Effects of Hearing and Aging on Sentence-Level Time-Gated Word Recognition

Abstract

Purpose

Aging is known to influence temporal processing, but its relationship to speech perception has not been clearly defined. To examine listeners' use of contextual and phonetic information, the Revised Speech Perception in Noise test (R-SPIN) was used to develop a time-gated word (TGW) task.

Method

In Experiment 1, R-SPIN sentence lists were matched on context, target-word length, and median word segment length necessary for target recognition. In Experiment 2, TGW recognition was assessed in quiet and in noise among adults of various ages with normal hearing to moderate hearing loss. Linear regression models of the minimum word duration necessary for correct identification and identification failure rates were developed. Age and hearing thresholds were modeled as continuous predictors with corrections for correlations among multiple measurements of the same participants.

Results

While aging and hearing loss both had significant impacts on task performance in the most adverse listening condition (low context, in noise), for most conditions, performance was limited primarily by hearing loss.

Conclusion

Whereas hearing loss was strongly related to target-word recognition, the effect of aging was only weakly related to task performance. These results have implications for the design and evaluation of studies of hearing and aging.

Accurate processing of ongoing speech requires rapid identification of words in sequence, often in less than ideal listening conditions. Fundamental to current models of spoken word recognition is the notion that information from phonemes near word onset leads the listener to make inferences about a word's identity in the context of other phonemically similar words (Dirks, Takayanagi, Moshfegh, Noffsinger, & Fausti, 2001; Marslen-Wilson, 1987; Marslen-Wilson & Welsh, 1978; Slowiaczek, Nusbaum, & Pisoni, 1987). For this reason, deficits in the rapid identification of the initial phonemes in spoken utterances could have a relatively larger impact on the ability to understand continuous speech than would result in difficulties identifying later portions of spoken words. The work presented here is an attempt to better understand the potential influences of aging and hearing loss on the ability to identify words based on linguistic context and on the evolving acoustic information available in progressively longer segments of speech.

In 1980, Grosjean introduced the time-gated word (TGW) paradigm as a way to assess “real-time” perceptual encoding of isolated words. The paradigm involves “gating-in” successive segments of the target word starting at its onset and asking listeners to identify the target based only on the initial phonemic information. Increasingly longer segments (or gates) are presented until the entire word is presented (or fully gated-in). Isolation point (IP) refers to the gate at which the listener initially correctly identifies the target word (presumably by isolating it from the competing possibilities in their lexicon). Several studies using the TGW paradigm have found that older listeners require significantly greater amounts of the target word to be gated-in (i.e., a greater IP) in order to achieve accurate identification (Craig, 1992; Craig, Kim, Rhyner, & Chirillo, 1993; Elliott, Hammer, & Evan, 1987).

To investigate the effects of higher level linguistic cues and aging on word recognition, Craig et al. (1993) used key words from the Revised Speech Perception in Noise (R-SPIN) test (Bilger, Nuetzel, Rabinowitz, & Rzeczkowski, 1984) as TGW targets. Each list in the R-SPIN test is composed of 50 sentences in which the target word is the final word in the sentence. In high-context (HC) sentences, target words are highly predictable based on the preceding context (“Stir your coffee with a spoon”), whereas in low-context (LC) sentences, the target word cannot be predicted based on the preceding context (“Bob could have known about the spoon”). They found that age-associated differences in IP were small but reliable, and older listeners were more likely to fail to identify a target word even after targets were fully gated-in. However, these results are difficult to interpret because the younger participants had pure-tone audiometric thresholds of 15 dB HL or better through 6000 Hz, whereas older listeners had average thresholds that were substantially higher at 4000 and 6000 Hz.

A handful of recent studies have employed the gating paradigm to investigate identification of initial consonants, isolated words, and final words in low- and high-predictability sentences in older hearing aid users or older listeners with nominally normal hearing (pure-tone average or PTA ~18 dB HL; Moradi, Lidestam, Hällgren, & Rönnberg, 2014) and young listeners (thresholds not provided) in quiet and noise (Moradi, Lidestam, & Rönnberg, 2013; Moradi, Lidestam, Saremi, & Rönnberg, 2014). The results demonstrated longer IPs in noise for the younger listeners and for the older hearing aid users in the low-predictability sentences compared with the older listeners with normal hearing. Although no direct comparisons were made between the younger and older listeners, for the low-predictability sentences in quiet, average reported IPs were shortest for the young listeners and longest for the older hearing aid users, with the older listeners with normal hearing falling between. Differences in IPs were minimal among the groups for high-predictability sentences. In addition, no significant correlations were observed between IPs for final words in sentences and measures of working memory or attention for younger or older listeners.

