Effect of Training on Word-Recognition …audres/Publications/humes/papers/132...Effect of Training...

Effect of Training on Word-Recognition Performancein Noise for Young Normal-Hearing and Older

Hearing-Impaired ListenersMatthew H. Burk, Larry E. Humes, Nathan E. Amos, and Lauren E. Strauser

Objective: The objective of this study was to evalu-ate the effectiveness of a training program for hear-ing-impaired listeners to improve their speech-rec-ognition performance within a background noisewhen listening to amplified speech. Both noise-masked young normal-hearing listeners, used tomodel the performance of elderly hearing-impairedlisteners, and a group of elderly hearing-impairedlisteners participated in the study. Of particularinterest was whether training on an isolated wordlist presented by a standardized talker can general-ize to everyday speech communication across noveltalkers.

Design: Word-recognition performance was mea-sured for both young normal-hearing (n � 16) andolder hearing-impaired (n � 7) adults. Listenerswere trained on a set of 75 monosyllabic wordsspoken by a single female talker over a 9- to 14-dayperiod. Performance for the familiar (trained)talker was measured before and after training inboth open-set and closed-set response conditions.Performance on the trained words of the familiartalker were then compared with those same wordsspoken by three novel talkers and to performanceon a second set of untrained words presented byboth the familiar and unfamiliar talkers. The hear-ing-impaired listeners returned 6 mo after theirinitial training to examine retention of the trainedwords as well as their ability to transfer any knowl-edge gained from word training to sentences con-taining both trained and untrained words.

Results: Both young normal-hearing and older hear-ing-impaired listeners performed significantly bet-ter on the word list in which they were trainedversus a second untrained list presented by thesame talker. Improvements on the untrained wordswere small but significant, indicating some gener-alization to novel words. The large increase in per-formance on the trained words, however, was main-tained across novel talkers, pointing to thelistener’s greater focus on lexical memorization ofthe words rather than a focus on talker-specificacoustic characteristics. On return in 6 mo, listen-ers performed significantly better on the trainedwords relative to their initial baseline performance.Although the listeners performed significantly better

on trained versus untrained words in isolation, oncethe trained words were embedded in sentences, noimprovement in recognition over untrained wordswithin the same sentences was shown.

Conclusions: Older hearing-impaired listeners wereable to significantly improve their word-recogni-tion abilities through training with one talker and tothe same degree as young normal-hearing listeners.The improved performance was maintained acrosstalkers and across time. This might imply that train-ing a listener using a standardized list and talker maystill provide benefit when these same words are pre-sented by novel talkers outside the clinic. However,training on isolated words was not sufficient to trans-fer to fluent speech for the specific sentence materialsused within this study. Further investigation isneeded regarding approaches to improve a hearingaid user’s speech understanding in everyday commu-nication situations.

(Ear & Hearing 2006;27;263–278)

Elderly hearing-impaired listeners often have dif-ficulty understanding speech within a backgroundnoise, even with amplification. These deficienciesmay be due in part to peripheral physiological prob-lems accompanying the sensorineural hearing loss(i.e., reduced frequency selectivity or temporal reso-lution), decreases in central or cognitive abilitiesassociated with aging, or some combination of thesefactors (Humes, 1996). Although hearing aids havebeen shown to increase speech understanding forelderly hearing-impaired listeners (Humes, Chris-tensen, Bess, & Hedley–Williams, 1997; Humes etal., 1999;Humes, Garner, Wilson, & Barlow, 2001;Larson et al., 2000), the associated benefit is oftennot what might be expected, given the increasedaudibility of the signal (Humes, 2002).

Is it realistic to expect older hearing aid wearers,especially new ones, to immediately perform as wellas they should, given the improved acoustics pro-vided by the hearing aids? On average, for example,the length of time from the initial signs of hearingloss to the point at which help is sought is commonly5 to 10 yr (Brooks, 1979; Smedley, 1990). Therefore,a new hearing aid user may not make optimal use ofauditory information made audible by a new pros-thetic device as a consequence of central or cognitive

Department of Speech and Hearing Sciences, Indiana University,Bloomington, Indiana.

0196/0202/06/2703-0263/0 • Ear & Hearing • Copyright © 2006 by Lippincott Williams & Wilkins • Printed in the U.S.A.

263

changes brought about by an impoverished acousticinput for many years. Although the audibility of aspeech signal can typically be restored to a sufficientdegree immediately, the listener may need time tolearn to interpret the newly restored information.This issue of training, or in essence retraining, thehearing aid user is generally overlooked within thehearing aid evaluation and rehabilitative processes.As will be discussed, training an individual withamplified speech may be needed to optimize thebenefits from a hearing aid.

The predominant factor contributing to an olderhearing-impaired listener’s difficulties with speechrecognition is the audibility of the speech signal(Humes, 1996). However, audibility alone may notaccount for all of a listener’s difficulties, particularlyin the case of a long-standing hearing loss in whichthe listener has not been exposed to many of theauditory cues present within the speech signal formany years. Prolonged high-frequency hearing lossmight predictably influence the listener’s ability tomake use of any newly restored high-frequencyinformation on the introduction of a hearing aid(e.g., Amos & Humes, 2001). A period of relearningmight be needed as the user is exposed to newauditory information. There is not strong support formost older hearing aid wearers’ ability to spontane-ously learn to make use of the new information overtime. This has generally been referred to in theaudiology literature as acclimatization of hearingaid benefit (Cox & Alexander, 1992; Gatehouse,1992). The consensus appears to be that such accli-matization effects are either small or nonexistent, atleast for study periods of up to 3 yr (e.g., Humes &Wilson, 2003; Humes, Wilson, Barlow, Garner, &Amos, 2002; Turner, Humes, Bentler, & Cox, 1996).

Acclimatization, however, typically is a conceptreserved for cases of spontaneous or self-taughtimprovements in speech understanding with in-creasing amounts of hearing aid usage. Training oraural rehabilitation programs, on the other hand,are designed to facilitate improvements in aidedspeech communication through a formal program.There has been some evidence that such trainingcan be effective in listeners with high-frequencysensorineural hearing loss like that found in mostolder adults (Walden, Erdman, Montgomery,Schwartz, & Prosek, 1981).

The training regimen of Walden et al. (1981) isgenerally representative of so-called analytic audi-tory training programs (Blamey & Alcantara, 1994).In such programs, listeners are trained first todiscriminate subtle differences in syllabic stimuli,with minimal contrasts, such as /of/ and /os/. Listen-ers may eventually proceed to training in the iden-tification or recognition of such low-contrast stimuli.

On the other end of the auditory training continuumare those approaches often identified as syntheticapproaches (Blamey & Alcantara, 1994). In thesynthetic approach, high-context materials are usedand the listener is trained to make use of thiscontext to better reconstruct the message from thefragments perceived.

Word-based auditory training would appear to liesomewhere between these two extremes. Wordsclearly have more context, even in isolation, thannonsense syllables. Yet, misperception of a singlephoneme comprising the target words can result inincorrect identification or classification, just as withnonsense syllables. The present study sought toexamine an approach to word-based auditory train-ing in elderly hearing-impaired adults listening toamplified speech. Moreover, the focus here was anattempt to improve the speech communication of thelisteners in the presence of background noise.

