VeteransAdministration

79

Journal of Rehabilitation Researchand Development, Vol . 23, No . 1, 1986pages 79-87

A Digital Master ~~~~~~~~~~~~ ~~^r~" " ~~ " x~ " ~~~ "~~~^"~°" Hearing Aid

Abstract -- The use of computer simulation in evaluating conven-tional and experimental hearing aids is described . Two illustrativeexamples are provided . The first involves the simulation of acon-ventional master hearing aid and its application in evaluatingdifferent adaptive strategies in the prescriptive fitting of hearingaids. The second example involves the simulation of an expori-nnoniel hearing aid embodying modern digital signal-processingtechniques for the reduction of background noise . A high-speedarray processor is used in order to accomplish these simulationsin real time.

INTRODUCTION

There are three basic approaches to the design of digital hearingaids . The first is to design instruments that essentially duplicatethe operation of conventional (analog) hearing aids . The secondapproach is to develop digital hearing aids that embody the samebasic principles as analog instruments, but make use of digitaltechniques to incorporate features that would have been difficultand/or innpnaotioal to achieve with conventional analog circuits.The third approach is to develop digital instruments that are con-ceptually very different from conventional analog hearing aids andwhich embody characteristics that, fora!! practical purposes, canbe achieved only by advanced digital signal processing techniques.

An example of the first approach is a digital hearing aid consist-ing of a conventional hearing-aid microphone and preamplifierfo!!ov0ed by an ana!oA'to'digital oonverter, ad!Aite! fi!ter, a digital-to-analog converter and a power amplifier driving a conventionalhearing aid receiver . Prototype units of this type have already beendeveloped ; see Levitt and Sullivan, 1984 (1).

An example of the second approach is edigita! hearing aid thatcan be programmed to best meet the needs of individual users . Afairly sophisticated digital filter may be required in order to obtainthe optimum frequency-gain characteristic for each individual . Al-though programmable analog hearing aids have already been de-veloped -- see Mangold and Leijon, 197S(2) — there are limitationsto the range and flexibility of the analog filters that can be used ina practical system. Digital filters have fewer limitations . Amongthe important advantage of programmable instruments is that theyallow a unified treatment of both the design and fitting of hearingaids: see Engebretson etal ., 1983 (3), and Pope!kaand Engebretson,1883(4).

The third approach offers the greatest potential for significantimprovements in aiding the hearing impaired . Innovative methodsof aignal processing that are currently being explored include

HARRY LEvnT n~D . bARLENE NEUMAN, Ph . D.RUSSELL MILLS, M.S.TERESASCHVVANDER .M.Ed.

Center for Research inSpeech & Hearing SciencesCity University of New York a33 West 42nd StreetNew York, N .Y . 10036

'The research reported hereported by Grant P01 NS17764 from theNational Institute of Neurological andCommunicative Disorders and Stroke(NINCDS), and by Grant No .6008302511to the Lexington Center from the Na-tional Institute of- Handicapped Re-search (NIHR).

bAddress all correspondence to : HarryLevitt, Ph . D ., CUNY Graduate Schooland University Center, 33 West 42ndStreet, New York, N .Y . 10036 .

80

LEVITT et al . : A Digital Master Hearing Aid

frequency lowering, enhancement of perceptuallyimportant speech characteristics, and automaticnoise reduction : e .g ., Levitt, Pickett and Houde, 1980(5) . Experimental systems for investigating differentforms of digital speech processing for the hearingimpaired are already being developed in several re-search laboratories and it is only a matter of timebefore experimental hearing aids embodying ad-vanced signal-processing techniques are developed.

In this context, with the application of advancedtechniques a subject of wide and intense currentinton*et, it is surprising that relatively little attentionseems to have been directed toward the use of acomputer tosimulate hearing aids. Computer simu-lation is particularly well suited for the evaluationof innovative experimental hearing aids ; e .g . . digitalhearing aids incorporating advanced aignal proc-essing techniques . Computer simulation may alsobe used to advantage in the simulation of conven-tional hearing aids for clinical applications, such asthe prescriptive fitting of hearing aids using a mas-ter hearing aid, as described in Levitt in 1978 (6).