The understanding of linguistically simple speech presented at a normal rate in quiet is reasonably well predicted by the audiogram (Humes, 1996). However, the audiogram is less effective at predicting speech intelligibility in the presence of environmental factors such as reverberation or background noise; stimulus factors such as rapid or unclear conversational speech; or in cases when other listener attributes, such as age or cognitive function, are also varying. For example, when compared with young adults with similar pure-tone thresholds, older adults have particular difficulty understanding speech in the presence of noise (Dubno, Dirks, & Morgan, 1984; Pichora-Fuller, Schneider, & Daneman, 1995; Plomp & Mimpen, 1979) or presented rapidly, as with time-compressed speech (Gordon-Salant & Fitzgibbons, 2001; Konkle, Beasley, & Bess, 1977; Pichora-Fuller, 2003). One factor hypothesized to account for the speech-understanding differences observed between older and younger listeners with similar thresholds is impaired temporal processing in the older auditory system (Humes, 1999; Jerger, 1973; Pichora-Fuller & Souza, 2003; Schneider, Speranza, & Pichora-Fuller, 1998; Wiley et al., 1998). As described in Moore (2014), there are two main time scales at which difficulties with auditory temporal processing could potentially arise: temporal fine structure, which usually refers to the very fast changes in the waveform of speech used to convey information such as pitch and interaural time differences, and envelope information, which relates more strongly to the boundaries between syllables. Gallun et al. (2014) examined sensitivity to what could be described as temporal fine structure and found impacts of age and hearing loss for both monaural and binaural abilities. Although there is currently little strong evidence relating such basic abilities to speech understanding, this is an important direction for future work.

An additional factor contributing to the difficulty older listeners have understanding speech may be slowing of nonauditory cognitive processes needed to comprehend speech (Salthouse, 2000). Consequently, variability in the extent to which previous studies have accounted for declines in hearing or cognition associated with normal aging may explain why age is correlated with speech-understanding ability in some studies but not others. Furthermore, the type of speech stimuli used can have a strong impact on whether or not such cognitive abilities account for substantial portions of the variance among listeners. Here, we employ three approaches to address this issue. First, participants were recruited such that the maximum degree of hearing loss was limited but not in a way that reduced the range of ages tested or that produced a group of older listeners with uncharacteristically good hearing. This was then combined with a continuous analysis via linear regression rather than a group analysis, which allowed the correlation of age and hearing ability present within the sample to be accounted for statistically rather than minimized in a group analysis in which the older group had nominally normal hearing that was nevertheless substantially worse than the younger group. The second approach was to assess working memory and the general mental function on all of the participants to have a quantitative measure of cognitive ability, as well as recruitment criteria that ensured generally normal mental function in all participants. The third aspect of the study that was used to examine cognitive influences on performance was the use of both HC and LC sentences. By comparing the relationships between age, hearing loss, and performance in the HC and LC conditions, it becomes possible to better distinguish issues of basic intelligibility, which dominate performance in the LC conditions, from the ability to predict an unheard or misunderstood word, which will dominate performance in the HC conditions.

The investigation described here consisted of two experiments. In Experiment 1, a TGW recognition task was developed and verified using the R-SPIN corpus (Bilger et al., 1984) following Craig et al. (1993). To allow for comparisons without potential confounds of word length or variability in individual target-word difficulty, two sets of HC and LC R-SPIN sentence lists were matched on these attributes. This first experiment was essential to the development of a set of materials that could clearly distinguish the impacts of intelligibility, as captured in the LC conditions, from the effects of linguistic prediction, as captured in the HC conditions. In Experiment 2, the matched sentence lists were used to characterize the effects of aging on word recognition in both quiet and noise while accounting for differences in hearing sensitivity among participants. Results of this experiment were assessed via a statistical model—similar to one employed by Gallun et al. (2014)—to address the correlation of age and hearing loss.

Experiment 1: SPIN-TGW Test Development

Method: Phase 1

Listeners were tested using a TGW task in which the time-gated targets were key words from a large sample of R-SPIN sentences. The results are described in terms of average segment duration required to accurately identify the target word (IP). The goal of the first phase of the experiment was to characterize expected performance for future participants and thus allow the construction of two sentence lists with different words but similar expected IP values.

Participants. Six young listeners with normal hearing between the ages of 23 and 27 years (M = 26.3 years) participated in Phase 1. All had audiometric thresholds ≤20 dB HL at all octave and interoctave frequencies from 250 to 8000 Hz in the test ear. Tympanometry was performed to rule out middle ear abnormalities, and no more than one air–bone gap greater than 10 dB was present at audiometric octave frequencies from 500 to 4000 Hz. Additionally, all participants achieved a passing score (24 or above) on the Mini-Mental State Examination (MMSE; Folstein, Folstein, & McHugh, 1975) to rule out effects of a major cognitive deficit. Participants were tested monaurally in the better-hearing ear, or the right ear when thresholds were similar between ears. For all experiments in this study, individuals consented to participate in accordance with the Institutional Review Board procedures at the Portland VA Medical Center.¹

Stimuli. A TGW task was developed using 104 sentences selected from Sentence Lists 1, 3, and 5 of the R-SPIN corpus. Similar to Craig et al. (1993), the final word in each sentence served as the time-gated target. All target words only appeared once, either in HC or LC sentences. Two example R-SPIN sentence waveforms are provided in Figure 1, with target waveforms shown at different gates. The sentence on the left is HC, whereas the sentence on the right is LC.

