SII (1986)].
797 J. Acoust. Sec. Am. 85 (2), February 1989

FIG. 1. (a) Schematic diagramof activeattenuation problem.(b) Block diagram active of attenuation problem showing general solution compento sation transfer for functions to loudspeaker due SanderrorpathE through the additionof duplicatetransferfunctions inverse or transferfunctions.
) 1989 Acoustical Society of America 797

0001-4966/89/020797-06500.80

feedback thatisdetermined anoff-line path on basis through calculations useof a trainingsignal. or

LMS algorithmwhenboth auxiliarypath and error path transfer functions present.A transfer are 0 function added is

Erikssonhas presented new techniquefor active ata to the input to the error correlators, whichrepresents the tenuation that effectively utilizesadaptive signalprocessing product the auxiliary of pathanderrorpathtransfer functo solve problem acoustic the of feedback fromthesecondary tions. Widrow Stearns have and .2 similarly discussed the sound source theinput to microphone. technique 6This uti- "filtered-X"LMS algorithm use for witha plantin theauxillizesa recursive-least-mean-squares (RLMS) algorithmdeiarypath..2 veloped Feintuch provide complete by 7to a pole-zero model Theseresults havebeenextended an infiniteimpulse to of the acousticalplant. The acousticfeedbackis considered response (IIR) adaptive filter usingthe RLMS algorithmby part of the adaptivemodel usedto model the plant. From Eriksson. speaker The transfer function anderrorpath S thisperspective, acoustic the feedback introduces fixedpoles transferfunctionE mustbe knownto compensate their for into the overall response the model, which may be reof effecton the convergence both the directand recursire of movedwith the pole-zeroresponse the RLMS algorithm. of elements the IIR filter. This can be donethrougheither of Using the configuration shown in Fig. l(b), the direct theadditionors andEinto theinputlinesto theerrorcorreacoustical pathP andfeedback acoustical Fare simulta- latorsor the addition the inverse path of transfer functions, Sneously modeledby the RLMS modelusingadaptivefilters andE-, into the errorpath,as shown Fig. 1(b). As in A andB, in series with the loudspeaker Perfectcancella- discussed S. above,theformertechnique beendescribed has by tion is obtainedwhen the overall model response matches Widrow andBurgessfor theLMS algorithm assures a and the response the plantor using transforms: of z that the error signaland input signalwill have the same relationship time.The lattertechnique beendescribed in has M r = MS/( 1 + FMS) = AS/( 1 -- B + FAS) = P, byMorganfortheLMS algorithm eliminates need 9 and the (1) for a modification the input signalin principle,but, in to practice,the lack of causalityfor the inversetransferfuncwhere

tions, andE- , requires Scompensation inputsigof the


nal to the error correlators delay,as will be discussed by in the following. Unfortunately,bothS and E are unknownand are time

M r = overallmodelresponse, M = responseof pole-zero recursive filter structure usedin RLMS algorithm, A = response all-zeroleast-mean-squares of (LMS) elementusedin directpath of pole-zero structure, B = response all-zeroLMS elementusedin recursive of path of pole-zerostructure, P = directpath acoustic plant, F--- feedback path acoustic plant,and S = response loudspeaker. of One solution to this equation is for 4 = P/S and B = PF. However,the actualmodelresponse a complex is function thespectral of content thesource theacoustiof and

varyingdue to effects suchas heat and agingon the loudspeaker dueto changes temperature flowin the and in and error path. Thus it is necessary obtain either direct or to inverse models S andE onanon-line of basis. Although they arenot shown explicitly Fig. 1(b), the errormicrophone in maybeconsidered partof theerrorpathtransfer as function E, and the input microphone simplyaddsan additional
transfer function in series with the RLMS model and loud-

speaker Since inputmicrophone S. the occurs prior to the adaptive model, does need becompensated in the it not to for
same manner as S and E.

cal plant of the system. The RLMS algorithm provides a fully adaptivemeansto simultaneously model the direct plantandfeedback plantwitha given source such wayas in a
to minimize the residual noise.