EXPERIMENT 1Answers to two questions were necessary before

examining the benefit of word-based auditory train-ing on everyday speech communication, typicallytaking place in a background of noise or competingspeech, in older hearing-impaired adults. First, canyoung normal-hearing listeners improve their word-recognition abilities for words presented in a back-ground of noise? To address this question, baselineopen- and closed-response performance were mea-sured both pre- and posttraining for two sets ofwords. Both sets of words were spoken by the sametalker, with the listeners only receiving training onone set. This allowed for a direct comparison of pre-and posttraining performance for a trained versusuntrained list of words presented by the same talkerunder identical listening conditions.

Second, does training, if observed, generalize tothose same trained words presented by novel talk-ers? Of key importance is whether any improvedperformance for the trained list is due to the listen-ers’ memorization of the lexical representation of thewords themselves or whether they were learning thetalking style or acoustic characteristics when thatword was spoken by the trained talker. The use ofboth a second set of untrained words spoken by thesame talker, as well as the trained words spoken bynovel talkers, allowed this distinction to be exam-ined. If the listeners were learning or focusing oncues specific to the talker, they might predictablyperform better on a new untrained list presented bythat same talker relative to novel talkers. Thismight inhibit any possible generalization to noveltalkers, making training with a single standardizedtalker unrealistic if improvements in everyday com-

264 EAR & HEARING / JUNE 2006

munication are the goal. However, if the learningaffected the lexical representation of the wordsthemselves, improvements in word-recognition abil-ities may also be evident when the same words arepresented by new talkers (including those outsidethe clinic). Therefore, the main focus of Experiment1 was to examine the effect of word-based auditorytraining on word-recognition abilities within a back-ground noise both across words and across talkersfor young normal-hearing listeners.

METHODS

Participants

A group of six male and ten female young normal-hearing (YNH) listeners (20 to 30 yr of age, mean �23.6 yr) participated in the study. Listeners wererecruited through general postings and e-mailswithin the Department of Speech and Hearing Sci-ences at Indiana University, Bloomington. All lis-teners passed an air-conduction screening at 20 dBHL (American National Standards Institute, 1996)at octave intervals from 250 through 8000 Hz. Lis-teners were paid $10 per session as well as a $35completion bonus on collection of all data. Duringthe training aspect of the study, listeners received abonus of $3 for each training block in which theirscore was higher than their initial closed-set base-line performance, for a possible bonus of $12 persession during training. Subjects were not informedof this incentive until after the baseline measureswere completed.

Stimuli

The monosyllabic words used in this study wereAB words (Boothroyd, 1995, 1999), as recorded onthe Q/Mass Speech Audiology, Volume 2 compactdisc (Massachusetts Eye and Ear Infirmary, 1991).The compact disc contained 15, 10-item isophonemicword lists of four different talkers, two women (F1and F2) and two men (M1 and M2), yielding thesame 150 words for each individual talker. This setof 150 words was then divided in half (words 1 to 75and words 76 to 150) to create Set A and Set B lists,respectively, for each talker. The original presenta-tions from the compact disc were converted to digitalwave files and the monosyllabic words were ex-tracted from the carrier phrase to create individualwave files for each word. The final result was digi-tized recordings of the individual Set A and Set Bwords for each of the four different talkers, or a totalof 600 files. The 600 wave files were then equated foraverage RMS using Cool Edit Pro (Syntrillium,Version 1.2a) software. Set A, spoken by FemaleTalker 1 (F1), was used for all training, as well as,

pre- and posttraining conditions. Set B words werealso administered pre- and posttraining; however,listeners were not trained by using this set of words.Set A, henceforth, will be considered the trainedwords, whereas Set B will be considered the un-trained words.

The background noise presented during all testconditions was extracted from the Computer-As-sisted Speech Perception Assessment (CASPA) soft-ware (Boothroyd, 1999), Version 2.2. This back-ground noise was selected because it was shaped tomatch the long-term spectrum of the first 50 wordsof the AB word-recognition materials.

All testing took place with the use of insertearphones (ER-3A, Etymotic Research) in a double-walled sound-treated room (Industrial AcousticsCompany) that met or exceeded ANSI guidelines forpermissible ambient noise (American NationalStandards Institute, 1999). The listeners wereseated in front of a computer monitor to receivefeedback during training and used a mouse to selectappropriate responses. The stimulus wave files werepresented digitally (16-bit D/A, 44.1 kHz samplingrate) via Matlab (The MathWorks, Version 6.5)running Tucker Davis Technologies (TDT) Sys-tem-II equipment. Channel 1 (speech) and channel 2(noise) were low-pass–filtered at 10 kHz (TDT FT5)and summed together (TDT SM3) for presentation tothe right ear only. The left ear was occluded duringtesting with the left insert earphone. Levels wereadjusted using an amplifier (Crown D-75) and aheadphone buffer (TDT HB6) to produce an overalllevel of 83 dB SPL for the speech and noise channelsas measured in an HA-2 2-cm3 coupler. An overallspeech-to-noise ratio (SNR) of 0 dB was used for alltest conditions in this experiment.

Procedure

Listeners took part in seven, 60-minute ses-sions over approximately a 2-week period. Nomore than one session per day was allowed with amaximum time between sessions of 3 days (gener-ally over a weekend). The specific conditions ad-ministered in each session are described in Table1. For the open-set conditions, the listeners wrotetheir responses on an answer sheet, whereas forthe closed-set conditions, all 75 words were pre-sented alphabetically in list form on the computerscreen from which the listener selected the appro-priate answer with the computer’s mouse. Presen-tation order was randomized for each listener andblock of trials during the closed-set testing andtraining. However, during the open-set blocks, asingle randomized order was used across all sub-jects. Although this previously generated random

EAR & HEARING, VOL. 27 NO. 3 265

order was later repeated while examining post-training open-set performance (in Session 5), itwas after several other truly random closed-setpresentations. Within Session 1, a short practicecondition containing 10 items was presented be-fore baseline closed-set testing to familiarize thelisteners with the computerized task. Sessions 2through 5 consisted of training for the Set A listspoken by F1. In addition, Session 5 also includedposttraining measures of performance for both thetrained (Set A) and untrained (Set B) lists spokenby F1. During training, the listeners were in-formed whether each response was correct orincorrect and, on an incorrect response, weregiven the correct answer orthographically. Noauditory feedback was provided. Sessions 6 and 7examined the ability to generalize training toseveral novel talkers (F2, M1, and M2) speakingboth the trained and untrained words. Half of thelisteners received the Set A and Set B wordsspoken by the novel talkers in a closed-set condi-tion in the posttraining sessions, whereas theother half received the same materials in anopen-set condition during the posttraining ses-sions. All sessions were self-paced by the partici-pants, requiring a key press to move on to the nextpresentation.