Figure 1.

High context (left) and low context (right) Speech Perception in Noise–time-gated word sentence waveforms shown with increasing gatings. Sentences were presented in background noise (not shown). Unaltered (fully gated) waveforms are shown in the top panels.

To create the speech stimuli, sentences were converted from the R-SPIN test CD to individual sentence wave files using Adobe Audition (Adobe Systems, San Jose, CA). Sentences were low-pass filtered with a 4-kHz cutoff (~24 dB/octave reduction) to reduce artifactual aliasing in the original stimuli and to ensure that threshold differences above 4 kHz among the participants in Experiment 2 would not impact TGW performance.

A system of three-way interrater judgment was used to determine the precise onset and offset of each target word. Two primary raters independently judged target onsets and offsets and compared their judgments. In cases of disagreement, a third rater helped form group consensus. Time-gated targets were created by gating-in only a brief segment of the target word beginning at its onset (i.e., at 0 ms). The segments were successively incremented until the gate contained the entire word; the initial gating increments were 12, 25, and 50 ms, and thereafter increments were increased by 50 ms. As a result, each target word had a total number of gates that varied from six to 24, depending on the fully gated-in word length. The majority of targets had between 10 and 14 total gates available.

All conditions in Experiment 1 were presented in background noise. The multitalker babble distracter from the R-SPIN test CD was replaced with continuous speech-shaped noise spectrally matched to the SPIN sentences used in this study. This reduced the chance that brief fluctuations in the masker would influence the amount of masking for the very brief targets. The noise was generated offline using Matlab (Mathworks, Natick, MA). A Fourier analysis of the entire sentence corpus was calculated, and then the resultant individual frequency components (complex vectors) were phase-randomized without altering their amplitude vectors. Finally, the entire complex of phase-randomized Fourier components was returned to the time domain via inverse Fourier transform. This process resulted in a sound file containing 30 s of continuous noise that was spectrally matched to the sentence corpus. Each sentence was presented with a 3-s noise masker drawn from a random starting point within the 30-s noise file. Noise maskers started 500 ms before R-SPIN sentence onset and ended approximately 500 ms after sentence offset, depending on sentence duration, such that each sentence presentation was completely embedded in noise.

Procedure. Listeners were tested individually in an Acoustic Systems RE-245S Double Wall sound-attenuated booth (Acoustic Systems, Austin, TX). Stimuli were controlled via a Windows desktop computer using E-Prime software (Psychology Software Tools, Sharpsburg, PA) and were presented monaurally to the better ear through an Etymotic ER-2 insert earphone (Etymotic Research, Elk Grove Village, IL). Sentences were presented at an overall level of 80 dB SPL. Background masking noise was presented at 75 dB SPL, resulting in a +5 signal-to-noise-ratio (SNR) for all experimental presentations.

The 104 R-SPIN sentences, adapted for the TGW task, were divided into four lists of 22 sentences each and one list of 16 sentences. Each of these lists contained an equal number of HC and LC sentences. Twelve additional sentences from the R-SPIN List 1 were reserved for a practice condition. Participants were provided with detailed instructions and practice prior to testing. Practice testing was completed as follows: (a) two nongated sentences in quiet (one HC/one LC); (b) five nongated sentences in background noise (two HC/three LC); and (c) five gated sentences in background noise (two HC/three LC). After completing the practice lists, participants listened to the five experimental sentence lists in random order. They were encouraged to identify the final target word in each sentence as best they could with the limited speech information heard and to rate the confidence of their responses on a numerical scale from 1 to 10, where 1 indicated they were not confident and 10 indicated they were very confident. Participants were instructed that all targets were real English words. They were further instructed that the same sentences could be repeated several times even if they had already correctly identified the target word, but that the task was always the same—to try to correctly identify the target word on each presentation. Participants were informed that with each repetition of a given sentence, the amount of speech information would be increased very slightly. If they could not formulate a guess, participants were instructed to say “I don't know” and were not required to provide a confidence rating. Target-word responses and confidence ratings were given verbally by participants and were scored via a custom PC interface in real time by the experimenter. Although we elicited confidence ratings from respondents in this experiment, the ratings were only recorded to evaluate the software performance and were not evaluated. Confidence ratings will be addressed in Experiment 2.