The purpose of this paper is to describe the applica-tion of computer-simulation techniques to thedevelopment of an extremely flexible, general-purpose master hearing aid . Two illustrative applica-tions of the system are described. The first exampleinvolves the simulation of a conventional masterhearing aid, but with the added feature that thesystem is computer controlled, thereby greatlyfacilitating the adjustment process . The secondexample involves the simulation of an experimentalhearing aid in which modern digital signal-process-ing techniques are used to suppress backgroundnoise.

TECHNICAL DESCRIPTION

A block diagram of the basic system is shown inFigure 1 . The simple diagram shows only the majorcomponents; actually the system is extremely flex-ible andaUovvomanyvariediona.

The input signals (typically but not necessarilyaudio signals) may be from a microphone, taperecorder, or FM receiver. The anti-aliasing filterlimits the bandwidth of the signals processed by thesystem to a value compatible with the samplingrate of the digital system . This sampling rate isunder computer control . Although a sampling rateas high as 125 kHz can be used, the time requiredto process signals in real time sets practical limitson the sampling rate . In general, the lower thesampling rate, the greater the amount of signal

processing that achieved . A useful andpractical compromise for this application has beena sampling rate of 12 kHz. A bandwidth of 6 kHz istheoretically usable with this sampling rate, but inpractice a bandwidth of just over 5 kHz is used inorder to allow for practical constraints such as roll-off in the anti-aliasing filters. This bandwidth isconsidered to be adequate for most hearing-aidsimulations . In special applications, such as simu-lating "wideband" hearing aids, a sampling rate ashigh as 25 kHz can be used, but then the amount ofsignal processing is limited to little more than fre-quency shaping.

The signals from the antialiasing filter are digitizedand fed into one or more storage buffers . A 12-bitanalog-to-digital converter is used . A multiplexer isemployed when simulations involve more than oneinput channel, as in the case of a binaural hearingaid.

The heart of the system is a MAP-300 array pro-cessor. This device allows high-speed vector opera-tions and is well suited to performing extremelyrapid Fourier transformations and convolutions . Forexample, a 1024-point fast Fourier transform (FFT),corresponding to 85 .3 msec of speech at a samplingrate of 12 kHz, can be performed in less than 10msecs. This capability allows much of the signalprocessing to be performed in the frequency domain,with many attendant advantages.

The array processor is controlled by a minicom-puter (DEC LSI-11/23) . The controlling computerfeeds programs to the array processor and in thisway controls what the array processor does . Thecontrolling computer can also be used to monitorthe subject's responses, analyze data, check theoutput of the array processor (for self-calibrationand checking) and for the running of experiments.In the latter application, the controlling computercan be used to implement an adaptive test strategywhereby, depending on the subject's responses,the simulated hearing aid converges on the optimumparameter values for that individual user.

The limitations of the system are determinedprimarily by the operating time of the array processorand its high-speed memory. The current system(which can be expanded and speeded up by addingmore high-speed memory) can perform at least eightFFT's per time window at a sampling rate of 12 kHz,and can store a filter impulse response of about200 msec for realtirne operation . For non-realtimeapplications, both the number of FFT's that can beimplemented per time window, and the length ofthe system impulse response, are essentiallyunlimited .

81

Journal of Rehabilitation RoaeamhandDovo!opmon/ Vol . oawo . 1 1980

~

~

MULTIPLEXPROCESSOR

o — x+ -

MULTIPLEX

CONTROLLING

COMPUTER _

_

-4 HA REC.

HEADPHONE

-0 FM

INPUT_

_

Awn-xuxmmoFILTERS

_

ANTI-IMAGINGFILTERS

- x«REC.

-0 HEADPHONE

-0 FM~

_

_

mi c.~

_

'~.

FIGURE 1Components of Computerized Syste

Note that in the realtime mode of operation, whenusing discrete time-windows, there is an inherenttime delay between input and output equal to theduration of at least two time windows . If recursivefiltering without buffering is used, the delay is re-duced to that of the sampling period, e .g ., roughly80 microseconds at a 12-kHz sampling rate . In itscurrent mode of operation, the discrete time-windowapproach is used . For the applications used todate, 256-point FFT's have been employed resultingin input/output delays of a little over 40 msec at the12-kHz sampling rate.