Poole al.3have et described a system using LMS the


algorithmin which a fixedcompensating inverse transfer function added theerrorpath.However, is to since and S-

E- arenoncausal, off-line an model a delayed of inverse


modelof theloudspeaker errorpathAS- ]E- I is deterand minedwhere A is the delay necessary make the inverse to modelcausal. The useof thisdelayed 4 inverse modelreduces errorpathtransfer the function a fixedpuredelayA. to As notedabove, thisapproach thenrequires additionof the the samedelayA to the input to the error correlators the of LMS algorithmasdescribed Widrow.8The primarydisby advantage thistechnique that it doesnot usean on-line, of is continuously adaptive modelof the loudspeaker error and path. Eriksson describedthree-microphone has a system usingtheRLMS algorithmin whichtheerrorplantismodeled
on line using either a direct or inverse model while the speakeris modeledoff line. However, there have not been

II. TRANSDUCER

MODELING

One of the problemswith this techniqueis that the RLMS algorithm requires knowledge thespeaker of transfer
function and error path transfer function for proper conver-

gence. Widrow has s shown theLMS algorithm be that can


usedwith a delayederror signalif the input to the error correlators alsodelayedby the sameamount.Similarly, is Morgan hasstated that, with propercompensation, LMS the algorithmcanalsobe usedifa transferfunction,suchasthat dueto the loudspeaker, in the auxiliarypath followingthe is adaptive filter. Propercompensation requires additionof the a transfer function theinputto theerrorcorrelators the in or
addition of an inverse transfer function in series with the

errorpath. Burgess discussed 9 has similarresults the for


798 J. Acoust. Soc. Am.,Vol. 85, No. 2, February1989

any previous approaches described that providean on-line model of the speakerand the error path that responds to changes their response in overtime.
798

L.J. Eriksson M. C. Allie:Randomnoisefor on-linemodeling and

III. MODELING

APPROACHES

The traditional solution is then to use either the direct or

inverse modeling approach shownin Fig. 2(a) and (b), reThere are two basictechniques available usein sysfor spectively, anoff-line on basis a broadband source with noise tem modeling usingadaptive filters.The directapproach N. Since isanoff-line it process, plantoutputy model the and places model parallel the in withtheunknown plantandis output arenotpresent. noise . The source thusallows N a

adapted such thatthedifference between outputs the the of plantandmodel minimized thesame is for signal. inThe verse approach places model series the in with theunknown plantsuch thatthedifference between outputof thissethe
riescombination a delayed and version the input signal of is minimized.In thiscase, response the adaptive the of model becomes delayedversionof the inverseof the unknown a plant response. As shownin Fig. 2(a), to determine speaker the and errorpathresponse, directmodelapproach the places the

precise determination eitherthespeaker errorpath of and response, or the delayed ,fiE, inverse modelof the speaker anderrorpath,AS- E . In the directapproach Fig. of 2(a), the response is fixedafter convergence then SE and
usedin the inputsto the error correlators the LMS or of RLMS algorithms. the inverse In approach Fig. 2(b), the of

adaptive model parallel in withthespeaker errorpath. and An errorsignal formedby subtracting adaptive the model outputfromthemicrophone outputismultiplied theinby putsignal formtheupdate to terms thecoefficientsthe for of adaptivemodel.The inverse modelapproach places the adaptive modelin series with thespeaker errorpath,as and shown Fig. 2(b). In thiscase, errorsignal in the formedby
subtracting adaptivemodeloutputfrom a delayedverthe sionof the noise inputismultipliedby the inputto theadaptivemodel formtheupdate to terms thecoefficients the for of adaptive model.Thusthe adaptive modelformsa delayed inverse modelof the speaker and the error path while attempting matchthe response thedelayed to of noise input.

response E- is alsofixedafterconvergence, the ASand modeling delay isused theinputs theerrorcorrelators A in to of the LMS or RLMS algorithms. Bothtechniques assume the useof a large-amplitude, broadband noise source an on
off-linebasis avoidcontamination the modeling to of process

byy or. andto avoid addition undesired the of noise during on-lineoperation the noise by source N.
IV. CONTINUOUS MODELING SYSTEM

A new approach the on-linemodeling S andE is to of shownin Fig. 3. An uncorrelated randomnoisesource is used excite series to the combination the speaker of followed by the error plant as well as adaptive modelC while the

system operating.Thisrandom is '5 noise source ultiwill


mately become source theresidual the of noise thesystem. of The directadaptive modelC is usedto obtaincoefficients describing response $ andE that canbeused the the of in input linesto the error correlators the primaryRLMS for algorithm. generalized The model outputandweight update equations the recursive for adaptivefilter may be written

following notation Widrow Stearnsas the of and j2


T Yk _ WkUk ,

(2)

Wk+ = W + 2MU;e , k

(3)

U= [u,u_, ....] , =cuk ,


where
^ +

(4) (5)

RANDOM

' )+
N

FIG. 2. (a) Directmodeling approach thedetermination thetransfer for of function thespeaker errorpathwithadaptive of and model (b) Inverse SE. modeling approach the determination the delayed for of inverse transfer