RESULTS AND DISCUSSION

Figure 1 presents the individual (top) and group(bottom) percent correct scores from the pretraining(far left) through post-train sessions. As can be seenfrom the mean data represented by the unfilledcircles in Figure 1 (lower panel), listeners steadilyimproved from an average score of 73.8% on theirfirst block of training to 86.6% on their last block oftraining. Before statistical analysis, all percent cor-rect scores were transformed into rationalized arc-sine units (RAUs; Studebaker, 1985) to stabilize theerror variance. The main effect of training blocks onword-recognition performance was significant(F(15,225) � 17.14, p � 0.001), as indicated by ageneral linear model repeated-measures analysis.Post hoc paired sample t-tests, using a Bonferroniadjustment for multiple comparisons, were thencomputed as needed, with all significant differencesdesignated at the p � 0.05 level. Performance on 13of the 14 training blocks (Training Block 2 throughTraining Block 14) was significantly better than thelisteners’ initial baseline closed-set performance(left filled circle). However, if one uses the initialtraining block as the baseline, performance did notimprove significantly until the completion of thesixth block of trials. At this point, which consisted of

TABLE 1. Presentation conditions for each testing session in Experiment 1

Talker List Response Condition Feedback

Session 1(Baseline) Female 1 Set A Open No

Female 1 Set B Open NoFemale 1 Practice Closed NoFemale 1 Set A Closed NoFemale 1 Set B Closed No

Sessions 2 through 4(Training) Female 1 Set A Closed Yes

Female 1 Set A Closed YesFemale 1 Set A Closed YesFemale 1 Set A Closed Yes

Session 5(Training) Female 1 Set A Closed Yes

Female 1 Set A Closed YesGroup 1/Group 2

(Post-train) Female 1 Set A Open/Closed NoFemale 1 Set B Open/Closed No

Session 6(Generalization to talkers) Female 1 Set A Closed/Open No

Male 1 Set A Closed/Open NoFemale 2 Set A Closed/Open NoMale 2 Set A Closed/Open No

Session 7(Generalization to words) Female 1 Set B Closed/Open No

Male 1 Set B Closed/Open NoFemale 2 Set B Closed/Open NoMale 2 Set B Closed/Open No


approximately 1.5 hr of training, performance wassignificantly better on all subsequent blocks relativeto the initial training session, as indicated by theasterisks in Figure 1. Although performance ap-peared to steadily improve through the last trainingsession, the overall improvement per hour of train-ing slowed. The rate of improvement decreased froma mean of 5.9% per hour after the first hour oftraining, to an improvement of 4.3% per hour for thesecond hour, to only about 0.5% per hour after that.Therefore, it might be reasonable to halt trainingafter 2 hr (8 blocks of training trials), in turn

causing only a small decrease in final performancewhile reducing the total amount of training timeneeded.

Figure 2 provides mean word-recognition scoresfor open-set and closed-set response formats for boththe trained and untrained words. Paired-samplet-tests were calculated for all pretraining and post-training word-recognition scores to examine the dif-ferences between trained and untrained words, withall significant differences designated at the p � 0.05level. Before training, there were no significantdifferences between the Set A and Set B words ineither the open-set (32.7% and 35.7%) or closed-setresponse formats (69.4% and 70.4%). Open- andclosed-set performance improved significantly aftertraining for both the trained and untrained words.Although there was a significant improvement inword-recognition scores after training for both setsof words, the magnitude of the differences weresmall for the untrained words relative to the trainedwords. The improvement in open-set performanceafter training, for example, was 52.5% for thetrained words compared with 11.1% for the un-trained words. This same trend held true for theclosed-set test conditions as well, with improve-ments of 16.7% for the trained words versus 9.2% forthe untrained words.

Figure 3 shows individual performance for the 8listeners in which the talker-generalization taskswere done using the closed-set response format.Analysis of the data in the left panel of Figure 3indicated there were no significant differences be-tween the trained list of words spoken by F1 (87.7%)and the same list spoken by the novel talkers(88.7%, 89.8%, and 86.8% for F2, M1, and M2,

Fig.1. Word-recognition performance for Set A words spokenby Female Talker Number 1, as a function of training block (1through 14) for the YNH listeners (SNR � 0 dB). Bothindividual data (upper panel) and mean data (lower panel) areshown. Significant differences identified are relative to thefirst block of training. *Significant at the 0.05 level.

Fig. 2. Word-recognition performance for Set A (trained) andSet B (untrained) words spoken by Female Talker Number 1,before and after training for the 16 YNH listeners (SNR � 0dB). Error bars equal �1 SD. *Significant at the 0.05 level.


respectively). For the closed-set scores on untrainedwords (Figure 3, right panel), the score for F1(80.2%) was significantly lower than that for M2(88.8%). No other differences between talkers weresignificant. If improved performance from trainingwas talker specific, listeners would be expected toperform better on a list presented by F1 relative tothe other three talkers, whether trained or other-wise. This was not the case for either set of words.

For the data obtained in the closed-set responseformat described above (Figure 3), ceiling effectsmay have prevented the occurrence of talker-specificdifferences, given the high levels of performance forall subjects when using this response format. Per-formance for the other eight listeners, however, wasassessed by using an open-set response format.Results for this group of eight subjects are shown inFigure 4. As can be seen, the trends are the same aswas previously shown for the closed-set group. Thatis, there was no significant difference in perfor-mance between talker F1 (80.2%) and the noveltalkers for the trained words (80.3%, 84.2%, and84.3% for F2, M1, and M2, respectively; see leftpanel of Figure 4). Further, performance for theuntrained words (right panel) was no higher for thetrained talker (F1; 47.3%) than the novel talkersspeaking those same words (52.8%, 60.7%, and62.2% for F2, M1, and M2, respectively). In fact,performance for the trained talker was significantlylower than that for each of the untrained talkers inthis condition.

In this experiment, the YNH listeners were ableto significantly improve their open- and closed-setword-recognition performance through training.This improved performance generalized to novelspeakers using both an open- and closed-set re-sponse format, implying that training a listener byusing a standardized list and talker may still pro-vide benefit when these same words are presentedby novel talkers in other contexts, including outsidethe laboratory or clinic. The larger increase in open-set performance for the trained lists versus theuntrained list spoken by the same talker wouldinitially point to a more lexically based representa-tion of the words rather than learning of specifictalker characteristics. Memorization of the lexicalrepresentation of spoken words would not rule outthe use of similar training techniques as a rehabil-itative procedure. It would, however, make thechoice of training words much more important. Onecould, for example, choose a restricted set of difficultto recognize, but frequently occurring words fortraining; words referred to as “hard words” in thelexical neighborhood theory (Luce & Pisoni, 1998).Further, it might be appropriate to assess errors fora standardized list of frequently occurring wordsand use these results to individually tailor subse-quent training. However, before developing morespecific word lists, the ability to replicate the pre-ceding study in older hearing-impaired listenersusing amplification required investigation.

Fig. 3. Word-recognition performance as a function of thetalker (Female 1, Male 1, Female 2, Male 2) for the trained(Set A) and untrained (Set B) words in a closed-set task for theYNH listeners (SNR � 0 dB). Individual data are shown.

Fig. 4. Word-recognition performance as a function of thetalker (Female 1, Male 1, Female 2, Male 2) for the trained(Set A) and untrained (Set B) words in an open-set task forYNH listeners (SNR� 0 dB). Individual data are shown.


EXPERIMENT 2As was shown, YNH listeners can improve their

word-recognition performance in a background ofspeech-shaped noise through training. Although thetraining was effective, the same may not hold truefor older hearing-impaired (OHI) listeners. There-fore, Experiment 2 examined the potential benefit oftraining for an OHI population. With the exceptionof the presentation levels, the procedures were iden-tical to Experiment 1.

METHODS

Participants

Seven OHI listeners (65 to 75 yr of age, mean �69.6 yr) with mild-to-moderate bilateral sensorineu-ral hearing loss participated in this study. Thegroup consisted of three men and four women. Meanpure-tone air-conduction thresholds are shown inFigure 5. The hearing-impaired listeners were re-cruited from a large-scale study within the Depart-ment of Speech and Hearing Sciences at IndianaUniversity, Bloomington. At the time of the study,none of the listeners were regular hearing aid users.Listeners were paid at the same rate per session andreceived the same bonuses as described in Experi-ment 1.

