Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
J Am Acad Audiol 12 : 233-244 (2001) Effects of Lexical Factors on Word Recognition Among Normal-Hearing and Hearing-Impaired Listeners Donald D . Dirks* Sumiko Takayanagi* Anahita Moshfegh' Abstract An investigation was conducted to examine the effects of lexical difficulty on spoken word recognition among young normal-hearing and middle-aged and older listeners with hearing loss . Two word lists, based on the lexical characteristics of word frequency and neighborhood density and frequency (Neighborhood Activation Model [NAM]), were developed : (1) lexically "easy" words with high word frequency and a low number and frequency of words phonemically similar to the target word and (2) lexically "hard" words with low word frequency and a high number and frequency of words phonemically similar to the target word . Simple and transformed up-down adaptive strategies were used to estimate performance levels at several locations on the performance-intensity functions of the words . The results verified predictions of the NAM and showed that easy words produced more favorable performance levels than hard words at an equal intelligibility . Although the slopes of the performanceintensity function for the hearing-impaired listeners were less steep than those of normal-hearing listeners, the effects of lexical difficulty on performance were similar for both groups . Key Words: Lexical stimuli, simple and transformed up-down adaptive strategies, word recognition Abbreviations : ANOVA = analysis of variance, CVC = consonant-vowel-consonant, NAM = Neighborhood Activation Model, NU-6 = Northwestern University Auditory Test No . 6, SNR = single-to-noise ratio ost theories of spoken word recognition involve several underlying perceptual processes in which the speech signal converted into an acoustic-phonetic repreis normalized for factors such as difsentation, ferences among talkers, speaking rate, and dialects and then matched to one of the thousands of items stored in long-term memory. Regarding the last step in this process, it has been shown that several lexical factors, including the frequency of occurrence of a word in M *National Center for Rehabilitative Auditory Research, Veterans Administration Medical Center, Portland, OR ; !Veterans Administration Greater Los Angeles Healthcare System ; 'Division of Head and Neck Surgery, UCLA School of Medicine, Los Angeles, California Reprint requests : Donald D . Dirks, Division of Head and Neck Surgery, UCLA School of Medicine, Rehabilitation Center, Bldg 31-24, Los Angeles, CA 90095 the lexicon and the number and frequency of other words phonemically similar to the target, affect the speed and accuracy of spoken word recognitions (Luce, 1986). The influence of these lexical factors on word recognition has been addressed in the Neighborhood Activation Model (NAM), which assumes that words that occur frequently and have few phonemically similar neighbors (lexically "easy" words) are recognized more accurately than words that occur less frequently but have a large number of phonemically similar neighbors (lexically "hard" words) . Most empirical support for the model has been found among young normal-hearing listeners (Luce, 1986 ; Sommers, 1996 ; Luce and Pisoni, 1998 ; Dirks et al, 2001). In addition, experimental results have also indicated that the NAM's principles can be generalized to children fitted with cochlear implants (Kirk et al, 1995), adults with acquired sensorineural hearing loss (Kirk et al, 1997 ; Dirks et al, 2001), and 233 Journal of the American Academy of Audiology/Volume 12, Number 5, May 2001 non-native listeners with normal hearing (Bradlow and Pisoni, 1999). Recognition performance, in the aforementioned investigations, has routinely been measured at a single speech presentation level or signal-to-noise ratio (SNR), frequently selected at or near ceiling levels for word recognition. Thus, comparative results between easy and hard word recognition are necessarily limited to one estimate on the underlying performance-intensity functions. As will be reviewed in detail, the results in all studies show higher performance for easy versus hard word recognition; however, the magnitude of the effect varies . To examine the effects of these lexically based words on speech recognition in a more comprehensive manner, in this investigation, intelligibility was measured at several locations on the response curves for easy and hard words among listeners with normal hearing and hearing loss . The NAM proposes that discrimination and identification of a particular target word are influenced by several lexical factors. Central to the model is the concept that words in the mental lexicon are organized into similarity neighborhoods, which can be created from the target item by adding, deleting, or substituting a single phoneme. Thus, words such as "nap," "sap," "lap," and "mad" would be considered neighbors of the word "map ." A similarity neighborhood is composed of two structural characteristics: (1) neighborhood density or the number of words that are phonemically similar to a particular target item and (2) neighborhood frequency, which refers to the average frequency of occurrence of items in the neighborhood . The frequency of occurrence of a word in the language has often been based on word frequency counts such as those provided by the Brown Corpus of printed text (Kucera and Francis, 1967). Target words with many neighbors (phonemically similar words) are located in "dense" neighborhoods and those with few neighbors in "sparse" neighborhoods . A third lexical factor, the frequency of occurrence of the target word itself, is considered a biasing factor, making it easier to recognize a word that occurs often in language as opposed to one that is seldom used . Although neighborhood density, neighborhood frequency, and word frequency each significantly influence word recognition, any one of these lexical properties is independently insufficient to account for the effect of lexical discrimination on word recognition. Rather, words are recognized relationally, that is, word frequency and the neighborhood structural effects of neighborhood density and 234 average neighborhood frequency covary (Meyer and Pisoni, 1999 ; Dirks et al, 2001). Further theoretical details of NAM have been described by Luce and Pisoni (1998) . According to NAM, words with high word frequency and low neighborhood density and neighborhood