FIG. 3. Newapproach on-line to modeling speaker anderrorpathE of S andusing results RLMS model in withacoustic feedback forma fully to
adaptive active attenuation system.

function thespeaker error of and path withadaptive model A(SE)


799 J. Acoust.Soc. Am., Vol. 85, No. 2, February 1989

L.J. Erikssonand M. C. Allie:Random noisefor on-linemodeling

799

Yk = scalarmodeloutputat discrete time k, Wk = generalized weight vector (includesdirect and


recursire coefficients),

V. RESULTS

W r= transpose of Wk+ = updated generalized weightvector,


M = convergence factormatrix, e = scalarerror signal,

U = generalized input vector(includes directand recursiveinput vectors), U , = compensated generalized inputvector, u;, = firstcomponent compensated vector, of input and C r= transpose weight of vector model of associated with transferfunctions auxiliarypath. in
The weightvectorof the adaptivemodelC is obtained on an on-linebasis usingan adaptivealgorithmsuchas the LMS or RLMS algorithm with the independent random noise source aninputandtheerrorsignal shown Fig. as as in 3. The amplitude thenoise of source keptverylow sothat is the finaleffect the residual on noise small.The plantnoise is y and model output are not presentat the input to the adaptive modelC andsowill not affect finalvalues the the of modelweights.
The use of an uncorrelated random noise source that is

The resultsof a computersimulationof the system shownin Fig. 3 confirmed that the algorithmproperlyconvergesfor either narrow-bandor broadbandinput signals. The coefficients the SE model properly describe SE of the plant,andthe coefficients the overallsystem of modelproperly describe F, and S. P, The approach shownin Fig. 3 hasalsobeenimplementedoncomplete acoustical systems using TMS320 family the of digitalsignalprocessing microprocessors, inputmiwith crophones,canceling loudspeakers,and error micro-

phones."?'a Initially,oneof these systems utilized was to


cancelelectroacoustically generated noisein a 12-in.-diam circular duct. The duct was about 25 ft long and unlined exceptfor a short4-ft-longadsorptive silencer near the primary noise source. Typical results after adaptation are shownin Fig. 4. The noisereductionobtainedwith the system operating a broadband for noiseinput is shownin Fig. 4(a). This curvewasobtained subtracting canceled by the

independent the input signalensures of that the speaker and error path will be correctlymodeled. The signals from the

spectrum from the uncanceled spectrum. maximaand The minima in the spectrum due to acoustical are resonances. The converged weightstructure the4,B, and C elements for of Fig. 3 is shownin Fig. 4 (b). The decayof the coefficients confirms that the filter lengthchosen wasadequate. The system is effective on broadband as well as narrow-band noise

plant (y) andmodel(.) represent noise the "plant"side on of the speaker/error path modeling process will not afthat fect the weightsof the direct model C usedto determine

SE.2 Thismodel then is copied theinput to lines theerror of


correlators the RLMS algorithm. of It shouldbe notedthat, althoughthe delayedadaptive inverse modelshownin Fig. 2(b) couldbe usedin a similar fashion,this will resultin decreased performance sincethe "noise"in the auxiliary path and error path due to y and also appearsat the input of the adaptivefilter due to the series arrangement. Thusthe autocorrelation function the of filter input is adversely affected, and the filter weightsare

and requires calibration trainingof any kind. no or Performance an actualindustrial or heating, in fan ventilating,and air conditioning ductis mademuchmoredifficult by the turbulentairflowandlargeductdimensions that areusuallyrequired. Goodsystem performance requires antiturbulence microphones well as large,powerful,lowas

frequency sources.Typical 9 results obtained shown are in


Fig. 5. The uncanceled autospectrum a linedsupply in duct (34 X 44 in.) approximately ft from a centrifugal is 40 fan
(a)

modified described Widrow and Stearns. If this as by 2


"noise"is largeenough, adaptive the modelmayfail to converge. Thusthedelayed adaptive inverse approach requires a much largeramplituderandomnoisesource that increases the residualnoiseand decreases overall systemquieting. In the directmodelsystem, shownin Fig. 3, the "noise" duetoy and. does affectthe final weights the adaptive not in model.In addition,the convergence the SE model is asof sured longastheinitialamplitudes withinthedynamic as are range and signal-to-noise ratio constraints the system. of Thus, with $E accurately determined, the overall system
model will converge,resultingin minimum residual noise. The randomnoisesource usedto modelSE may be read-

-10

0
()

HZ

2OO

ily obtainedthroughthe useof a varietyof methods. One simple approach is to generatea Galois sequence using

methods described Schroeder.A Galois by 6 sequencea is


pseudorandomsequence that repeats after 2'"1 points,

where isthenumber stages a shiftregister. iseasy rn of in It to calculate caneasily and havea period muchlonger thanthe response time of the System. this'study,31 stages In
(m = 31 } were used.