Presentation Levels and Filter Design

To ensure audibility of the test materials for thehearing-impaired listeners, levels of the unshapedspeech signal (overall level � 67.4 dB SPL) wereadjusted by using a quasi-DSL (Seewald, Hudson,

Gagne, & Zelisko, 1992) approach, in which the RMSspectrum of the test materials would be at least 10dB above the listener’s hearing thresholds, from 200to 4000 Hz. Each listener’s hearing thresholds in dBSPL for one-third–octave bands from 200 to 4000 Hzwere compared with the corresponding one-third–octave bands from the long-term average speechspectrum for the AB words. If the speech signal wasat least 10 dB above the hearing threshold, no gainwas provided. When the signal fell below the listen-er’s hearing threshold for a particular one-third–octave band, the appropriate gain for a 10 dB SLsignal was applied. Figure 6 shows a representativeamplified speech level based on the average hearingimpairment of the seven listeners. The selection ofmild to moderately impaired listeners for this studyresulted in minimal gain at most frequencies except4000 Hz, where approximately 20 dB of gain wasneeded. Individual digital filters created for eachhearing-impaired listener were accomplished usinga software based digital programmable filter (TDTPF1). The experimental design, which required au-dibility (�10 dB SL) of the speech signal through4000 Hz, together with the maximum availableoutput of the stimulus presentation system, limitedthe maximum degree of hearing loss for the partic-ipants to moderate levels. Based on pilot testing, thespeech signal for the OHI listeners was presented ata more advantageous SNR (�5 dB) to eliminate flooreffects and perceptually equate performance withthe YNH listeners from Experiment 1.


Figure 7 shows individual (top panel) and mean(bottom panel) word-recognition scores from pre-

Fig. 5. Mean pure-tone thresholds (dB HL) for seven olderhearing-impaired listeners. Error bars indicate �1 standarddeviation (dB).

Fig. 6. Example of the amplification provided to maintain anaudible signal of at least 10 dB SL across a wide bandwidth forthe mean hearing loss of the seven older hearing-impairedlisteners.


training (far left), through training, to posttraining(far right) for the OHI listeners. The main effect oftraining session on word-recognition performance(in RAUs) was significant (F(15, 90) � 5.81, p �0.001). As shown in Figure 7, the mean data (bottompanel) show that improvement was relatively stableat the end of 10 training blocks when compared withperformance for the first training block, as indicatedby the asterisks. The percent improvement per hourof training did not decrease to a negligible amount(0.4% per hour) until the last hour of training. Therate of learning did diminish over time, however,

from 7.6% per hour for the first hour to 3% and 1.9%per hour for the second and third hours of training,respectively.

Figure 8 compares pretraining and posttrainingword-recognition scores for both trained (Set A) anduntrained (Set B) words, in open- and closed-setresponse formats. Separate paired-sample t-testswere calculated for all pretraining and posttrainingword-recognition scores to examine differences be-tween the trained and untrained words. There wasno significant difference (p � 0.05) between the SetA and Set B words in either the open- or closed-setconditions before training. Open-set word-recogni-tion performance improved significantly from 38.3 to83.6% after training for the trained words (t ��18.8, df � 6, p � 0.001) and from 44.0 to 50.9% forthe untrained words (t � -2.98, df � 6, p � 0.05).However, the degree of improvement was small forthe untrained words (6.9%) compared with thetrained words (45.3%). For the closed-set conditions,the improvement on the trained words of 11.0%(72.4 to 83.4%) was significant (t � �3.94, df � 6, p� 0.05), whereas the 5.3% improvement for theuntrained words (77.0 to 82.3%) was not (t � �2.12.df � 6, p � 0.05).

Individual differences were examined by usingRAU-based 95% CD (14.8 RAU); four of the sevensubjects showed significantly higher scores thanbaseline on a consistent basis by the second block ofsession 3, or approximately 1.5 hr of training. Whenexamining the individual data, it was generally thecase that if initial baseline scores were less thanapproximately 75 RAU, scores tended to improvesignificantly with training (three of four such sub-

Fig. 7. Word-recognition performance for Set A words spokenby Female Talker Number 1, as a function of training block (1through 14) for the OHI listeners (SNR � �5 dB). Bothindividual data (upper panel) and mean data (lower panel) areshown. Significant differences identified are relative to thefirst block of training. *Significant at the 0.05 level.

Fig. 8. Word-recognition performance for Set A (trained) andSet B (untrained) words spoken by Female Talker Number 1,before and after training for the seven OHI listeners (SNR ��5 dB). Error bars equal �1 SD. *Significant at the 0.05 level.


jects), but if initial scores were greater than this,significant improvements were not observed (onlyone of three such subjects). Still, all seven subjectsshowed significant improvements from initial tofinal open-set scores.

For the OHI listeners, the generalization wasexamined only for the closed-set response format. Aspreviously shown in Experiment 1 for the YNHlisteners, the inclusion of the open-set generaliza-tion tasks yielded a very similar pattern of results tothe closed-set conditions but at lower overall levelsof performance.

Figure 9 shows the individual posttraining word-recognition scores for trained words (left panel) anduntrained words (right panel) spoken by each of thefour talkers (F1, F2, M1, M2). For comparison, thepretest scores for talker F1 also appear in eachpanel. For the trained words, there were no signifi-cant differences between the trained talker (F1;83.4%) and the three novel talkers (89.1%, 86.9%,and 87.6% for F2, M1, and M2, respectively). For theuntrained words, there was no advantage associatedwith words presented by the talker used in training.In fact, word-recognition performance on the list ofuntrained words spoken by Female Talker 1 (82.3%)was significantly lower (p � 0.05) than the wordsspoken by both male talkers (88.0% and 92.2%).

The goal of this experiment was to examinewhether the older listeners with hearing impair-ment could improve their performance with word-based auditory training in a manner similar to thatof the YNH listeners. The word-recognition perfor-

mance of the OHI listeners using amplified speechpresented at a more favorable SNR (�5 dB) followedthe same trend during training as the YNH listeners(at 0 dB SNR). That is, the OHI listeners were ableto significantly improve their open- and closed-setword-recognition performance through training.The OHI listeners, however, needed more trainingtime to achieve significant improvements in perfor-mance when compared with the YNH listeners.Whether the need for greater training time was dueto the hearing impairment, reduced cognitive abili-ties associated with aging, or both, remains to beseen. Nonetheless, the preceding training protocolseems effective at improving word-recognition abil-ities in noise, not only for the talker used in trainingbut for novel talkers as well. That is, in this trainingprotocol, the improved performance tended to beword-specific rather than talker-specific and ap-pears to be lexically based. However, if a trainingprotocol is to benefit a patient in everyday commu-nication situations, the ability to retain what hasbeen learned and to generalize it to these samewords used in running speech is paramount. Theseissues were addressed in the next experiment.

EXPERIMENT 3As shown in Experiment 2, OHI listeners can

benefit from training to improve their word-recogni-tion performance in noise. As was the case with theYNH listeners, improvements for the OHI listenerswere not restricted to talker F1 but generalized tonovel male and female talkers speaking the samewords. However, improved performance on isolatedwords does not necessarily indicate improved perfor-mance for more everyday stimuli (such as sen-tences). The key issue of any training procedurewould be its effect or benefit outside the laboratoryor clinic with conversational speech. Another rele-vant aspect of the evaluation of a training programis the listener’s ability to retain the benefits oftraining over an extended period of time. Therefore,Experiment 3 examined the OHI listeners’ retentionof the trained words over time, as well as theirability to transfer this training to those same wordsembedded within sentences.