Sentences within each list were presented using a “duration blocking” paradigm in which different blocks corresponded to different gating lengths (target-word segment durations), starting with the shortest gate. Within a block, sentences were presented at random and then rerandomized with each increase in gating length. Walley, Michela, and Wood (1995) found no difference between duration blocking and a “successive blocking” method in which a target segment is progressively increased until the entire word is gated-in. In order to shorten the test time, relative to procedures used in previous work, listeners were not required to respond to all of the possible gates for each target. Rather, target-word duration required for correct recognition at two consecutive gates was determined, and the shorter of these two gates was taken as the IP. Once IP was achieved, the sentence was no longer presented in subsequent blocks. The impact of this change on the duration blocking method is that the later blocks contained comparatively fewer sentences than the earlier blocks due to the elimination of target words that had already been correctly identified on two successive previous blocks. If listeners did not correctly identify the target word until the target was fully gated-in, they were provided with a second chance to correctly identify the fully gated-in target word in the subsequent gating block. If, at that point, IP was achieved, it was recorded as the target-word length. For words not correctly identified even after being fully gated-in, a separate index recorded failure to achieve IP.

Rules were developed for scoring listener responses that approximated the target word. For example, a response that incorporated the target and an additional phoneme at the end of the word was scored as a correct response if it did not change the meaning of the target word (such as “oxen” for the target word “ox”). In addition, plurals or verb tense suffixes (e.g., -ing) added to a target word were scored as a correct response. However, additions, alterations, or deletions of phonemes that altered the meaning of the root target word were scored as an incorrect response. For example, “barn” for the target word “bar” or “feet” for the target word “fee” were both scored as incorrect responses even though a listener correctly identified the beginning phonemes. We reasoned that such errors would have resulted in the incorrect understanding of the meaning of the sentence in a real-world scenario.

Results: Phase 1

Effects of predictability (context) and target-word length. Figure 2 shows the means of median IP obtained for each target word across all subjects for HC and LC sentences. Most HC targets were correctly identified within the first few gates, indicating that there was enough contextual information available in the preceding words of the sentence that little or no additional information about the target was needed for it to be guessed correctly. In contrast, most participants required at least 200–300 ms gated-in for LC targets to be correctly identified. There was also greater variability among listeners on the LC target sentences than the HC target sentences. For LC sentences, targets with longer word lengths were identified at longer IPs; this was not true for HC targets. Pearson's correlation coefficient relating IP to target length was .49 (p < .001) for the LC sentences and .05 (p = .42) for the HC sentences. Although SPIN-TGW performance in the present study was examined in background noise, these IP results for HC and LC time-gated words are generally consistent with earlier observations for young adult listeners in quiet (Craig et al., 1993).

Figure 2.

Distributions of the means of median isolation point (IP; in ms) for each target word across all subjects in HC (left) and LC (right) SPIN sentences.

Results of the matching procedure. As shown in Figure 3, the matched sentence pairs were created by plotting all of the sentences in a space representing the median IP, target-word length, and context, and then finding the pairs with the smallest Euclidian distance in this space. Lines connect a best matched target pair, indicating that the two connected targets would be assigned to different lists. Each matched sentence pair was equated to be sufficiently similar with respect to context (10 HC and 10 LC each per list), target-word length, and mean IP as observed in our sample of six young listeners with normal hearing. Not all of the targets were amenable to matching, which reduced the original pool of 104 sentences used in Phase 1 to the 80 sentences comprising the SPIN-TGW corpus used in Experiment 2. One member of each matched pair was allocated to List A, the other member to List B, and each matched pair was then randomly allocated to List Set 1 or 2. This resulted in four lists, each of 20 sentences, denoted 1A, 1B, 2A, and 2B, where List 1A was matched to List 1B, and List 2A was matched to List 2B.

Figure 3.

Results of sentence matching based on sentence context, median IP, and target length. Lines connect best-matched sentences.

Method: Phase 2

IP was obtained in a separate sample of listeners using the newly developed SPIN-TGW sentence lists. Equivalent IP performance across lists was taken as evidence that the sentence list matching was effective.

Participants. A new sample of eight young listeners with normal hearing between the ages of 22 and 32 years (M = 27.2 years) participated in this phase of the study. Inclusion criteria were identical to Phase 1. The sample size was chosen via simulation to provide 86% power to reject the null hypothesis of nonequivalence among sentence lists, using the model-based approach described below.

Stimuli. The stimuli were the 80 sentences identified in Phase 1 as previously described, divided into two sets of two matched lists (1A/1B, 2A/2B). Practice conditions in Phase 2 were expanded by incorporating some of the 24 sentences tested in Phase 1 that were excluded from the matched lists. Practice conditions were completed as follows: (a) 10 nongated sentences in quiet (five HC/five LC); (b) five gated sentences in quiet (two HC/three LC); (c) 10 nongated sentences in background noise (five HC/five LC); and (d) five gated sentences in background noise (two HC/three LC). Sentences for the practice conditions and the two sets of matched lists are provided in the Appendix.