The output of the array processor is stored tem-porarily in buffers, after which the signals are con-verted to analog form by a 12-bit digital-to-analogconverter . A multiplexer is used for multichannelapplications . In order to avoid the problem of tran-sients between successive time windows, anoverlap-add procedure is used in which two con-current time waveforms using overlapping timewindows are generated and then added, as describedby Allen in 1979 (7).

The output of the analog-to-digital converter isfed to an appropriate acoustic delivery system (e .g .,headphones, hearing-aid receiver, FM transmitter).A major concern regarding the clinical use of adesktop or rack-mounted master hearing aid is thatthe acoustic characteristics of a wearable personal

hearing aid are inherently different from that of anon-wearable instrument . Not only are head andbody baffle effects difficult to take into account,but there appears to be a complex interaction be-tween head movements, which can be exquisitelyfine, and associated auditory perceptions . In orderto circumvent differences between wearable andnonwearable hearing aids, an FM signal transmis-sion may be used because this can make the simu-lated hearing aid acoustically identical to that ofewearable personal hearing aid.

For applications requiring the simulation of wear-able personal hearing aids, a conventional hearing-aidnniorophnneandnaoeivorhovnboannnounLeUina standard casing; thus far, a standard postauricularhearing-aid case has been used . The same approachcan also be used for simulating an in-the-ear hear-ing aid that requires a custom-made mold .or thepostauricular simulation, the output of the hearing-aid receiver is led to a custom earmold using stan-dard acoustic tubing.)

In a typical simulation, the electrical signalgenerated by the subject's wearable microphone ispre-amplified and led a pocket-size FM transmit-ter vvhooetranann!tLodaignal is picked up by an FMreceiver located at the input to the computer . Theoutput of the computer is transmitted back to thesubject to be picked up by a pocketsize FM receiver/

82

LEVITT et al . : A Digital Master Hearing Aid

amplifier, the output of which is fed directly to thehearing-aid receiver.

For all practical purposes, the acoustic inputand output stages of such a simulated hearing aidare identical to those of a conventional personalhearing aid . However, the electronic processing ofthe signals that takes place between microphoneand output transducer has been transferred to thecomputer and back by means of the FM transmissionsystem . (It is important to note that two independenttransmitter-receiver systems, with non-interferingcarriers, are used so that the hearing aid can trans-mit undnaoeiveaithoaanneUnne .)

EXAMPLES OF SIMULATION

1 . Conventional Master Hearing Aid

The potential of the coml , uter-controlled hearingaid simulation system as a research and clinicaltool is best illustrated by means of examples.

In the first example, the system is used to simulatea conventional master hearing aid under computercontrol . In addition to providing frequency-selectiveamplification, the system has been programmed togenerate such test stimuli as tones or bands of noise,as needed . In this application the system serves asboth a hearing aid and an audiometer.

A block diagram of the simulated master hearingaid (Fig . 2) shows that the input from each channel

of this two-channel system is multiplexed and fedinto one of four buffers, two buffers being used foreach channel . The method of simulation uses theoverlap-add procedure developed by Allen (7).

This system is in use in an experiment comparingdifferent adaptive strategies far adjusting thefrequency-gain characteristic of a master hearingaid . A general approach to the prescriptive fittingof frequency-gain characteristics is to begin with afirst estimate of the optimum frequency-gain char-acteristic based on psychoacoustic considerations,followed by systematic adjustment of these char-acteristics to arrive at an improved estimate (6).Typically, a listener's judgment of relative speechintelligibility is used as the criterion for an"improved" estimate, although other criteria suchas relative speech quality may also be used . Theadjustment procedure is continued iteratively untilno further improvements in the prescribed criterioncan be obtained.

The specific implementation of this general pro-cedure began in the current experiment with a firstestimate based on the method developed by Pascoe(8) . According to that technique the subject's thres-hold, loudness discomfort level, and several inter-mediate levels (very soft, soft, comfortable, loudand very loud) are obtained using third-octavebands of noise . The test stimuli are generated andcontrolled through the computer, using an adaptive

INPUT BUFFERS

OUTPUT BUFFERS

CHANNEL , .#,

mSPECTRUM SHAPING

GENERALIZED

mULnBANo

COMPRESSION

SWITCHINGCONTROL

LSI 11/23

FFT

INPUT ,

AID

FIGURE 2Block diagram of master hearing aid .