METHODS

Participants

Five of the seven OHI listeners (68 to 75 yr of age,mean � 71 yr) returned approximately 6 mo aftercompletion of Experiment 2 for further testing. Thisgroup consisted of two men and three women. Lis-teners once again received the same pay rate persession including bonuses as described in Experi-

Fig. 9. Word-recognition performance as a function of thetalker (Female 1, Male 1, Female 2, Male 2) for the trained(Set A) and untrained (Set B) words in a closed-set task for theOHI listeners (SNR � �5 dB). Individual data are shown.


ment 1. A second group of nine YNH listeners (21 to33 yr of age, mean � 25.3 yr) was recruited to collectinitial pilot data for the second aspect of the exper-iment involving the generalization of training tosentences, as described below. The YNH listenerscompleted data collection in one 60-minute sessionin which they were paid $10.00.

Stimuli

The stimuli for the third experiment consisted ofboth words (AB words as described in Experiments 1and 2) and sentences. The sentence stimuli weredrawn from the TIMIT database (Garofolo et al.,1993). The goal was to find prerecorded sentencesthat contained a trained word (one from the Set Awords) as well as separate sentences that containeda word from the Set B or untrained words. Thesentences from the TIMIT database were chosenaccording to the following criteria: (1) sentencesspoken by speakers of North Midland dialect; (2)sentence length of 12 words or less to minimizememory demands; (3) as many male talkers asfemale talkers; and (4) balancing the location of thetrained word within the sentence across the set ofsentences (i.e., first third, middle, or last third ofsentence). This selection process initially yielded 50TIMIT sentences for Set A words (Sentence Set A)and 41 TIMIT sentences for Set B words (SentenceSet B).

Each sentence in Sentence Set A contained atrained word from the Set A words used in Experi-ments 1 and 2 and an untrained control word,excluding articles (chosen arbitrarily), allowing foran examination within each sentence of enhancedrecognition of trained versus untrained words. Thesentences for Sentence Set B contained a word fromthe untrained Set B word list used in the prior twoexperiments as well as a second untrained controlword. Although listeners were not intentionallytrained on the Set B words, by the time the sen-tences were presented, the word list had been re-peated several times during the generalization tasksof Experiment 2. Further, there were some small butsignificant training effects for Set B (untrained)words in the prior experiments. Because some train-ing may have occurred for the Set B words by theend of Experiment 2, the inclusion of these sen-tences allowed for an extra set of control conditions.

The sentences were designed to permit compari-sons of performance for trained words to untrainedcontrol words. An assumption was that the wordsselected in each sentence were equally intelligiblewithout training. Pilot data were collected from theuntrained YNH listeners to establish the equiva-lency of the Set A/Set B and control words in each

sentence. Those sentences that yielded differencesin mean keyword scores for Set A/Set B and controlwords exceeding 1 standard deviation (of the SetA/Set B word score) were eliminated. That is, sen-tences were eliminated if Set A and Set B wordswere either easier or harder to identify than thecorresponding control words in the same sentences.Collection of pilot data using the YNH listenersfurther reduced Sentence Set A to 33 sentences andthe Sentence Set B to 25 sentences. Despite thedesire to have an equal number of talkers from eachsex, in the end, sentence Set A was spoken by 20men and 13 women, and Sentence Set B containedsentences spoken by 16 men and 9 women.

The selected sentences were converted to wavefiles, at which point they were matched for averageRMS amplitude (�1 dB), using Cool Edit Pro soft-ware. The sentences were then recorded to channel 1of a compact disc with the overall presentation levelof the compact disc approximating (�3 dB) that usedfor the AB words in Experiment 2. A new speech-shaped noise was created to match (�1 dB) thelong-term average spectrum of all the sentencesconcatenated together. This spectrum matching ofspeech and competition was also accomplished usingCool Edit Pro. This speech-shaped noise was alsomatched in overall level to the speech-shaped noiseused in Experiment 2 and recorded to the secondchannel of the compact disc. The background noisebegan 1 second (500-msec fall-time) before the onsetof the sentence and ended 1 second (500-msec fall-time) after sentence completion.

Presentation Levels

Both the AB words and the sentences receivedspectral shaping of the speech sounds to ensureaudibility for the OHI listeners as described inExperiment 2. The same individually programmedfilters were used, once again making use of anunshaped (unaided) speech level of 67.4 dB SPL andquasi-DSL gain, as measured in an HA-2 2-cm3

coupler for both the words and the sentences. Basedon the pilot data collection for the TIMIT sentences,the background noise was adjusted to SNR values of0 dB and �5 dB for the OHI listeners. The sentencespresented to the YNH listeners during pilot datacollection were unshaped, low-pass–filtered at 10kHz, and had an overall level of 81 dB SPL. Thebackground noise for the YNH listeners was ad-justed to SNR values of –5 dB and 0 dB.

Procedures

Table 2 outlines the test conditions of Experiment3 for the OHI listeners. The baseline measures andtraining procedures for the word-recognition portion


of the experiment were once again identical to Ex-periment 2. After the baseline word-recognitionsscores were recorded, the number of training ses-sions for each OHI listener was dependent on thetraining necessary to raise their performance towithin the 95% critical difference (Thornton & Raf-fin, 1978) of their posttraining scores from Experi-ment 2. An open- and closed-set presentation of thetrained words by F1 at the end of Session 2 showedall five OHI listeners meeting this criterion at thattime. During Session 3, Sentence Set A and Sen-tence Set B were presented at 0 dB SNR for the firstthree OHI listeners and �5 dB SNR for the remain-ing two OHI listeners. Pilot data suggested that anSNR of 0 dB would be appropriate to avoid ceilingand floor effects and approximate the performanceof the YNH at a -5 dB SNR. However, when the firstthree OHI listeners failed to demonstrate transfer tosentences, as described below, it was decided thatthose remaining subjects would be tested at thesame SNR as used in Experiment 2. Both sentencesets were administered twice within the same ses-sion. Once again, the sentences were presented tothe right ear through the same equipment used forthe word presentation. After presentation of eachsentence, the listener repeated the sentence to theexperimenter. They were told to repeat back anyword they heard even if they could not repeat thewhole sentence. The experimenter, seated next tothe subject in the sound booth, recorded the correctidentification of the target word (the word fromSentence Set A or B) and the untrained control word

within each sentence. There was a 6-second re-sponse interval between each sentence.

RESULTS AND DISCUSSIONS

Figure 10 shows mean open-set word-recognitionscores for the OHI listeners on the trained anduntrained lists at pretraining, immediate posttrain-ing, and at the 6-month follow-up. The pretrain andposttrain bars are replots of the mean data fromExperiment 2 for the five returning OHI listeners.Word-recognition performance was significantly bet-ter on the trained words after 6 mo had lapsed

TABLE 2. Presentation conditions for each testing session in Experiment 2

Talker List Response condition Feedback

Session 1(Baseline after 6 mo) Female 1 Set A Open No

Female 1 Set B Open NoFemale 1 Practice Closed NoFemale 1 Set A Closed NoFemale 1 Set B Closed No

Sessions 2 through 4(Retraining) Female 1 Set A Closed Yes

Female 1 Set A Closed YesFemale 1 Set A Closed YesFemale 1 Set A Closed YesFemale 1 Set A Open NoFemale 1 Set A Closed No

Session 3(After training) Female 1 Set A Open No

Female 1 Set B Open NoFemale 1 Set A Closed NoFemale 1 Set B Closed No

(Sentences) – Sentence Set A Open No– Sentence Set B Open No– Sentence Set A Open No– Sentence Set B Open No

Fig. 10. Performance on the trained and untrained wordsbefore training, immediately after training, and at a 6-monthfollow-up for the five returning OHI listeners. Error barsequal �1 SD.