Procedure. Aside from differences in sentence materials as noted above, procedures for test administration and response recording were identical to Phase 1. Briefly, all sentences were presented monaurally in the better hearing ear (or right ear if thresholds were identical) at a level of 80 dB SPL at a +5 dB SNR.

Results: Phase 2

The goal of Experiment 1 was to construct and demonstrate the equivalence of two lists of time-gated word sentence stimuli. Effective equivalence is demonstrated if the average IPs for Lists A and B within each List 1 and 2 differ by less than the variability found for a given target word across the individual listeners. Using Phase 1 analyzed data to determine an appropriate cutoff, it was observed that group median IPs for individual HC words spanned about a 100-ms range. On the basis of this observation, a difference of 75 ms was chosen to be stricter yet to extend beyond the 50-ms gate-length increment used in all but the initial two gating blocks.

To determine whether or not the individual IP values for the matched target words in the A and B lists differed by more than 75 ms, confidence intervals of the difference were calculated for each list and for HC and LC sentences. Confidence intervals were estimated by fitting a linear mixed model to the observed IP, with list (1 vs. 2) and context (HC vs. LC) as factors. A subject-level random factor was also included. A 90% confidence interval, suitable for Two One-Sided Test (TOST) equivalence testing (Schuirmann, 1987), was computed from fitted model coefficients. A simulation approach to joint confidence-interval estimation was used to preserve a .05 nominal test level. As can be seen from the confidence intervals shown in Figure 4, IPs for matched sentences from both pairs of lists (1 and 2) differed by less than 75 ms for both HC and LC sentences. Furthermore, results indicated no evidence of systematic bias, a situation in which all listeners always performed better for a given list. This demonstrates that the matching procedure was successful and that it is appropriate to use the matched lists to compare the impact of additional factors, such as SNR, within the same listeners.

Figure 4.

Confidence intervals showing the difference in mean IP in ms. Equivalence is demonstrated by simultaneous confidence intervals for the 1A/1B and 2A/2B mean IP that are within 75-ms control limits.

Experiment 2: Effects of Aging and Hearing Loss on TGW-Recognition Performance

Method

The goal of Experiment 2 was to investigate the separate effects of aging and hearing loss on listeners' performance on a TGW task in quiet and in noise. Stimuli were the sentence lists derived from the Experiment 1 matching process.

Participants. TGW recognition performance was assessed in a sample of 61 participants ranging in age from 18 to 75 years (M = 48.4 years, SD = 16.5 years). Mean pure-tone thresholds across 500, 1000, 2000, and 4000 Hz (PTA) were 13.2 dB HL in the left ear (SD = 9.0) and 12.9 dB HL in the right ear (SD = 8.5). Participants with conductive or mixed hearing losses were excluded from the study. Participants were also excluded who had thresholds averaged across 3, 4, and 6 kHz greater than 50 dB HL. This exclusion criterion was essential for ensuring that the correlations between age and hearing were not so great that they could not practically be distinguished even with an individualized statistical modeling approach. Tympanometry was performed to rule out middle ear abnormalities, and no more than one air–bone gap greater than 10 dB was present at audiometric octave frequencies from 500 to 4000 Hz. Figure 5 displays individual and mean pure-tone thresholds for individual ears of all participants. Although age was analyzed as a continuous variable, for clarity of display, individual audiograms are presented grouped in separate panels for younger (18–40 years), middle-aged (41–60 years), and older (61–75 years) listeners. Thin lines indicate individual ears, and thick lines indicate average group thresholds. The three panels of this figure illustrate the overall changes in auditory sensitivity observed in this sample with increasing age. In the youngest group, all participants had thresholds better than 20 dB HL—a common cutoff for normal hearing—for all audiometric frequencies. Most of the middle-aged participants had thresholds better than 20 dB HL for frequencies 2 kHz or below, but overall these thresholds were higher even at lower frequencies. In the oldest group, many participants still had hearing within normal limits (i.e., ≤20 HL) at frequencies below 4 kHz, but even greater losses at high frequencies. Note that for the most part, the best thresholds in the middle-aged and older groups are only as good as the average thresholds of the younger group. Nonetheless, it is essential to note that the range of hearing across all groups was fairly small, relative to the range found in the general population, and that none of the participants would be characterized as having any more than a mild loss in the frequency regions most important for speech understanding.

Figure 5.

Individual subject (thin lines) and mean (thick lines) pure-tone thresholds as a function of frequency for the left and right ears. Results are displayed separately for the younger (18–40 years), middle-aged (41–60 years), and older (61–75 years) listeners. Shaded regions indicate 1 SEM.

As for Experiment 1, participants were required to have a passing score (24 or above) on the MMSE, a screening test for gross cognitive dysfunction. Participants' auditory working memory was also assessed using the Letter-Number Sequencing (LNS) subtest of the Wechsler Memory Scale–III (Crowe, 2000; Wechsler, 1997). All participants had age-appropriate scores on the LNS test (within 1 SD of the mean), and further, our participants displayed above-average mean scaled scores overall (12.15) compared with the mean normative score (10). Although this reduces the ability of this experiment to reveal much about the potential influences of cognition on the TGW test, it does allow the potential impacts of age and hearing loss to be more clearly revealed.