83

o Rehabilitat on Research and Deve opment Vol . 23 No . 1986

up-down strategy to converge efficiently on thelevels indicated . The first estimate of the optimumfrequency-gain characteristic is obtained by findingthe frequency-gain curve that places the rms levelof the speech stimulus at the average comfort levelover the frequency range of interest, 200 to 5000 Hz.Subsequent estimates of the optimum frequency-gain characteristic are then obtained by systemati-cally adjusting this frequency-gain characteristicand testing for improved intelligibility.

Methods of frequency-gain adjustment compared—An

performed to comparedifferent methods of adjustment . One method, re-ferred to as the "round-robin'' procedure, was usedin earlier investigations with a manually controlledmaster hearing aid and was described by Sullivanet al . in 1983 (9) . A 3 x 3 matrix was formed with thecentral cell of the matrix corresponding to the firstestimate . The rows of the matrix correspond tochanges in low-frequency gain, the columns corre-spond to changes in high-frequency gain . Paired-comparison judgments of relative intelligibilitywere obtained for all possible pairs of cells in thematrix . The frequency-gain characteristic of thecell showing the largest number of "wins" in thepaired comparisons became the second estimateof the optimum frequency-gain characteristic. Thenthe round-robin strategy was repeated using thissecond estimate as the central cell . Unless contra-indicated by the data, it was usually possible toterminate testing after the second or third estimateof the frequency-gain characteristic had beenobtained.

The second adaptive procedure used a double-elimination paired-comparison strategy of the typedescribed by Montgomery in 1982 (10) ; this pro-cedure is referred to as the "tournament" strategy.A 4 x 4 matrix of cells was formed by adding arow and column to the original 3 x3 matrix . Asbefore, rows of the matrix correspond to changesin low-frequency gain and columns correspondto changes in high-frequency gain . The tournamentwas then performed on all of the cells in the matrixusing paired-comparison judgments of relativeintelligibility . The frequency-gain characteristiccorresponding to the winning cell became thesecond and final estimate of the optimum frequency-gain characteristic.

The third adaptive strategy used a variation ofthe simplex procedure (6) and is referred to as the"modified simplex" method . A two-way matrix ofpossible cells was formed using the same basicdesign as for the round-robin and tournament

LOUDNESS LEVELS IN THE HIGH FREOUENCIES

1

2

3

4

5

LOUDNESS LEVELSo=AVERAGE COMFORT

1 HIGHEST VERY SOFT

*=HIGHEST COMFORTu=AVERAGE SOFT

w=AVERAGE LOUD

FIGURE 3Typical data for comparison among adaptive strategies

strategies, i .e ., rows correspond to changes inlow-frequency gain, and columns to changes inhigh-frequency gain . An elemental L-shaped set ofthree adjacent cells was formed with the vertexcorresponding to the first estimate of the optimumfrequency-gain characteristic. Paired comparisonsof relative intelligibility were then obtained betweenthe vertex and the two adjacent cells . Dependingon the results of these paired comparisons, a newL-shaped set was formed in an adjacent regionaway from less intelligible cells of the old L-shapedset . A sequence of L-shaped sets thus generatedshould converge on an improved estimate of theoptimum frequency-gain characteristic . The criterionfor convergence was three reversals in the directionof movement along each axis (i .e ., across rows andcolumns of the large matrix of possible cells).

The results obtained for two typical subjects areshown in Figure 3 . The diagram shows the set ofpossible cells centered on the first estimate of theoptimum frequency-gain characteristic . The symbolsA and B identify the two subjects ; the suffixes R, T,S identify the round-robin (R), tournament (1) andmodified simplex (S) strategies, respectively.

The location of each symbol identifies the cellcorresponding to the final estimate of the optimumfrequency-gain characteristic . In this case, for each

1

2

3

4

5

84

L*vDT 'g al . : A Digital Master Hearing Aid

subject, all three strategies converged on the samecell . However, preliminary results on additionalsubjects indicate that this pattern does not neces-sarily hold, although the differences in the estimatesobtained by different adjustment strategies arerelatively small . Differences between estimates areto be expected when the optimum condition beingsought is represented by a relatively shallow peakor ridge in the response surface.