(62.9%) relative to the listeners’ initial pretrainingperformance of 37.6% (p � 0.05), indicating thatsome of the effects of training from 6 mo previouslyhad been retained. However, they still performedsignificantly worse than their maximum perfor-mance of 83.5%, obtained at the end of Experiment 2(p � 0.05).

After the initial baseline testing, listeners wereagain trained by using Set A, spoken by Female 1.As mentioned previously, the goal was to return thelisteners to within the 95% critical difference of theirposttraining scores at the end of Experiment 2. Notonly did the listeners’ perform significantly betterthan their initial pretraining baseline in Experi-ment 2, but no more than 1 hour of training (4 blocksof 75 words) was needed to return them to theirprevious posttraining performance levels.

Figure 11 provides the mean data for the un-trained YNH group and individual data for thetrained OHI listeners for each of the sentence con-ditions. The data from the untrained YNH listenersindicated no significant difference in the identifica-tion rate between the trained words and controlwords within Sentence Set A or between the un-trained words and the control words within Sen-tence Set B. For the three OHI listeners in Figure 11who listened to the sentences at 0 dB SNR, nostatistical analyses were completed because of thesmall number of listeners. However, visual inspec-tion of the data suggests no discernible difference inrecognition of a trained word versus a control oruntrained word embedded in a sentence. Figure 12shows similar data for the two OHI listeners testedat the same SNR in which they were trained inExperiment 2 (�5 dB SNR). The additional 5 dB

improvement to the signal creates a better perceptualmatch to the YNH listeners, but, once again, the trendshows no improvement in performance for trainedwords within sentences compared with the untrainedand control words. Replication of the sentence testinga second time failed to alter the pattern of resultsapparent in Figures 11 and 12.

The results of this experiment indicate that, onreturn in 6 mo, the OHI listeners were able tomaintain about half of the improvement in perfor-mance achieved earlier in training. At the start ofthe session, most of the listeners commented thatthey remembered none of the training from 6 mopreviously and would probably perform poorly. Al-though their word-recognition scores at the 6-monthfollow-up (�60%) did drop significantly from theirposttraining scores of Experiment 2 (�80%), theirword-recognition scores were still significantly bet-ter than on their initial baseline performance(�40%). This may indicate that a short trainingregimen after the hearing aid fitting may providelong-lasting increases in performance.

Based on the results obtained with the sets ofTIMIT sentences, however, training on isolatedwords did not transfer to fluent speech. This may bedue to one of the following reasons. First, the talkerswithin the sentences were not the same talkers forwhich training occurred. Although there was gener-alization to novel talkers when examining isolatedwords, it may not be appropriate to assume the samefor words in sentences. Had the initial words used inthe training been extracted from the sentencesthemselves, perhaps the results may have beendifferent. Regardless, the idea was to train listenerson an isolated word list and then evaluate the

Fig. 11. Mean percentage of key words identified correctly forYNH listeners (hexagons) and individual data for OHI listen-ers on trained versus control words within a sentence (SNR �0 dB). Error bars indicate �1 standard deviation (dB) re: themean for the YNH listeners.

Fig.12. Mean percentage of key words identified correctly forYNH listeners (hexagons; SNR � 0 dB) and individual data forOHI listeners (SNR � �5 dB) on trained versus control wordswithin a sentence. Error bars indicate �1 standard deviation(dB) re: the mean for the YNH listeners.


transfer of this training to sentences used in every-day speech. In this experiment, “everyday speech”was composed of sentence materials spoken by awide range of talkers. Either of these factors, wordsversus sentences and narrow versus wide range oftalkers, could have hindered the transfer of training.

A second possible reason for the lack of transfer oftraining in this study deals more directly with howlisteners recognize speech. That is, do they perceiveisolated words in the same way as words presentedwithin a sentence? This issue was discussed byGreenspan, Nusbaum, & Pisoni (1988), while inves-tigating perceptual learning of words and sentencespresented using synthetic speech. Although theirstudy used low-quality synthetic speech, a possibleconnection could be made between this poor-qualityspeech signal and the distorted speech signal withinnoise perceived by OHI listeners. Their results weresimilar to the data presented here, in that trainingwith isolated phonetically balanced words failed toimprove sentence intelligibility. However, when lis-teners were trained by using synthetic sentencesrather than words, improvements were noted forboth sentence and isolated word intelligibility.

Nygaard & Pisoni (1998) also examined the abil-ity to generalize knowledge acquired from word-based training to sentences, as well as the ability toimprove word-recognition abilities through sentencetraining. They used natural speech to train listenersto identify talkers of specific speech stimuli. Onceagain, they found word-recognition performance didimprove for words trained and presented in isola-tion; however, there was a large benefit in sentencetranscription only when the listeners were trainedspecifically with sentences. Unlike the Greenspan etal. (1988) paper, training with sentences did notgeneralize back to isolated words. More recently,Hirata (2004) trained native English speakers toperceive Japanese vowel and consonant contrasts.In this case, the listeners’ performance improvedafter training whether training was accomplishedvia words or sentences. However, the data “suggeststhat, given training in only one context, the degreeof generalization to the untrained context wasgreater for the sentence training than the wordtraining” (pg. 2392).

Several theories of speech perception propose thatimprovements in the recognition of words presentedin isolation should also improve performance forthose words in sentences (Greenspan et al., 1988).However, this was not the case when training usingsynthetic speech (Greenspan et al., 1988), normal-hearing listeners (Hirata, 2004; Nygaard & Pisoni,1998), or when training hearing-impaired listenersby using natural speech within a noise background(this study). These results suggest that a more

promising technique may be to train with sentencesthemselves, similar to synthetic approaches to audi-tory training.

Because learning in the present study appears tobe lexical in nature, another approach may be totrain using lexically “hard” words, as identified byusing the Neighborhood Activation Model (Luce, &Pisoni, 1998), presented by multiple talkers. It maybe that training using lexically hard words couldgeneralize back to easy words, while also bettertransferring to sentences. Regardless of the ap-proach, the ideal goal would be to obtain significantimprovements in everyday speech communication,including within a background noise, which mightbe attained within reasonable periods of trainingand retained indefinitely after training. Several is-sues arose, based on the results of the previous threeexperiments, particularly regarding the apparentlexical nature of the training, as well as the degreeto which performance on a set of words can bemaintained on the introduction of a new set ofwords.

EXPERIMENT 4The results from Experiments 2 and 3 suggest

that the training effects observed are mediated bythe lexical properties of the 75-item word list. Oneway to further examine this issue was to identify theaspects of the training protocol that helped improvethe listeners’ word-recognition performance in noise(i.e., the orthographic feedback, the acoustic proper-ties of the words themselves, or simply presentingthe list in a closed-set paradigm). Experiment 4looked at three separate issues: (1) would removingthe orthographic feedback have any effect on theclosed-set performance over time; (2) would trainingusing an all open-set condition with no feedback(i.e., purely acoustic information) show any increasein performance over time; and (3) once trained on alist of 75 words (Set A), if these lexical items areessentially memorized by the subject, would imme-diate training on another set of 75 words (Set B)diminish performance for the Set A words, or couldthe listeners maintain both the previous and newlytrained words?

METHODS

The previous three experiments showed similartrends and performance for the YNH and OHIlisteners. Thus, YNH listeners were used through-out Experiment 4. All stimuli, equipment, partici-pant selection, and payments were identical to theprevious experiments using the YNH listeners.