Stimuli. The two sets of matched SPIN-TGW test lists developed in Experiment 1 (1A/1B and 2A/2B) were used as stimuli, and performance was separately assessed for each participant in both ears in quiet and noise for a total of four test conditions. The same phase-randomized background noise generated for Experiment 1 was used in the noise conditions. Testing for each listening condition within each ear was completed using a matched list pair. As each pair of lists contained 20 LC and 20 HC sentences, each participant provided data for 40 targets presented to each ear: 10 LC in quiet, 10 HC in quiet, 10 LC in noise, and 10 HC in noise. For each participant, a set of matched lists (1 or 2) was randomly assigned to be presented to the left or right ear, and then the lists (A or B) were randomly assigned to be presented in quiet or noise (i.e., if List 1A was presented to the left ear in quiet, then List 1B was presented in that ear in noise). All sentences were presented at a level of 80 dB SPL, regardless of hearing thresholds. When present, background masking noise was presented at 75 dB SPL, resulting in a +5 SNR. Stimuli were presented monaurally to each ear in turn via Etymotic ER-2 insert earphones. Monthly coupler checks were performed to ensure consistent presentation levels at the output of the earphones.

Practice testing was identical to Phase 2 of Experiment 1 and was completed in the better hearing ear, or in the right ear when hearing was similar in both ears. At a minimum, listeners were required to correctly identify the two HC gated target words in the +5 SNR condition to proceed with experimental testing. All listeners passed the practice testing and did not report any discomfort with the levels, demonstrating that all understood the task, cognitive function was sufficient for performing the testing, and that the 80 dB SPL level provided sufficient audibility without introducing distortion or uncomfortably loud signals. For the experimental test sessions, assignment of matched list (1 vs. 2) and ear order (right vs. left) was randomly predetermined. Testing for an ear always began with a sentence list presented in quiet followed by testing in noise. Instructions and other procedures were identical to those used in Phase 2 of Experiment 1.

Results

Isolation point. As in Experiment 1, IP was taken as the shorter of two consecutive gates corresponding to a correct response. Mean IP for each ear for the four combinations of HC versus LC sentences and quiet versus background noise are shown as a function of age and PTA in Figures 6 and 7, respectively. Although there is quite a bit of variability in performance within each condition and a fair amount of overlap between conditions presented in quiet and in noise, there is no overlap between IPs for the LC and HC stimuli. The longest average IP for HC stimuli is still substantially shorter than the shortest IP for the LC stimuli.

Figure 6.

Mean IP in HC and LC sentences in quiet and noise as a function of age. Pearson's correlations and p values are listed by condition; significant correlations are noted with an asterisk. Data points marked with an X indicate outlier subjects not included in analyses.

Figure 7.

Mean IP in HC and LC sentences in quiet and noise as a function of pure-tone average (PTA). Pearson's correlations and p values are listed by condition, significant correlations are noted with an asterisk. Data points marked with an X indicate outlier subjects not included in analyses.

Within each listening condition, modest but significant positive correlations were observed between age and IP for both LC conditions (quiet: r = .24, p < .01; noise: r = .18, p < .05), and between PTA and IP for both LC conditions (quiet: r = .28, p < .01; noise: r = .25, p < .01) and in the HC condition in noise (r = .18, p < .05). It was determined that the correlations were substantially reduced by the presence of two younger participants (indicated by Xs on Figures 6 and 7) with unusually poor performance given their age and hearing. Adjusted correlations following the removal of these two participants revealed the same relationships as the unadjusted correlations (age and LC quiet: r = .35, p < .001; age and LC noise: r = .32, p < .001; PTA and LC quiet: r = .36, p < .001; PTA and LC noise: r = .37, p < .001; PTA and HC noise: r = .22, p < .05), with the exception that HC in quiet and PTA were now significantly correlated as well (r = .18, p < .05).

Because age and hearing is correlated in our sample (see Figure 5), it is not possible to know from the raw correlations the independent effects of each. One alternative approach often employed to distinguish among the effects of multiple predictors (each of which may be correlated with the others) is a partial-correlation analysis, in which the effects of one factor are “partialled out” to allow an estimate of the impact of the other. Although there are multiple examples in the literature in which this analysis is used to distinguish the effects of age and hearing loss on test performance (e.g., Hopkins & Moore, 2011), there are several reasons to avoid this approach. The most serious issue is that in a partial regression, performance is first averaged across multiple measurements, and this averaged value is then used to evaluate the impact of age or hearing loss on performance. This data reduction method yields two problems. First, it decreases the number of samples available for analysis, thus reducing the power of the analysis. Second, while this approach allows basic relationships to be described, it becomes impossible to communicate precisely the size of the changes in performance associated with age and hearing loss separately. Therefore, to further examine the effects of age and hearing loss, a linear mixed model of the type presented in Gallun et al. (2014) was developed and applied to the data presented here, with the two young outliers removed. A linear mixed model has the advantages of using each measurement as a data point, thereby preserving potential power of the analysis, and it allows the independent contributions of individual factors to be estimated. This approach is especially useful when applied to factors such as age and hearing that are extremely difficult to isolate.