frequency should be identified more accurately than familiar words with low word frequency and high neighborhood density and neighborhood frequency. In investigations by Luce and Pisoni (1998) and Dirks et al (2001), the predictions from NAM theory were tested comprehensively by measuring performance for experimental conditions that were formed orthogonally by combining two levels of stimulus and frequency (high and low), neighborhood density (high and low), and average neighborhood frequency (high and low) . Combining the three independent lexical variables and two levels resulted in eight possible lexical conditions . In the Luce and Pisoni (1998) study, subjects were normal-hearing listeners, whereas Dirks et al (2001) included normal and hearingimpaired listeners . Dirks et al (2001) attempted to minimize ceiling effects by word presentations at SNRs or speech presentation levels where performance resulted in scores between 40 and 70 percent correct . Although the recognition scores for normal and hearing-impaired listeners varied slightly, the results demonstrated that the highest performance scores were always obtained for the high word frequency and low neighborhood density and neighborhood frequency condition . For both groups of subjects, scores from this condition were, on average, 15 percent higher than scores from the condition with the lowest performance scores, the low word frequency and high neighborhood density and neighborhood frequency condition . In the Luce and Pisoni (1998) study, the aforementioned conditions also produced the highest and lowest scores . Because of the large and consistent difference in performance between these conditions for both normal and hearing-impaired listeners, in the current experiment, lexically easy words were derived from a database of consonant-vowel-consonants (CVCs) that were used to form the high word frequency and low neighborhood density and neighborhood frequency condition, whereas lexically hard words were those incorporated in a database from the low word frequency and high neighborhood density and neighborhood frequency condition . The definition of lexically easy and hard words has varied somewhat among investigators, some stressing word frequency and neigh- Effects of Lexical Factors/Dirks et al borhood density (Kirk et al, 1997) and others neighborhood density and neighborhood frequency, with word frequency itself controlled (Sommers, 1996 ; Sommers et al, 1997) . In the studies by Luce and Pisoni (1998), Dirks et al (2001), and Bradlow and Pisoni (1999), the three lexical factors were varied to form the experimental conditions . Regardless of the variations in definition, however, performance scores for lexically easy words have always been found to be superior to those for hard words, although the difference in scores between the easy and hard word lists has varied . As indicated earlier, in experiments concerned with the effects of lexical difficulty on word recognition, the stimuli have been administered at a single presentation level or SNR . As a consequence, performance could only be evaluated from one location on each of the underlying performance-intensity functions . In the previously cited experiments, it was not unusual to present the speech at a high comfortable presentation level or favorable SNR . Thus, performance ceiling effects probably influenced the difference in scores between the lexically easy and hard word conditions in such experiments . For hearing-impaired listeners, the problem of choosing the appropriate speech presentation level to avoid ceiling or floor effects is especially difficult because of the large variations in the magnitude of the hearing loss and the audiometric configuration . Even individuals with similar magnitudes of hearing loss may have different performance-intensity functions and maximally obtainable speech recognition scores . In the current investigation, simple and transformed up-down adaptive strategies (Levitt, 1971) were used to estimate 29 .3, 50 .0, and 70 .7 percent correct on the response curves for easy and hard words . Thus, performance was measured at equal intelligibility for easy and hard word recognition among both normal-hearing and hearing-impaired subjects . There are at least two advantages in choosing adaptive strategies for this investigation . First, performance is measured at equal intelligibility on the response curve for all subjects . Thus, at least one common performance reference exists for comparison between results from normal-hearing versus hearing-impaired individuals . Second, by choosing adaptive strategies that measure word recognition at points between -30 and 70 percent (the linear portion of the response curve), it is possible to minimize the risk of measuring performance at presentation levels where scores begin to transition from the linear to the non- linear segment of the performance function, often making comparisons between group results problematic . The primary purpose of this investigation was to examine the performance-intensity functions for lexically easy and hard words for normal-hearing subjects and individuals with mild-to-moderate sensorineural hearing loss . Because the data of Sommers (1996) suggested that age itself may affect the ability to isolate an individual lexical representation from among phonetically similar competitions, half of the hearing-impaired subjects in the current investigation were middle-aged between 33 and 60 years and the other half were older listeners between the ages of 61 and 80 years . Scores were obtained using adaptive strategies, which converged on 70 .7, 50 .0, and 29 .3 percent response points on the performance-intensity curves . Thus, comparisons could be made between performance for the lexically easy and hard words at several locations on the response curves where ceiling and floor effects were absent or minimized . METHOD Participants Participants in the study were 20 adults with normal hearing and 20 adults with sensorineural hearing loss . The 20 adults (10 male and 10 females) with normal hearing had a mean age of 22 years (range 18-33 years) and were recruited from the student and employee population at UCLA . These subjects had normal hearing (<_15 dB HL at octave test frequencies from 250 to 6000 Hz) and scored 94 percent or higher on the Northwestern University Auditory Test No . 