FIG. 4. (a) Noisereduction with activeattenuation system for bandlimon ited (15-200 Hz) pink noiseinput signal(no flow--128 averages).(b) Filter coefficients to obtaintheresults used shown (a) for adaptive in filters .4 (32 taps),B (64 taps), and C (64 taps).

800

J.Acoust. Am., 85,No. February Sec. Vol. 2, 1989

L.J.Eriksson M.C.Allio: and Random foron-line noise modeling

800

- 40 ('

20 (<

WV V TM

-90
o

-10

HZ

200

HZ

200

-40

1.00 (d)

-90 t
0

HZ

200

0.80 .......
0

HZ

200

FIG. 5. (a) Relativesound-pressure spectrum discharge in duct of centrifugal with activeattenuation fan system [Mach number(M): 0.04-128 off averages]. Relative (b) sound-pressure spectrum discharge in ductof centrifugal with activeattenuation fan system (M = 0.04-128averages (c) Noise on ). reduction obtained fromFig. 5(a) and (b). (d) Typicalcoherence between inputmicrophone errormicrophone and before cancellation the used obtain to
the results Fig. 5(c) (M = 0.04-128 averages). of

shown Fig. 5(a). In addition a very-low-frequency in to peak at about8 Hz, there is a broadpeakof noisefrom about40
Hz to about 140 Hz. With the active soundcontrol system,

is essentially noiseaddedat any frequency. no Additional

performance have results been presented elsewhere. 3'2


Although theseresultsare substantially better than typicalpassive silencers, theredo not appear beany reato sonswhy the performance this systemcannotbe inof creased. primaryareasof potentialimprovement to The are obtainbettercoherence between input and error microthe phones through useof improved the antiturbulence microphones to increase speed computation and the of through the useof more powerfulmicroprocessors.
Vl. CONCLUSIONS

this broadpeak was reduced,as shownin Fig. 5(b). The noise reduction plottedin Fig. 5(c). Thereisa broadrange is of attenuation from about40 to 140 Hz peakingat about 18 dB. System performance limitedby theeffectiveness the is of antiturbulence microphones. There is minimal attenuation in the8- to 40-Hz rangeduetOthelackof coherence between the input and error microphones thesefrequencies, at as shown Fig. 5(d). It should notedthat it hasbeenfound in be that themagnitude the coherence of mustbeon the orderof 0.95 or greater cancellation beeffective. excellent for to The
coherence from about 40-140 Hz is consistent with the at-

A completeactive attenuationsystemhas been describedin which acoustic feedback modeledas part of an is

tenuation shown Fig. 5(c). It wouldbe difficultto obtain in this performance usinga conventional passive silencer. In addition,the activeattenuation system results essentially in no restrictionto the flow, thus avoidingthe needto modify thefandrive.As before, system fully adaptive, the the is and resultsshownwere obtainedwith no training or calibration beforeoperating system. the To demonstratethe effectiveness the systemon narof row-bandaswell asbroadbandnoise,an electronictonegenerator was usedwith a loudspeaker introducea tone at to about70 Hz at the fan while the HVAC systemwasoperating. The noisereduction this combined on broadband fan
noise and narrow-band electronic tone is shown in Fig. 6. The tone is reducedabout 30 dB, while the broadband noise

40

-1

HZ

200

is reducedabout 15 dB. It is importantto note alsothat, in additionto the attenuation shownin Figs. 5(c) and 6, there
801 J. Acoust.Soc. Am., Vol. 85, No. 2, February 1989

FIG. 6. Noisereductionfor an input signalconsisting broadband of noise from a centrifugalfan combinedwith a tone generated a speaker by (M = 0.04-128 averages). 801

L.J. Erikssonand M. C. Allie: Random noisefor on-linemodeling

adaptive filterbased theRLMS algorithm. effects on The of


sound source errorpathtransfer and functions adaptiveare ly determined linethrough useof a second on the LMS algorithm that usesan independent low-level random noise sourceto modelthe soundsourceand error path while the system operating. is The combined system fully adaptive, is compensates changes all transducers, source, for in the and
acousticalelements,is effectiveon broadbandas well asnar-

7p.L. Feintuch, adaptive "An recursive filter,"Proc. LMS IEEE 64 ( 11),


1622-1624 (1976).