Because the observed raw correlations were significant in all four LC conditions—indicating possible roles of age, hearing, or both in performance on those materials—a linear regression model was fit to the observed LC sentence IP performance. The model included PTA and age as independent variables and a subject-specific random intercept to account for variability among study participants in their ability to perform the TGW task. The model also included a target-word–specific random intercept to adjust for variability in the overall difficulty with which certain targets could be identified. Table 1 shows predicted changes in IP per year of age or dB of hearing loss for LC sentences with and without background noise. PTA effects are statistically significant (p < .05), whereas age effects approached but did not reach significance (p < .1).

Table 1.

Model results for linear regression analysis of mean isolated point (IP) on low-context sentences.

	Quiet				Noise
	Slope	p value	df	SE	Slope	p value	df	SE
Age	0.37	.071	1055	.21	0.40	.088	748	.23
PTA	0.76	.041	1055	.37	0.92	.038	748	.44

Note. Slope is in milliseconds/year or milliseconds/dB and indicates the change in IP per unit increase in age or pure-tone average (PTA). df = degree of freedom; SE = standard error. Interactions are not significant.

Failure rate. For LC sentences in noise, listeners failed to correctly report the target word for a substantial portion of observations (28.36%) at any gate duration compared to LC sentences in quiet (2.21%) or HC sentences in quiet (0.08%) or noise (0.57%). As the failure rate may have influenced the correlations, a logistic regression model was fit to the observed mean failure rates in the LC conditions. As with the linear regression model described above, PTA and age were independent variables, and subject-specific as well as target-word–specific random intercepts were included. The same two highly influential outlier subjects were removed from the analysis.

Table 2 shows the odds ratio (OR), or approximate relative risk of test failure (i.e., failing to correctly identify the target word) per unit increase in the value of independent variable (PTA or age). PTA predicted increased failure rates in quiet (OR = 1.15; p < .001) and noise (OR = 1.06; p < .001). These ORs correspond to an increase in failure rate of approximately 15% and 6%, respectively, per dB increase in subject PTA. The impact of age was not significant.

Table 2.

Model 2 results for logistic regression analysis of IP failure rate.

	Quiet		Noise
	OR	p value	OR	p value
Age	0.97	.247	0.99	.370
PTA	1.15	<.001	1.06	<.001

Note. OR = odds ratio, or approximate relative risk of test failure per unit increase in the value of independent variable.

Confidence. Table 3 presents the correlations of age and PTA with the mean confidence ratings of each participant for all listening conditions. Past work by Craig (1992) has suggested that listeners' confidence about their TGW performance accuracy is influenced by age, with older listeners reporting lower confidence at the gate corresponding to IP compared with younger adult listeners. This could provide an indication that older listeners tend to have less confidence in their speech understanding ability. However, correlations between hearing loss and confidence ratings were not examined in that study. In the present study, confidence was generally unrelated to listener age in all four listening conditions, suggesting that the older listeners were not less or more confident than younger listeners. In contrast, we found significant correlations between confidence at IP and PTA in three of the four listening conditions, but paradoxically, the correlations were all positive indicating that listeners with relatively more hearing loss were more confident in their responses. A likely explanation for this finding is that confidence covaries with IP such that in general when more of the target word is gated-in, listener confidence is higher.

Table 3.

Correlations (r) of age and PTA with mean confidence ratings.

	HC		LC
	Quiet	Noise	Quiet	Noise
Age	.10	.10	.11	.00
PTA	.23*	.23*	.26*	.19

Note. HC = high-context sentences; LC = low-context sentences.

^*p ≤ .01

General Discussion

In this study, we sought to determine whether aging would influence sentence-based estimates of word recognition speed when the independent effects of hearing were considered. Following Craig et al. (1993), we adapted the R-SPIN test by gating-in the final target word within the SPIN sentences to determine how TGW-recognition performance is influenced by target-word predictability. We extended previous work by including background noise conditions as a means of increasing the difficulty of the task to improve the chances of detecting deficits among individuals who are older and/or hearing impaired. For comparison with prior work, analyses include simple correlations; however, we also used regression modeling to separately assess effects of hearing and age, and to control for correlations among multiple measures from each subject and the influence of individual target words.