Procedures

To address the first question regarding the effec-tiveness of the orthographic feedback, five YNHlisteners duplicated Experiment 1 with the trial-to-trial orthographic feedback turned off. On comple-tion of the seven training and posttraining evalua-tion sessions, this group of five listeners wasimmediately started on a second training protocolusing the 75 Set-B words to address the question ofretention with an increased memory load or inter-ference of new lexical items with previously learneditems. A separate group of four YNH listeners alsoduplicated Experiment 1 using all open-responsetraining without feedback. This group had no expo-sure to the lexical items, either through the 75-itemresponse form on the computer screen or throughorthographic feedback. For this group, the closed-setbaseline testing was also removed to eliminate ex-posure to lexical information during the training.


Figure 13 shows group data for both the listenerswho received no orthographic feedback in the closed-set condition as well as the group of listeners whoreceived no feedback in an all open-set training-condition. Percent correct scores are shown pre-training through posttraining. For the group trainedin a closed-set paradigm, the main effect of training

block on word-recognition performance (after trans-formation to RAUs) was significant (F(13,52) � 4.41,p � 0.05). However, there were no significant differ-ences between the score on training block 1 and eachsubsequent training block, as identified by a posthoc paired-sample t-tests, using a Bonferroni adjust-ment for multiple comparisons. Mean word-recogni-tion performance on the trained words improvedsignificantly, by an average of 56.6% (from 22.1 to78.7%) and 22.6% (from 58.7 to 81.3%) for the open-and closed-set conditions, respectively. The un-trained (Set B) words showed no significant im-provements before training to after training foreither the open-set (from 27.2 to 37.3%) or theclosed-set (from 66.7 to 74.7%) words. These resultswere very similar to the improvements shown on thetrained words in Experiment 1 (52.5 and 16.7% forthe open- and closed-set conditions), using ortho-graphic feedback. In this case, the removal of thetrial-to-trial feedback had little effect on the overallimprovements shown for the trained words but didlimit generalization to untrained words.

The group who received no feedback or lexicalinformation from the printed list (i.e., all open-set)showed less but still significant improvementthroughout the training protocol. Although meanopen-set word-recognition performance improvedsignificantly from 29 to 43.3% for the trained list,the improvements were much smaller than thosegroups who received lexical input through theclosed-set presentation protocol and orthographicfeedback. There was no significant improvementfrom before training (33.7%) to after training(36.3%) for the untrained words in the open-setcondition. This group of listeners did not show asignificant main effect of training block on word-recognition performance (F(13, 39) � 1.03, p � 0.05).In summary, although this group improved signifi-cantly from before training to after training, theoverall improvements were smaller than those ob-tained with feedback, and, there were no improve-ments (generalization) to the untrained (Set B)words.

The elimination of the feedback from both theopen- and closed-set protocol also helped to examinethe effect of true training versus procedural learn-ing. As can be seen in Figure 13, the initial increasein performance from baseline testing to the scoreobtained on the first block of training, which basi-cally represents immediate or short-term retest,does point to some degree of procedural learning.Word-recognition performance immediately im-proved by 7% and 9% for the open- and closed-setgroups, respectively, on the second block of testing.After the initial increase in performance, a furthergain of 7.3% for the open-set group and 13.6% for the

Fig. 13. Word-recognition performance for Set A wordsspoken by Female Talker Number 1, as a function of trainingblock (1 through 14) for the YNH listeners (SNR � 0 dB).Mean data for the group who received closed-set presenta-tions with no feedback (circles) and the group who receivedan open-set presentations with no-feedback (squares) areshown.


closed-set group was noted. The initial increase of 7to 9% is very similar to the acclimatization effectnoted for new hearing aid users over time (Turner,Humes, Bentler, & Cox, 1996). It may be that theacclimatization effect is a consequence of procedurallearning rather than true acclimatization.

The third question addressed the listener’s abilityto maintain performance on previously learnedwords as new words were introduced. The listeners’mean open-set performance on the Set A (trained)words on starting the training protocol for the Set B(untrained) words was 74.7%. After training on theSet B words, the mean posttraining score for theinitial trained (Set A) words dropped by only 2.7%,which was not significant (p � 0.05). In this case, theaddition of 75 newly trained words did not signifi-cantly diminish performance on the previouslylearned Set A words, whereas improvements to thenewly trained Set B words followed the same trendsas the original protocol from Experiment 1.

Experiment 4 demonstrates that repeated expo-sures to speech stimuli in noise can lead to improvedperformance in and of itself, but the size of thetraining or learning effect is enhanced if the subjectsare provided with lexical information about thestimulus items. The lexical information can be pro-vided by visualizing the set of possible stimuli orthrough trial-to-trial orthographic feedback. Theability to effectively train a listener on a greaternumber of words (up to 150) also enhances thetraining possibilities by expanding the set of possi-ble words trained without acquisition of the new setinterfering with retention of the former set.

GENERAL DISCUSSION

Experiments 1 and 2 examined the efficacy of aword-based auditory training protocol. Both theYNH and OHI listeners were able to improve theirclosed- and open-set performance substantiallythrough training through repeated presentation ofthe test words. Although the OHI listeners were ableto improve their performance over time in much thesame way as that of the YNH listeners, they didneed a more advantageous SNR and more trainingtime to do so.

The improved word-recognition performance forthe OHI listeners could be explained several ways.As stated previously, the learning seemed to bemore lexically based, pointing to a memorizationof the trained words as presented by FemaleTalker 1. This explanation seems reasonable, asthe trained word set went from a true open-set, onthe first presentation, to what essentially becamea closed-set on completion of the training protocol.The positive effects of reduction in set size (in this

case from an unlimited open set to a closed set of75 words) on word-recognition performance arewell documented (Miller, Heise, & Lichten, 1951).However, pure memorization of the words them-selves cannot account for the significant (albeitsmaller) improvements in open-set performanceon the Set B words for the listeners after training.Further, generalization of training to untrainedtalkers in Experiments 1 and 2 argues that thememorization was lexical rather than acoustic innature. The learning that takes place is more thanlikely a combination of the process of lexical mem-orization and the ability to make use of newacoustic cues through repeated exposures (open-set training from Experiment 4). However, thelistener’s ability to retain a significant amount ofknowledge garnered from training for at least 6mo also points to a process of learning beyondimmediate or short-term memorization of thewords.

The ability of the OHI listeners to both improvetheir word-recognition performance in noise whilealso retaining significant knowledge of their train-ing over extended periods of time, point to auditorytraining procedures as a potentially viable rehabili-tative technique. However, several issues still re-main, not only regarding the specific words or talk-ers best suited for word-based training, but alsoconcerning the lack of transfer to sentences. Forexample, would training using multiple talkers bet-ter generalize to novel words and talkers more sothan training using an individual talker? And, if so,is the potentially shorter training time associatedwith a single talker versus training with multipletalkers a reasonable trade-off (Lively, Logan, &Pisoni, 1993; Lively, Pisoni, Yamada, Tohkura, &Yamada, 1994; Logan, Lively, & Pisoni, 1991)?

Rehabilitative training procedures have been gar-nering more interest in recent years, mainly due totechnological advances enabling a hearing aid userto accomplish training while at home, using a per-sonal computer. The data discussed in this studyrepresent a preliminary step in the design of word-based training that ultimately could be produced aseither a standardized or individually tailored train-ing protocol. Once again, the major focus of anauditory rehabilitative technique is its ability togeneralize to everyday speech. Although word train-ing appears effective at improving recognition of thetrained words, it does not appear to be an effectivemethod for improving patients’ real-world commu-nication abilities, using sentences. Additional ques-tions must be resolved before a specific auditorytraining protocol could be recommended.