8B.Widrow,"Adaptive filters," Aspects Network System in of and Theory,


editedby R. E. Kalman and N. Declaris(Holt, Rinehart,and Winston,
New York, 1971).

9D.R. Morgan, "Analysis multiple of correlation cancellation witha loop


filter in the auxiliarypath," IEEE Trans.Acoust.Speech SignalProcess.
ASSP-28 (4), 454-467 (1980).

toj.C.Burgess, "Active adaptive sound control aduct: computer in A simulation," J. Acoust. Soc. Am. 70, 715-726 (1981).

row-bandnoise,and requires calibrationor trainingprono

cedures prior to operation.


ACKNOWLEDGMENTS

B. Widrow, D. Shur,and S. Shaffer, "On adaptive inverse control,"in Proceedings the 15th ,4silomar of Conference Circuits, on Systems and Computers, Nov. 1981,Pacific 9-11 Grove,CA, pp. 185-189.

2B. Widrow S.D. Stearns, and 4daptiue Signal Processing (Prentice-Hall,


Englewood Cliffs,NJ, 1985), p. 288.

The authorsgratefullyacknowledge assistance the of Cary Bremigan, James Gilbert, andPatriciaSteaffens the of NelsonIndustries,Inc. CorporateResearch Department.

3L.A. Poole, E. Warnaka, R. C. Cutter,"The implementation G. and of digitalfiltersusinga modified Widrow-Hoffalgorithm the adaptive for
cancellation acoustic of noise,"Proc. IEEE ICASSP 84, SanDiego, Vol. 2, pp. 21.7.1-21.7.4.

14]]. Widrow,J. M. McCool,and B. P. Medoff,"Adaptivecontrolby inverse modeling," Proceedings the12thAsilomar in of Conference Ciron cuits,Systems, Computers, November1978,PacificGrove, CA, and 6-8
pp. 90-94. SL. J. Eriksson,"Active attenuationsystemwith on-line modellingof speaker, errorpath,andfeedback path,"U.S. PatentNo. 4,677,676(30
June 1987).

L. J. Eriksson, "Activesound attenuation usingadaptive digitalsignal processing techniques," Ph.D. thesis, University Wisconsin--Madison of
(August 1985).

2L.J. Eriksson, C. AIlie,andR. A. Greiner, M. "Theselection applicaand tion of an IIR adaptive filter for usein activesoundattention,"IEEE Trans.Acoust. Speech Signal Process. ASSP-35(4), 433-437 (1987). 3L.J. Eriksson M. C. Allie, "A practical and system active for attenuation
in ducts," Sound Vib. 22(2), 30-34 (1988).

6M. R. Schroeder, Number Theoryin Science and Communications


(Springer, Berlin, 1984).

t7M.C. Allie,C. D. Bremigan, J. Eriksson, R. A. Greiner, L. and "Hardware and software considerationsfor active noise control," Proc. IEEE

ICASSP 88, New York, Vol. V, PaperA3.6, pp. 2598-2601.

4L. J. Eriksson, C. Allie, C. D. Bremigan, M. and R. A. Greiner,"Active noisecontrol usingadaptivedigital signalprocessing," Proc. IEEE ICASSP 88, New York, Vol. V, PaperA3.5, pp. 259--2597.

Digisonix Division, Nelson Industries, "Digisonix Inc., digital sound cancellation systems activenoise for control,"Pub.No. DX-DS-1/987.

9L.J. Eriksson M. C. Allie,"A digital and sound control system use for in
turbulent flows," Proceedings NOISE-CON 87, State College, PA, pp.
365-370.

SL.J. Eriksson M. C. Allie,"System and considerations adaptive for modelling applied active to noise control," Proc.IEEE ISCAS88,Espoo, Finland, Vol. 3, pp. 2387-2390. 6L. J. Eriksson, "Active soundattenuation system with on-lineadaptive
feedbackcancellation," U.S. Patent No. 4,677,677 (30 June 1987).

:L.J.Eriksson, C. Allie,C. D. Bremigan, J.A. Gilbert,"Theuse M. and of


activenoise controlfor industrialfan noise,"presented the 1988ASME at Winter Annual Meeting,Chicago, Paper88-WA/NCA-4. IL,

802

J. Acoust.Soc. Am., Vol. 85, No. 2, February1989

L.J. Eriksson and M. C. Allie:Randomnoisefor on-linemodeling

802

Potrebbero piacerti anche