6 (NU-6) (Department of Veterans Affairs, Version 1 .1, 1991) with words presented at 40 dB SL re : spondee threshold . The 20 hearing-impaired adults (7 male and 13 females) had a mean age of 62 years (range 33-81 years) . Half of the subjects were less than 60 years, and half were over 60 years of age . The average age for the middleaged subgroup was 50 years and for the older subgroup was 75 .1 years . Subjects with mildto-moderate sensorineural hearing loss were recruited from patients seen for otologic and audiologic assessment at the UCLA Medical Center or the Veterans Administration Greater Los Angeles Healthcare System . Among the older participants (>60 years), the etiology of the hearing loss was diagnosed either as pres- 235 Journal of the American Academy of Audiology/Volume 12, Number 5, May 2001 bycusis or as presbycusis combined with noiseinduced hearing loss . Among the 10 middleaged patients, 5 had a history of noise exposure, and the remainder had hearing loss of unknown origin . All subjects denied serious neurologic problems (such as stroke, central nervous system lesions, or cognitive disorders), and none had been tested for central nervous system problems or serious mental disorders . For most subjects, the hearing loss was bilaterally symmetric . If there was a difference between ears for an individual subject, the ear with the more sensitive thresholds was tested during the experiment . Table 1 presents the mean auditory test thresholds of subjects along with results from speech audiometry. All subjects were native English speakers and were paid for their participation . imental conditions in which the three lexical conditions were systematically manipulated . A total of 114 CVC words from the database were found to conform to these lexical characteristics and were considered easy words for the current study. The earlier experimental results also demonstrated that words with low word frequency and high neighborhood density and frequency produced the poorest recognition scores . A total of 123 CVC words that possessed these lexical characteristics were found and considered hard words. From the available words in each lexical category, 75 easy and 75 hard words were selected for use in the current experiment. In generating the specific 75 easy and 75 hard words, attention was given to the types of phonemes contained in each word list . Although no attempt was made to phonetically balance the words, each list contained a representative sample of phonemes . Because persons with hearing loss may find special difficulty in the recognition of certain consonants (e .g., unvoiced fricative), an attempt was made to control the phonetic content of the words between the two lists. There were, however, only a modest number of CVC words available for each lexical group, so it was not possible to equate the phonemic content between the lists completely. Table 2 provides a frequency count of the linguistic features for the easy and hard word lists. The phonemic content (percent of phonemes in the initial and final positions) between the two lists varied modestly, but the differences were not large. The specific words used in this investigation are reported in the Appendix. Table 3 provides descriptive statistics concerning the lexical characteristics for the easy Test Stimuli Lexically easy and hard words were selected using an online lexical database (the 20,000word Hoosier Mental Lexicon; Nusbaum et al, 1984) derived from Webster's 1967 Pocket Dictionary. The database contains orthographic and phonetic transcriptions, a word frequency count based on Kucera and Francis (1967), a subjective rating of familiarity (Nusbaum et al, 1984 ; the rating scale ranged from 1.0 "don't know the word" to 7 .0 "know the word and its meaning"), and quantitative estimates of neighborhood lexical density and mean neighborhood frequency. As indicated previously, earlier results (Dirks et al, 2001) demonstrated that CVC words with high word frequency counts and low neighborhood density and frequency always produced the highest recognition scores among eight exper- Table 1 Means and SDs for Auditory and Spondee Word Thresholds and Speech Recognition Scores for Hearing-Impaired Listeners Auditory Thresholds (dB HL) Frequency (kHz) 0 .25 0.5 Middle-aged hearing impaired (n = 10) 16 .5 (12 .3) 17 .5 (10 .3) Older hearing impaired (n = 10) 18 .5 (9 .4) 20 .5 (12 .3) Total hearing impaired (n = 20) 17 .5 (10 .7) 19 .0 (11 .2) 236 1 .0 2.0 4 .0 6.0 Spondee NU-6 Word Recognition Threshold Score (%) 25 .0 (12 .9) 33 .5 (9 .1) 47 .0 (9 .2) 52 .5 (4 .9) 25 .2 (8 .3) 28 .5 (17 .0) 38 .5 (10 .3) 55 .5 (11 .7) 67 .5 (10 .1) 28 .3 (12 .4) 90 .1 (7 .1) 26 .7 (14.8) 36.0(9 .8) 51 .2(11 .1) 60 .0(10 .9) 26 .8 (10 .4) 91 .4 (6 .1) 92 .8 (4 .7) Effects of Lexical Factors/Dirks et al Table 2 Summary of the Phonemic Content of the 75 Easy and 75 Hard Words Initial Position (%) Phoneme Type IPA Symbol Voiced plosives Unvoiced plosives /b/, /d/ ./g/ Unvoiced fricatives Nasals I fl, l0/, /s/, /S/, /hl Voiced fricatives Approximants Voiced affricates Unvoiced affricates /P/. /t/, /k/ /v/, /8/ , / Z/ /m/, /n/, /rq/ /I/, /r/, /wL /j/ /tS/ /dZ/ Final Position (%) Easy Hard Easy Hard 20 .0 20 .0 22 .7 21 .3 9 .3 21 .3 13 .3 30 .7 10 .7 24 .0 33 .3 5 .3 1 .3 5 .3 5 .3 0 .0 0.0 2.7 0 .0 14 .7 14 .7 6 .7 10 .7 17 .3 24 .0 25 .3 5 .3 5 .3 0.0 6.7 6 .7 16 .0 */h/ is grouped as fricative as per categorization by Ladefoged (1993) . IPA = International Phonetic Alphabet . and hard word lists . Words were selected so that the median neighborhood density of the easy word list was considerably lower than the median neighborhood density of the hard word list (15 .0 vs 26 .0) . The median neighborhood frequency for the easy words was also much lower than the median neighborhood frequency of the hard words (30 .6 vs 180 .4) . Finally, median word frequency for the easy list was much higher than the median word frequency of the hard list (77 .0 vs 6 .0) . Median values were used in developing the word lists rather than the mean because the distribution of the individual lexical variables was not always symmetric around the mean . In summary, the easy word list is characterized as a set of words that occur frequently in the language and have low-frequency neighbors . In contrast, the hard word list contained words that occur much less frequently in the language than the easy words and have many neighbors that are high in frequency relative to the easy word list . Table 3 Lexical Factors Easy words Median Mean Range Hard words Median Mean Range Recording Procedure An adult female