An important difference between the two subjectsis that for one subject, A, the final estimate of theoptimum frequency-gain characteristic is the sameas the iniUal estimate (i .e ., the central oe!l in Figure3 corresponds to the first estimate) . For the secondsubject, B .thecell corresponding to the final esti-mate of the optimum frequency-gain characteristicdiffered from the initial estimate. (Subject B alsoconsistently judged speech to be more intelligiblefor the final estimate of the optimum frequeney-gaincharacteristic than for the initial estimate .)

All of the paired-comparison judgments of relativeintelligibility were obtained by having the listenerjudge the intelligibility of a continuous discoursepassage presented against a background of cafete-ria noise. This is an extremely efficient method oftenting, but the judgments are necessarily subjec-tive . However, when a final check was performedusing a more time-consuming objective measure ofspeech-recognition performance (NorthwesternUniversity Auditory Test #6), the scores supportedthe results obtained with themore efficient paired-comparison teohnique.

The relative efficiency of the three methods ofadjustment is shown in Table 1 . The round-robinstrategy, which can be implemented manually with-out too much difficulty, was the least efficient ofthe three procedures both in terms of time takenand the average number of trials required . The

TABLE 1Comparison Between Adaptive Strategies*

Subject A

Subject B

Round-Robin

Number of trials 216 216Time in minutes 97 125

Tournament

Number of trials 93 93Time in minutes 39 49

Modified Simplex

Number of trials 18 15Time in minutes 7 7

*All data are for three replications per paired comparison .

most efficient strategy was the modified simplexprocedure, which converged on the estimatedoptimum frequency-gain characteristic in a fractionof the time of the slower round-robin procedure.The modified simplex is a difficult technique toadminister manually, however, and some deOree ofcomputerization is needed for its implementation.(Although it is not as efficient as the modifiedsimplex procedure, the tournament strategy is alsoa relatively efficient technique, and it is also a rela-tively robust technique which can, if necessary, beadministered manually using a conventional masterhearing aid .)

2 . Simulating background-noisecomputer system has also been used to simulatean experimental hearing aid that would be exceed-ingly difficult, if not impossible, to construct inanalog form . The design goal is a hearing aid thatwill automatically reduce background noise . Back-ground noise is particularly damaging to speechintelligibility for the hearing impaired, and systemsfor reducing background noise are thus of greatinterest.

The method of noise reduction that has beenimplemented is that developed by Weiss andAschkenasy (11). Their method was chosen becauseof its potential power and also because it hasachieved some practical success; it is currentlyused for specialized applications in which normal-hearing listeners are required to monitor speech-in-noise for long periods of time.

The input and output stages of the system areessentially the same as those shown in Figure 2,except that the system is monaural and that a tri-angular weighting function is applied to each inputbuffer. The central part of the system is shown inexpanded form in Figure 4: after the first fast Fouriertransform (FFT), the phase information is stored forlater use while the square root of the amplitudespectrum is converted back to the time domain byan inverse FFT.

An estimate of the noise component in thiscepstrum-like signal is obtained by time-weightedaveraging over successive windows. This trans-formed noise spectrum is subtracted from thetransformed speech-plus-noise spectrum . The resultis an enhanced representation of the signal inwhich the noise content is greatly reduced. Byinverse transformation, an enhanced amplitudespectrum is obtained which, with the addition ofphase information, is converted back to conventionaltime-dependent signals . The output of the final FFTis led to the output buffers, as shown in Figure 2.

85

Journal of Rehabilitation Research and Development Vol . 23 No . 11988

FIGURE 4Block diagram of noise reduction hearing aid .

SUBTRACT

NOISECOMPONENT

FFT

RESTORE

PHASE

INFO.FFT

FFT

ESTIMATE

NOISECOMPONENT

STORE

PHASE/wFo.

FFT

mumcAMPL.

SPECTRU

As before, the overlap-add technique is used toreduce transients.

A pilot study has been performed to evaluate theeffectiveness of the noise-reduction process fornormal-hearing listeners, in preparation for a studywith hearing-impaired listeners . Ten normal-hearinglisteners served as subjects . Hearing levels werewithin 15 dB of normal audiometric zero for allsubjects . The subjects ranged in age from 16 to 35years.