ACKNOWLEDGMENTS

A portion of this paper was presented at the InternationalHearing Aid Research Conference (IHCON) in Lake Tahoe, Cal-ifornia, 2004. This work was supported in part by NIH grantR01-AG08293, awarded to the second author, and by an NIHpostdoctoral training grant T32 DC00012 (first author). Theauthors thank Chris Turner and Andrea Pittman for their helpfulcomments throughout the revision process.

Address for correspondence: Matthew H. Burk, Ph.D., Depart-ment of Speech and Hearing Sciences, Indiana University, 200South Jordan Avenue, Bloomington, IN 47405-7002. E-mail:[email protected]

Received November 25, 2004; accepted October 31, 2005.

REFERENCES

American National Standards Institute. (1996). Specifications foraudiometers, (S3.6-1996). New York: American National Stan-dards Institute.

American National Standards Institute. (1999). Maximum per-missible ambient noise levels for audiometric test rooms, (S3.1-1999). New York: American National Standards Institute.

Amos, N., & Humes, L. (2001). The contribution of high frequenciesto speech recognition in sensorineural hearing loss. In A. Hout-sma, A. Kohlrausch, V. Prijs & R. Schoonhoven (Eds.), Physiolog-ical and Psychophysical Bases of Auditory Function. (pp. 437–444). Maastricht, The Netherlands: Shaker Publishing BV.

Blamey, P., & Alcantara, J. (1994). Research in auditory training.Journal of the Academy of Rehabilitative Audiology, Supple-ment, 27, 161–191.

Boothroyd, A. (1995). Speech perception tests and hearing-im-paired children. In G. Plant & K. E. Spens (Eds.), ProfoundDeafness and Speech communication (pp. 345–371). London:Whurr Publishers.

Boothroyd, A. (1999). Computer-Assisted Speech Perception As-sessment (CASPA v2.2). San Diego, CA: Boothroyd, A.

Brooks, D. N. (1979). Hearing aid candidates: some relevantfeatures. British Journal of Audiology, 13, 81–84.

Cox, R., & Alexander, G. (1992). Maturation of hearing aid benefit:objective and subjective measurements. Ear and Hearing, 13,131–141.

Garofolo, J., Lamel, L., Fisher, W., Fiscus, J., Pallett, D., &Dahlgren, N. (1993). DARPA TIMIT acoustic-phonetic contin-uous speech corpus CD-ROM. Technical Report NISTIR 4930,National Institute of Standards and Technology.

Gatehouse, S. (1992). The time course and magnitude of percep-tual acclimatization to frequency responses: evidence frommonaural fitting of hearing aids. Journal of the AcousticalSociety of America, 92, 1258–1268.

Greenspan, S., Nusbaum, H., & Pisoni, D. (1988). Perceptuallearning of synthetic speech produced by rule. Journal ofExperimental Psychology: Learning, Memory, and Cognition,14, 421–433.

Hirata, Y. (2004). Training native English speakers to perceiveJapanese length contrasts in word versus sentence contexts.Journal of the Acoustical Society of America, 116, 2384–2394.

Humes, L. (1996). Speech understanding in the elderly. Journalof the American Academy of Audiology, 7, 161–167.

Humes, L. (2002). Factors underlying the speech-recognitionperformance of elderly hearing-aid wearers. Journal of theAcoustical Society of America, 112, 1112–1132.

Humes, L., Christensen, L., Bess, F., & Hedley-Williams, A.(1997). A comparison of the benefit provided by well-fit linearhearing aids and instruments with automatic reductions oflow-frequency gain. Journal of Speech, Language, and HearingResearch, 40, 666–685.

Humes, L., Christensen, L., Thomas, T., Bess, F., Hedley-Wil-liams, A., & Bentler, R. (1999). A comparison of the aidedperformance and benefit provided by a linear and a two-channel wide dynamic range compression hearing aid. Journalof Speech, Language, and Hearing Research, 42, 65–79.

Humes, L., Garner, C., Wilson, D., & Barlow, N. (2001). Hearing-aid outcome measured following one month of hearing aid useby the elderly. Journal of Speech, Language, and HearingResearch, 44, 469–486.

Humes, L., & Wilson, L. (2003). An examination of changes inhearing-aid performance and benefit in the elderly over a3-year period of hearing-aid use. Journal of Speech, Language,and Hearing Research, 46, 137–145.

Humes, L., Wilson, D., Barlow, N., Garner, C., & Amos, N. (2002).Longitudinal changes in hearing aid satisfaction and usage inthe elderly over a period of one or two years after hearing aiddelivery. Ear and Hearing, 23, 428–438.

Larson, V., Williams, D., Henderson, W., Luethke, L., Beck, L.,Noffsinger, D., et al. Efficacy of 3 commonly used hearing aidcircuits: a crossover trial. NIDCD/VA hearing aid clinical trialgroup. The Journal of the American Medical Association, 2000;284, 1806–1813.

Lively, S., Logan, J., & Pisoni, D. (1993). Training Japanese listen-ers to identify English /r/ and /l/, II: the role of phonetic environ-ment and talker variability in learning new perceptual categories.Journal of the Acoustical Society of America, 94, 1242–1255.

Lively, S., Pisoni, D., Yamada, R., Tohkura, Y., & Yamada, T.(1994). Training Japanese listeners to identify English /r/ and/l/, III: long-term retention of new phonetic categories. Journalof the Acoustical Society of America, 96, 2076–2087.

Logan, J., Lively, S., & Pisoni, D. (1991). Training Japaneselisteners to identify English /r/ and /l/: a first report. Journal ofthe Acoustical Society of America, 89, 874–886.

Luce, P., & Pisoni, D. (1998). Recognizing spoken words: theneighborhood activation model. Ear and Hearing, 19, 1–36.

Massachusetts Eye and Ear Infirmary. (1991). Q/Mass SpeechAudiology, Volume 2. Minneapolis: Qualitone.

Miller, G., Heise, G., and Lichten, W. (1951). The intelligibility ofspeech as a function of the context of the test materials.Journal of Experimental Psychology, 41, 329–335.

Nygaard, L., & Pisoni, D. (1998). Talker-specific learning inspeech perception. Perceptual Psychophysics, 60, 355–376.

Seewald, R., Hudson, S., Gagne, J., & Zelisko, D. (1992). Com-parison of two methods for estimating the sensation level ofamplified speech. Ear and Hearing, 13, 142–149.

Smedley, T. C. (1990). Self-assessed satisfaction levels in elderlyhearing aid, eyeglass, and denture wearers: a cross-modalitycomparison. Ear and Hearing, 11, 41S–47S.

Studebaker, G. (1985). A rationalized arcsine transform. Journalof Speech and Hearing Research, 28, 455–462.

Thornton, A., & Raffin, M. (1978). Speech-discrimination scoresmodeled as a binomial variable. Journal of Speech and HearingResearch, 21, 507–518.

Turner, C., Humes, L., Bentler, R., & Cox, R. (1996). A review ofpast research on changes in hearing aid benefit over time. Earand Hearing, 17, 14S–25S.

Walden, B., Erdman, S., Montgomery, A., Schwartz, D., & Prosek,R. (1981). Some effects of training on speech recognition byhearing-impaired adults. Journal of Speech and Hearing Re-search, 24, 207–216.


Effect of Training on Word-Recognition …audres/Publications/humes/papers/132...Effect of Training...

Documents

Transcript of Effect of Training on Word-Recognition …audres/Publications/humes/papers/132...Effect of Training...