talker with General American English recorded both the easy and hard words. The talker was seated in a double-walled sound attenuation room (Industrial Acoustics, Model l204A) and read the words, from sheets placed in front of her, at a normal conversational level. The stimuli were transduced via a microphone (AKG, C460b, flat frequency response) and associated amplifier onto a digital audiotape recorder (Sony, DAT DTC-46) . The productions were monitored by three laboratory staff, and occasionally words were recorded several times because of misarticulations or mispronunciations . The words stored on the audiotape were digitized via a Computerized Speech Laboratory System (CSL model-4300, Kay Electronic Systems) at a sampling rate of 25 kHz with 16-bit resolution . The items were "down-sampled" to 24 kHz for compatibility with the sound processing board used to deliver the stimuli dur- Descriptive Statistics for the Easy and Hard Word Lists Word Frequency Neighborhood Density Neighborhood Frequency 77 .0 15 .0 30 .6 878 .0 16 .0 81 .5 141 .5 6 .0 6 .6 15 .0 14 .2 26 .0 27 .0 16 .0 36 .8 180 .4 276 .7 1009 .4 Median word frequency is reported in occurrences per million, median neighborhood density is the number of lexical neighbors, and median neighborhood frequency is the average frequency of all of the neighbors. 237 Journal of the American Academy of Audiology/Volume 12, Number 5, May 2001 ing the experiment . The words were edited, and the speech files were then converted to WAV format and prepared for presentation . The long-term root mean square (rms) spectrum of the talker was measured by playing a concatenated string of the 75 easy or hard words and averaged on a real-time analyzer (LarsonDavis 4300). The spectrum was typical of that for female talkers. This particular talker had recorded words for a previous study, and the detailed spectral characteristics have been reported (Dirks et al, 2001, Fig. 1) . Spectral analysis of the word lists indicated that there was essentially no difference between the long-term rms level of the easy and hard words. In addition, the rms level of each individual word was also measured using a software analysis program (Cool Edit Pro Version 1 .0, 1997). From this analysis, it was determined that the individual words varied in rms level over a range of 6 dB . This range is similar to that reported in a previous study (Dirks et al, 2001) for words spoken under similar conditions and at a conversational level. The speech levels for the isolated words were essentially equivalent for the two lexical lists. Procedure The words were stored on the hard disk of a personal computer (Gateway E-3111) . An in-house software program was developed to deliver the words randomly via a digital signal processing board (Starky HW DWP) during each test trial. The software program automatically adjusted the intensity level of the signal depending on the strategy for the specific adaptive procedure . Subjects sat in a double-walled sound attenuation room for the test sessions . A routine audiometric examination was conducted to determine auditory thresholds for pure-tone stimuli followed by measurement of spondee threshold and recognition ability for NU-6 words. For the experimental tests, the speech signal was delivered to one channel of an audiometer (GrasonStadler 16), where the overall output level of the signal could be adjusted, and then passed to a single earphone (TDH-50 encased in a model 51 cushion) . The speech level was specified using a calibrated tone in a NBS 9A coupler and the output read on a sound level meter (LarsonDavis 800B) set for a "C" scale reading. The subjects responded verbally after each word presentation, and the tester pressed designated keys on the keyboard for a correct or incorrect response, which then initiated a new trial. The 238 entire protocol was conducted within a 2-hour test session with several rest periods . As indicated earlier, adaptive procedures were used to determine the speech performance levels that estimated 29 .3, 50 .0, and 70 .7 percent correct on the individual response curve. For estimating 50 percent correct, a simple up-down procedure (Levitt, 1971) was used . Transformed up-down methods, following the strategies described by Levitt (1971, Table 1), were used to estimate the 29 .3 or 70 .7 percent responses on the curve . When applying the simple up-down procedure to estimate 50 percent correct, the stimulus level was decreased by a fixed amount if a positive response was obtained and increased if a negative (incorrect) response occurred . For the 29 .3 percent estimates, stimulus level was increased after two incorrect responses and decreased either following a correct response or an incorrect/correct response sequence . For the 70 .7 percent estimates, stimulus level was increased following either an incorrect response or a correct/incorrect response and decreased following two correct responses. For both the simple and transformed up-down procedures, 20 of the 75 words available (either the easy or hard lists) were used to determine an initial starting level that was the best prior estimate of the intended outcome. A large step size of 6 dB was used for this portion of the test, followed by initiation of the experimental sequence in which a 3-dB step size was employed . For the experimental sequence, the initial run (a run is a sequence of changes in stimulus level in one direction only) was disregarded by the program, and the final estimate (calculated by the software program) was based on the results from the remaining runs . The test was terminated after responses were made to the remaining 55 of the 75 words available. Generally, a final estimate was based on results from 12 or more runs, slightly more for the simple up-down procedure and less for the transformed up-down procedure. Although this procedure was not the most efficient use of the adaptive strategies, it was decided to administer all of the words available to estimate a particular point on the response curve. Hearing-impaired listeners often have particular difficulty perceiving certain phonemes (e .g., unvoiced fricatives) . If a particular randomization resulted in a sequence of words heavily weighted with these phonemes, the final estimate might have had an underlying bias . Thus, all 75 words from either the easy or hard lists were used to obtain the final estimate level. Because the number of available Effects of Lexical Factors/Dirks et al CVCs that fell