Ten subsets of the City University of New YorkNonsense Syllable Test (NST), as described in 1978by Levitt and Resnick (12), were chosen as the teststimulus for measurement of speech recognition.The NST consists of nonsense syllables (VC or CVconstruction) recorded by a male talker in the carrierphrase "You will mark —, please ." Response foilsinclude nonsense syllables which differ in mannerandlor place of articulation, but not in voicing . Thesubject marks on his answer sheet which of theseven-to-nine possible syllables has been presented.The 10 subsets include voiced and voiceless con-sonants in final position in combination with the

initial position with the vowel l a / . The nonsensesyllable test is well suited to a repeated-measuresexperiment; the overall score on the test is highlyreliable, as is the pattern of errors made by a givensubject ; see Dubno and Dirks, 1982, (13), and Dubno,Dirks, and Langhofer, 1982 (1 4).

Subjects were first familiarized with the format

of the NGT. Performance on the NST was thenevaluated under three experimental conditions (i)the unprocessed speech-in-noise (+5 dB SIN), (ii)the processed signal, and (iii) unprocessed speech-in-noise with f17 dB SIN at the input. The latterwas included as a reference condition, since thenoise-reduction process increases the signal-to-noise ratio by approximately 12 dB . All testing wasdone at the subject's most-comfortable listeninglevel (MCL). Presentation of the three experimentalconditions was randomized.

Syllable-recognition scores on the nonsense-syllable test were calculated in pvopodiono, con-verted to arcsine units, and a one-way analysis ofvariance was performed to determine the effect ofexperimental condition . The results of the analysisrevealed that the factor of condition was highlysignificant [F (2,18) = 178 .6, p< .001] . The Scheffemultiple-range test (used to test for differencesbetween the three mean scores) revealed that onlyone of the means differed significantly from theothers, that of the +17 dB S/N condition (p< .01).The mean scores for the +5 dB SIN unprocessedand for the processed conditions did not differsignificantly . The failure of processing to improvesyllable-recognition scores despite measurableimprovements in signal-to-noise ratio, is in agree-ment with the findings of other investigators whohave used digital noise reduction techniques withdifferent speech materials and different types ofnoise: i .e ., Lim and Oppenheim in 1979 (15) .

86

LEVITT et al . : A Digital Master Hearing Aid

While the overall scores on the nonsense syllabletest did not change significantly as a function ofprocessing, the pattern of errors did change and, infact, showed certain improvements in perceptionafter processing . For instance, place/manner con-fusions decreased with processing for the subsetcontaining voiced final consonants following / u I.For voiced final consonants in combination withthe vovvn! /i/ . place and manner errors both de-creased with processing . These improvements inthe transmission of certain phonetic features areencouraging, since this may enhance the intelligi-bility of speech in context . Continued work isplanned to evaluate the effectiveness of this typeof processing for hearing-impaired listeners.

DISCUSSION

The preceding two examples show the power ofcomputer simulation in two very different applica-tions . The first example was clinically oriented andshowed how computer simulation could be used tofacilitate the prescriptive fitting of conventionalhearing aids. The advantages of the computersystem in this application are twofold . Because thehearing aids are simulated, it is not necessary todesign and construct an experimental master hear-ing aid system. Further, the range and versatility ofthe computerized system far exceeds that whichcould be obtained with conventional master hearingaids. Secondly, the system is computer controlled,greatly simplifying the task of adjusting the masterhearing aid . As a result, extremely efficient adaptivepaired-comparison strategies can be used to con-verge on the estimated optimum electroacousticcharacteristics.

In the example given, only the frequency-gaincharacteristic was adjusted, this being the thrustof much recent research. Using the method of com-puter it is a relatively simple matter tomanipulate other electroacoustic characteristicssuch as frequency-dependent amplitude limiting,compression amplification, and other forms ofacoustic processing.

The second illustrative example shows the levelof sophistication that can be reached in real-timecomputer simulation of experimental hearing aids.In this case, the design and construction of aprototype noise-reduction unit would be a formidabletask, and further changes in the design of the proto-type which might follow from the results of theexperimental evaluation would be likely to involveaubotenUal additional effort in design and con-struction .