within the lexically easy or hard criterion was modest (114 easy words, 123 hard words), it was not possible to have a sufficient number of different words for estimation of all three performance levels on the response curve . As a consequence, estimates for 29 .3 percent correct (where many words were missed) was always measured initially, followed by estimates of 50 percent correct and finally 70 .7 percent correct . This procedure was followed to minimize contamination of the estimates by correct guesses based on knowledge of the test items . The rationale was that an item that is heard correctly at a low level will generally be understood at a higher stimulus level, so that the subject's estimate should not be significantly changed at one stimulus level by the fact that he or she has already identified the items at a lower presentation level . The presentation of the hard or easy list was alternated among subjects so that half started with the easy list while the other half heard the hard words . For each estimate, a different randomization of the words was used . To summarize, each subject provided an estimate for the easy and the hard words to obtain the stimulus levels for 29 .3, 50 .0, and 70 .7 percent correct on the individual response curve . Data Analysis The overall analysis was conducted on the measured performance levels and consisted of a 2 (group : normal hearing, hearing impaired) x 2 (lexical category : easy, hard words) x 3 (per- formance conditions : 29 .3, 50 .0, and 70.7 percent) mixed analysis of variance (ANOVA) . A post hoc pairwise contrast analysis was carried out to determine if the differences between easy versus hard words were significantly different for either or both the normal and hearing-impaired listeners at each performance level . Finally, a 3 (performance level) x 2 (groups) mixed ANOVA was conducted on the difference scores between easy and hard word recognition to determine if there was any statistical difference in performance levels between normal and hearingimpaired subjects . Statistical significance was evaluated at an alpha level of .05 . RESULTS igure 1 shows the overall mean perforF mance levels for easy and hard words at the 29 .3, 50 .0, and 70 .7 percent locations on the performance-intensity functions. The results for normal-hearing listeners are illustrated in panel A and for subjects with hearing impairment in panel B . Three hearing-impaired listeners were unable to complete the test for the 70 .7 percent hard word condition . The maximum output of our test system was set at 110 dB SPL to minimize presentation of speech at levels that would be uncomfortable for hearingimpaired subjects . During the "up" sequences required by the adaptive strategy for the 70 .7 percent correct condition, the saturation level of the equipment was occasionally reached for these subjects, and the program could not be completed . The ages of these subjects were 54, 77, and 79 years, and the speech thresholds were among the highest in the group (38-52 dB) . Inspection of the individual data indicated that performance was, on average, measured within the linear portion of the response curves for the hearing-impaired and the normal-hearing listeners, that is, the decibel difference between results from 29 .3 to 50 .0 percent are practically the same as those measured between 50 .0 and 70 .7 percent . We assumed that a similar growth function applied to the aforementioned three subjects and that the test at 70 .7 was incomplete because of limitation in the intensity range available in the equipment . The individual difference levels between 29 .3 and 50 .0 percent conditions for each subject were calculated . This difference score was then applied to the level measured at 50.0 percent to estimate the result for the 70 .7 percent condition . For the remainder of the analysis, these three estimated levels were used to form a complete and more appropriate set of data. Table 4 shows the means and standard deviations for the experimental conditions for both normal and hearing-impaired subjects . As anticipated, from earlier experimental results of the effects of lexical characteristics on word recognition (Lute, 1986 ; Luce and Pisoni, 1998 ; Dirks et al, 2001), mean performance levels were always lower for the easy as compared with the hard words at each experimental condition and for both groups of subjects . As shown in Table 4, the difference in performance levels between easy and hard words ranged from 3 .0 to 5.8 dB, averaging approximately 4.0 dB . Thus, the difference in word recognition for easy versus hard words is a consistent and robust effect and can be generalized to listeners with hearing impairment and normal-hearing listeners. As shown in Figure 1, the performance-intensity functions, estimated from the mean scores at each condition, are characterized by much steeper growth functions for normal-hearing listeners 239 Journal of the American Academy of Audiology/Volume 12, Number 5, May 2001 teners (as also reflected in the observed difference in slopes of the performance-intensity functions between the two experimental groups). Separate 2 (easy, hard words) x 3 (performance conditions : 29 .3, 50 .0, 70 .7%) ANOVAs were also conducted on data for normal and hearing-impaired listeners. The results demonstrated a significant difference between easy and hard word recognition for normal listeners (F = 178 .53, df = 1, 19, p < .001) and for hearing-impaired listeners (F = 34 .03, df = 1, 19, p < .001), together with a significant difference among the measured levels for the performance conditions of normal (F = 136.8, df = 2, 38, p < .001) and hearing-impaired listeners (F = 132.26, df = 2, 38, p < .001) . A post hoc paired contrast analysis showed that the difference between easy and hard words at each performance level was significantly different for both normal (F = 30 .69 [70.7%1, 58 .0 [50.0%1, 70 .0 [29.3%1, p < .001) and hearing-impaired listeners (F = 21 .2 [70.7%1, 12 .8 [50 .0%1, 9.94 [29.3%1, p < .005). In a final ANOVA, the results showed that the mean difference scores (hard and easy word recognition) at each performance level were not significantly different between normal and hearing-impaired subjects, even though the difference score was always slightly larger for the hearing-impaired versus the normal-hearing listeners (see Table 4) . To examine the effects of age on the ability to recognize easy and hard words, the results from the hearing-impaired listeners were subdivided between measured performance levels for the middle-aged (<60 years) and old (>60 years) subjects . The average age of the middleaged group was 50 years. Only one subject was younger (23 years of age) than 40 years. The average age for the older group was 75 .1 years. The pure-tone and speech audiometric results from these subgroups have already been described (see Table 1) . Table 5 shows the results for easy and hard word recognition for the young and middle-aged listeners. The middle-aged 0 50 .0 29 .3 Figure 1 Mean word recognition performance for normal-hearing (A) and hearing-impaired listeners (B) at three locations (29 .3, 50.0, 70 .7 percent correct) on the performance-intensity function for lexically easy and hard words. than for listeners with mild-to-moderate sensorineural hearing loss . For normal-hearing listeners, the slope of the performance-intensity function (between 70 .7 and 29 .3%) for easy words was 4.1 percent per dB and for hard words was 3.8 percent per dB . For listeners with hearing impairment, the slope of the response curve for easy words was 2.1 percent per dB and for hard words 1.9 percent per dB . Similar results showing a much reduced growth function for monosyllabic words for individuals with sensorineural hearing loss versus normal-hearing listeners have been previously reported (Wilson et al, 1976 ; Beattie and Warren, 1983 ; Beattie, 1989). The general observations concerning easy versus hard word recognition were confirmed by a 2 x 2 x 3 mixed ANOVA with groups (normal, hearing impaired), lexical category (easy, hard), and performance conditions (29.3, 50 .0, 70 .7%) as within-subject variables. The results showed main effects for groups (F = 171.37, df = 1, 38, p < .001), for lexical category (F = 101.19, df= 1, 38, p < .001), and for performance condition (F = 246 .1, df = 2, 76, p < .001) . There was a significant group by performance interaction (F = 26 .44, df = 2, 76, p < .001) owing to differences in the rate of increase in word recognition for normal and hearing-impaired lisTable 4 Percent Correct Performance Levels (Means and SDs) for Easy and Hard Word Recognition Among Normal and Hearing-Impaired Listeners at Each Experimental Condition Normal Hearing Stimuli Easy words Hard words Difference scores (hard-easy) 240 70 .7 Hearing Impaired 50.0 29 .3 70 .7 50.0 29.3 35 .3 (3 .8) 39 .3 (3 .6) 29 .4 (3 .4) 33 .1 (3 .6) 25 .2 (3 .8) 28 .4 (3 .5) 72 .7 (13 .3) 78 .5 (14 .2) 62 .8 (10 .3) 67 .0 (11 .5) 53 .4 (9.5) 56 .5 (10 .9) 4 .0 (3 .2) 3 .7 (2 .2) 3 .2 (1 .7) 5 .8 (5 .6) 4 .2 (5 .2) 3 .1 (4 .4) Effects of Lexical Factors/Dirks et al Table 5 Mean Performance Levels (Means and SDs) for Easy and Hard Word Recognition for Middle-Aged (<60 Years) and Old (>60 Years) Hearing-Impaired Listeners Group/Condition Middle-aged 70 .7* 50 .0 29 .3 70 .7 Older Difference (older-middle-age) Hard Easy 50 .0 29 .3 70 .7 50 .0 29 .3 69 .3 59.5 51 .4 76 .0 (12 .3)t (8 .6) (6 .9) (14 .1) 66 .2 (11 .1) 55 .4 (11 .6) 6 .71 6 .7 4.0 Difference (Hard-Easy) 74 .7 (13.3) 63 .2 (10 .4) 54 .2 (8 .4) 83 .2 (15 .8) 5 .4t 3 .7 2 .8 7 .2 70 .7 (11 .7) 4.5 58 .8 (13 .0) 3 .4 8 .5 7 .5 4.6 *Percent correct recognition ; rdB SPL, 'relative dB, group showed slightly better performance levels than the older group for each easy and hard word recognition condition . As described in Table 5, the differences in performance between the old and middle-aged subjects ranged from 4 .0 to 8 .5 (consistently smaller for the middle-aged group) . The difference in results between the middle-aged and older subjects is likely related to the more sensitive average pure-tone and speech recognition thresholds observed in the middle-aged group . The spondee thresholds were 3 .1 dB more sensitive for the middle-aged than the older group . In addition, pure-tone thresholds above 2 .0 kHz were slightly higher for the older than the middle-aged group . Of somewhat greater interest in the current investigation were the differences in hard and easy word recognition for the middle-aged and older subjects (see Table 5) . Although the difference scores were consistently larger for the older population than the middle-aged group, over the three performance conditions, these differences averaged only 1 .0 dB . To assess these differences, while adjusting for the difference in hearing loss between the groups, an analysis of covariance was conducted between the scores for the middle-aged and older hearing-impaired subjects with pure-tone average (average of thresholds for 0 .5, 1 .0, 2 .0, and 3 .0 kHz) as the covariate . The results (F = 0 .26, df = 1, 17, p = .617) indicated that the difference in performance (hard-easy) between the middle-aged and older subjects was not statistically significant . Interestingly, the difference scores between hard and easy word recognition among the normalhearing listeners (see Table 4) were also simi- lar to those observed in both subgroups of hearing-impaired listeners (with the possible exception of the 70 .7, hard word condition for the older subjects) . There was a tendency for larger hard-easy difference scores among the older as compared with the middle-aged hearingimpaired subjects or normal-hearing listeners . However, statistical analysis of the data for hearing-impaired subjects did not support the hypothesis that the effects of lexical difficulty were disproportionately greater for older than younger subjects . DISCUSSION he purpose of this investigation was to T examine the relative performance of normal and hearing-impaired listeners for lexically easy versus hard word recognition . As predicted by the NAM theory, the current results verified earlier findings that both normal-hearing listeners and individuals with acquired hearing loss were more accurate at recognizing lexically easy than hard words. An earlier investigation (Dirks et al, 2001) from this laboratory, conducted at fixed presentation levels, indicated that the effects of neighborhood structure and word frequency applied similarly to normal-hearing and elderly listeners with acquired hearing loss . The current data demonstrate that the same effects are generalizable to middle-aged subjects and elderly individuals with hearing loss . Specifically, the effects of lexical difficulty favoring easy over hard word recognition averaged 4.0 dB when performance was measured at the locations of performanceintensity functions not limited by ceiling or floor effects. Thus, despite reduced audibility and other processing difficulties associated with acquired sensorineural hearing loss, word recognition was influenced by neighborhood structure and word frequency factors in essentially the same manner for persons with hearing loss as for young listeners with normal hearing. 