The programming ofsystem, although considerably less difficult thanbuilding hardware prototypes, is not a simplematter. By analogy, economy of programming effortcan be achieved by developing the software fornon-realtime simulations . This type of programmingis not difficult, particularly if a high-level languagegeared to simulation of speech processing systemsis available . If the non-realtime simulation appearspromising, the decision may then be made to investthe additional effort needed to get the system torun in real time.

Although the experimental results obtained withthe computer simulation of noise suppression weremixed (Le ., subjects preferred the processed signalsbut objective measures of intelligibility were lower),there is good reason to continue research in thearea. Almost all of the previous research on noisesuppression has been geared to the needs of normal-hearing listeners, according to Lim and Oppenheim,1979 (15) . Since the detrimental effect of noise onspeech is greater for the hearing impaired than fornormal-hearing listeners, any noise reduction thatcan be achieved is likely to be of greater benefit tothe hearing impaired

Acknowledgments

The authors would like to thank Mark Weiss andErnest Aschkenasy for their valuable cooperationin the implementation of their noise-reductionprocess . We wish also to thank Ms . Sharon Axelrodand Mr. Joseph Isaacs for their assistance in settingup the equipment and running the experiments andMs. Marilyn Halprin for her help in preparing themanuscript.

The authors are also grateful to the Com-TekCompany for their donation of the FM transmissionsystem and to Knowles, Inc ., and Precision Acous-tics, Inc., for their donation of hearing-aid trans-ducers .

REFERENCES

1. Levitt H, Sullivan JA: Developmentof New-Generation Hearing Aids.Annual Progress Report, Rehabilita-tion Engineering Center, LexingtonCenter, Inc., Jackson Heights, NewYork 11372, 1984.

2. Mangold S, Leijon A : Programmablehearing aid with multichannel com-pression . Scand Audiol 8 :121-126,1979.

3. G . A unified digital hearing-aid designand fitting procedure . ASHA 25 :101(abstract) 1983 .

87

Journal of Rehabilitation Research and Development Vol . 23 No . 1 1986

4. Popelka GR, Engebretson AM : A com-puter-based system for hearing aidassessment . Hearing Instruments34 :6-9, 44, 1983.

5. Levitt H, Pickett JM, Houde RA : Sen-sory Aids for the Hearing Impaired.New York: IEEE Press, 1980.

6. Levitt H : Methods for the evaluationof hearing aids, and adaptive testingin audiology, In Sensorineural Hear-ing Impairment and Hearing Aids.C . Ludvigsen and J . Barfod, Eds .,Scand Audiol, Suppl 6 :199-239, 241-291 . 1978.

7. Allen JB : FASTFILT—An FFT-BasedFiltering Program . In Programs forDigital Processing, edited by theDigital Signal Processing Committee,IEEE Acoustics Speech and SignalProcessing Society. New York: IEEEPress, 1979.

8. Pascoe DP: An approach to head ng-aid selection . Hearing Instruments29 :12-16, 36-37, 1978.

9. Sullivan JA, Neuman AC, Levitt H:Adaptive estimation of frequency-gain characteristics for binauralhearing aids . ASHA convention(abstract), 1983.

10. Montgomery A, Schwartz D, Punch J:Tournament strategies in hearingaid selection . J Speech Hear Disord47 :363-372, 1982.

11. Weiss M, Aschkenasy E : WidebandSpeech Enhancement, RADC-TR-81-53 Final Technical Report, May 1981.

12. Levitt H, Resnick SB : Speech recep-tion by the hearing impaired : Meth-ods of testing and the developmentof new tests . In Sensorineural Hear-ing Impairment and Hearing Aids, C.Ludvigsen and J . Barfod, Eds., ScandAudiol Suppl 6 :107-130, 1978.

13. Dubno JR, Dirks DD : Evaluation ofhearing-impaired listeners using anonsense-syllable test ; I . Test relia-bility. J Speech Hear Res 25 :135-141,1982.

14. Dubno JR, Dirks DD, Langhofer LR:Evaluation of hearing-impaired listen-ers using a nonsense-syllable test;II . Syllable recognition and conson-ant confusion patterns. J SpeechHear Res 25 :141-148, 1982.

15. Lim, JS, Oppenheim AV : Enchance-ment and bandwidth compression ofnoisy speech. Proc . IEEE, 67:1586-1604, 1979.