241 Journal of the American Academy of Audiology/Volume 12, Number 5, May 2001 Traditionally, word recognition tests, such as the NU-6 or CID W-22, have been designed primarily with the assumption that phonetic or phonemic discrimination was the critical dimension in the assessment of speech recognition ability. Without denying the significance of phonemic discrimination ability, the results from the current and other experiments on NAM suggest that lexical factors also play an important role in word recognition. The magnitude of the effect of lexical factors on word recognition relative to acoustic-phonetic factors is not entirely clear. For a more comprehensive and valid assessment of word recognition, however, future developments of speech materials should include consideration of higher-level properties (lexical factors) on "bottom-up" acoustic-phonetic processing. Quantitative values for neighborhood density and frequency and word frequency are available, which make it possible to incorporate these factors in future developments of word recognition tests. Earlier findings from Sommers (1996) suggested that neighborhood structure, as defined by the NAM, had differential effects on recognition in older than younger adults with normal hearing. Specifically, those results showed an age-related deficit in the ability to identify hard words for conditions in which performance between the older and younger adults was essentially identical for easy word recognition . Although there was a trend, in the current study, for larger differences between performance for easy versus hard words for the older subjects with hearing loss, there was no statistical difference between performance for the older than for the middle-aged adults with acquired hearing loss . This result is similar to earlier findings (Dirks et al, 2001) from this laboratory, which showed that older subjects with hearing loss performed in essentially the same manner on lexically based word lists as young normal-hearing listeners once speech presentation level was adjusted for the hearing loss in the former. Because of differences in design and subjects, the results from our laboratory and the findings of Sommers (1996) are not straightforward . Several of these differences are described in detail because the issues involved impact other related studies in this investigative area . First, there are differences in the characterization of easy and hard words, not only between the current study and Sommers' (1996) investigation but also more generally among experiments in which easy versus hard word 242 recognition has been the topic for study. In the current investigation, easy and hard words were differentiated on the basis of three lexical factors : word frequency of the target, neighborhood density, and average neighborhood frequency (easy words had a high word frequency count with low neighborhood density and frequency, whereas hard words had low word frequency and high neighborhood density and frequency) . This characterization is the same as was reported by Toretta (1995) in the development of a multitalker speech database of easy and hard words and was also used in the investigation of Bradlow and Pisoni (1999) . Sommers, however, emphasized differences in neighborhood structure, which are the central factors in the NAM theory, and essentially controlled for word frequency. Although the word frequency of the target words was not specifically reported in the Sommers' paper, a later report (Sommers, 1998), in which the same word sets were used, indicated that the frequency of the easy words was 51 .2, whereas the frequency of the hard words was 43 .4 (Sommers, 1998). The intent was most likely to control the word frequency factor and vary the neighborhood structural characteristics of the easy and hard words . Analysis of results from Dirks et al (2001) indicated that each of the three lexical characteristics do not affect word recognition independently but rather that the words are recognized relationally, that is, the factors covary. Those results indicated that word frequency accounts for a larger proportion of the variance among the experimental word lists (each containing different combinations of the lexical factors) than did neighborhood density and mean neighborhood frequency. Because the word frequency count for easy and hard words was controlled in the Sommers (1996) investigation, its effects on easy versus hard word recognition may not be the same as in the current investigation or that of Bradlow and Pisoni (1999) . Because of the differences in characterization of easy and hard words among investigators, there are some limitations in comparing results among the experiments where this classification system has been used . Whether the differences in the lexical characterization of easy and hard words among the investigations are related to the finding of an age-related deficit in hard word recognition is unclear but provides an interesting topic for future research . Second, the populations of the adults who served as subjects in these studies also differed . Sommers (1996) compared recognition between Effects of Lexical Factors/Dirks et al younger and older adults with essentially normal hearing (however, the older subjects did have reduced hearing relative to the younger group at frequencies above 1 .0 kHz) . In the current study, the age-related issue was primarily assessed by comparing recognition performance between older and middle-aged adults with hearing losses . Attempts to develop a group of elderly subjects with normal hearing often require a wider tolerance of hearing threshold than for a younger subject population . The presence of a high-frequency loss among the elderly group often compromises the audibility of some phonemic elements among these listeners . A third issue concerns differences in the phonemic structure of the words used in studies in which easy and hard recognition has been assessed . In the current study, the phonemic structure for the easy and hard words was controlled within limits imposed by the small CVC database available from the corpus of words that conform to the lexical criteria . Phonemic control was also imposed for the word sets used by Sommers (1996) . As indicated previously, because of differences in the definition of easy and hard words, the words available and chosen by Sommers as easy and hard were, in general, different than those available for the current investigation . The words themselves and their phonemic structure are understandably somewhat different between the two studies . Whether the phonemic structure of the words interacts with lexical characteristics to influence intelligibility is not entirely clear, but, especially when adults with hearing loss are used as subjects, this issue needs to be addressed . In summary, an investigation was conducted to determine the effects of several lexical properties of words on recognition . The words chosen were divided into two sets of lexically easy and hard words based on the NAM theory and word frequency. Adaptive procedures were used to measure performance for these word sets at several locations on the linear portion of the respective performanceintensity functions for normal and hearingimpaired listeners . The findings verified the NAM predictions, demonstrating that easy words produce higher recognition performance than hard words and that the results are generalizable to middle-aged and older adults with acquired hearing loss . The results indicate that word recognition is influenced by the lexical properties of words and the more well-known acoustic-phonemic properties . Acknowledgment . We thank P Douglas Noffsinger from the Veterans Administration Greater Los Angeles Healthcare System and Stephan Fausti, director for National Center for Rehabilitative Auditory Research, Veterans Administration Medical Center, for their support and helpful comments on this research . We are grateful to the audiologists at Veterans Administration Greater Los Angeles Healthcare System, West Los Angeles Medical Center, and the UCLAAudiology Clinic for assistance in recruiting subjects with hearing impairment . We also thank Richard Wilson and Anne Strouse for several helpful suggestions during the review. This work was supported by grants from the Department of Veterans Affairs Rehabilitative Research and Development Service to the National Center for Rehabilitative Auditory Research (RCTR 597-1060), a Veterans Affairs Merit Review Award (C2225R), and the Hope for Hearing Research Foundation. Portions of this article were presented at the Veterans Affairs Rehabilitative Research and Development Service Second Annual Meeting, Crystal City, MD, February 2000 . REFERENCES Beattie RC . (1989) . Word recognition functions for the CID W-22 test in multitalker noise for normally hearing and hearing-impaired subjects . J Speech Hear Disord 54 :20-33 . Beattie RC, Warren V (1983) . Slope characteristic of CID W-22 word functions in elderly hearing-impaired listeners. J Speech Hear Disord 48 :119-127 . Bradlow AR, Pisoni DB . (1999) . Recognition of spoken words by native and non-native listeners: talker-, listener- and item-related factors . J Acoust Soc Am 106:2074-2086 . Department of Veterans Affairs. (1991) . Speech recognition and identification materials . (CD 1.1). Long Beach, CA : VA Medical Center. Dirks DD, Takayangi S, Moshfegh A, Noffsinger PD, Fausti SA . (2001) . Examination of neighborhood activation theory in normal and hearing-impaired listeners . Ear Hear (in press). Kirk Kl, Pisoni DB, Osberger MJ . (1995) . Lexical effects on spoken word recognition by pediatric cochlear implant users. Ear Hear 16 :470-481 . Kirk KI, Pisoni DB, Miyamoto RC . (1997) . Effects of stimulus variability on speech perception in listeners with hearing impairment . J Speech Lang Hear Res 40 :1395-1405 . Kucera F, Francis W. (1967) . Computational Analysis of Present Day American English. Providence, Rl : Brown University Press. Ladefoged P. (1993) . A Course in Phonetics . Fort Worth, TX: Harcourt Brace College. Levitt H . (1971) . Transformed up-down methods in psychoacoustics . JAcou.st Soc Am 49 :467-477 . Luce PA. (1986) . A computational analysis of uniqueness points in auditory word recognition. Percept Psychophys 39 :155-158 . Luce PA, Pisoni DB . (1998) . Recognizing spoken words : the Neighborhood Activation Model. Ear Hear 19 :1-36. 243 Journal of the American Academy of Audiology/Volume 12, Number 5, May 2001 Meyer TA, Pisoni DB . (1999) . Some computational analysis of the PBK tests : effects of frequency and lexical density on spoken word recognition . Ear Hear 20 :363-371 . Nusbaum HC, Pisoni DB, Davis CK. (1984). Sizing up the Hoosier Mental Lexicon: measuring the familiarity of 20,000 words. Research on Speech Perception Progress Report No . 10 . Bloomington, IN : Speech Research Laboratory, Psychology Department, Indiana University. 122-134. Sommers MS . (1996) . The structural organization of the mental lexicon and its contribution to age-related changes in spoken word recognition . Psychol Aging 11 :333-341 . Sommers MS . (1998) . Spoken word recognition in individuals with dementia of the Alzheimer's type : changes in taker normalization and lexical discrimination . Psychol Aging 13 :631-646 . Sommers MS, Kirk KI, Pisoni DB . (1997) . Some considerations in evaluating spoken word recognition by normal-hearing, noise masked normal-hearing, and cochlear implant listeners. I: the effects of response format. Ear Hear 18 :89-99 . Torretta GM. (1995) . The easy-hard word multi-talker speech database : an initial report. Research on Spoken Language Processing, Progress Report 20 . Bloomington, IN : Indiana University, 321-334. Wilson RH, Coley KE, Haenel JL, Browining KM . (1976) . Northwestern University Auditory Test No . 0.6 : normative and comparative intelligibility functions. JAm Audiol Soc 1:221-228 . APPENDIX Stimulus List Easy Words boss cause chain cheap check church coach death deck deep dirt dive does dog down feed firm food gang gap gas girl give god hung join joke judge june king learn leg live lodge long loud love mouth move neck noise nurse page palm path peace pool pull put reach real ridge rob roof rough sang shape Hard Words shirt shop soil tape teeth thick thing tongue tough turn vain vote wash watch wife work wrong young bait ban bead beak boot cake chat cheer chill chore cod con cone doom dot dumb dune fade fake fin goat gore gut hash hick but lad lame lice lime mace mare mice mid moan mum nut pawn pin rum rut sane sill soak soar suck tan ear tile ton tune wad keen kin kit knit raid rat rhyme rim wail watt weed wick cot jot dill lace dad dame debt den pun roar wade