We found that when the listening condition was not challenging (HC in quiet), little relationship was observed between age and TGW performance, whether performance was evaluated as the earliest gate in which the target word was recognized (IP) or as the rate at which targets were never correctly identified by a subject (failure rate). Thus, in cases in which targets were highly predictable based on the sentence context, there was no penalty or advantage of age. Consistent with previous findings (Craig et al., 1993; Moradi, Lidestam, Hällgren, & Rönnberg, 2014; Moradi, Lidestam, Saremi, & Rönnberg, 2014), we found that word predictability tended to assist subjects in the early sensory perception of time-gated HC words but that older adults' increased language experience did not facilitate higher performance in the TGW-recognition paradigm.

A significant relationship with hearing loss was always observed, however, when task difficulty increased, either through the addition of noise or the removal of supporting sentential context, even though the majority of participants were without substantial hearing loss. In these more challenging conditions, the relationship between age and IP is statistically significant only for raw correlations and only borderline significant for LC conditions in quiet and in noise. Effects of age on hearing-adjusted failure rate did not reach significance.

Thus when the separate effects of age and hearing loss are considered, both have an effect on performance, but only the effect of hearing loss is significant. This has significant implications for the interpretation of studies that use an older group with thresholds that fall within clinically normal limits but nevertheless are demonstrably worse than the younger group. Our results indicate that in these cases when memory and attentional effects are controlled, there remains a possibility that small group differences attributed to aging may be further reduced or eliminated when even slight hearing loss is accounted for (Dubno, Ahlstrom, & Horwitz, 2000).

To aid in visualizing the relative effects of age and hearing, graphical representations of model predictions including both age and PTA provided in Table 1 are shown in Figures 8 and 9, respectively. In Figure 8, each line represents the impact of age on performance predicted by the model for two PTAs: 5 dB (bold lines) and 20 dB (thin lines). Although the model predicts that mean IP will increase with age, this increase only approached significance. In Figure 9, each line represents the impact of PTA on performance predicted by the model for two ages: 30 years (bold lines) and 60 years (thin lines) in quiet (solid lines) or noise (dashed lines). The lines in each figure extend across data ranges that were used in the model, which represent the ranges for which the accuracy of model predictions is greatest. These representations illustrate the separate influence of PTA and age on IP: At any given fixed age, IP increases about 1 ms per dB increase in PTA; at any given fixed PTA, IP increases about 10 ms in 30 years. Therefore, although the IP of the older listeners will be longer regardless of PTA, an equal increment in PTA will have the same size effect on measured IP regardless of age.

Figure 8.

Graphical representation of model predictions for linear regression analysis of mean IP. Each line represents the impact of age on performance predicted by the model for two PTAs: PTA = 5 (bold lines) and PTA = 20 (thin lines) in quiet (solid lines) or noise (dashed lines).

Figure 9.

As in Figure 8, but each line represents the impact of PTA on performance predicted by the model for two ages: an age of 30 years (bold lines) and an age of 60 years (thin lines).

A number of confounds may have affected earlier comparisons of TGW performance between younger and older listeners that generally found that older adults required more of a target word to be gated-in in order to correctly recognize the target and its beginning phoneme, and had higher failure rates for correctly identifying targets (Craig, 1992; Craig et al., 1993; Elliott et al., 1987; Marshall, Duke, & Walley, 1996). Perhaps the most obvious potential confound is related to the association between aging and hearing loss, which is often observed even among study participants with clinically normal hearing. To avoid this, the approach taken here was to test listeners who varied in age and hearing, and then develop statistical models of TGW performance based on multiple regression with age and PTA included to adjust for age and hearing threshold differences among participants. This approach leads us to conclude that the differences in time-gated word performance observed here are mediated primarily by degree of hearing loss rather than participant age.

Conclusions

In this study, we reached the following conclusions:

1. The SPIN-TGW task measures the duration of the target word required for its recognition within a sentence under conditions that vary the linguistic (HC vs. LC sentences) and acoustic (quiet vs. noise) difficulty of the task.

2. Metrics derived from this task, IP and failure rate, can be used to estimate the speed and accuracy with which words in connected speech can be understood by an individual listener, both of which appear to be influenced by hearing loss in LC and/or noise conditions.

3. Consistent with a large body of speech perception literature, the presence of background noise increased the size of effects attributable to hearing loss; however, we found that effects of age were only borderline significant in the most difficult conditions (LC and/or in noise). Only context from the sentence appeared to drive IP and failure rate in quiet conditions.

4. Results provide evidence that aging and pure-tone threshold elevation separately contribute to impaired speech understanding among older listeners; however, these data indicate that hearing had a stronger relationship to TGW performance than age.

5. Caution must be exercised when conducting and evaluating studies that attempt to assess the independent effects of age and hearing loss.

Appendix

SPIN-TGW Sentence Lists

Sentence	Target	Context	R-SPIN form
Practice 1 (quiet, no gating)
1. Nancy didn't discuss the	SKIRT	LC	1
2. Hold the baby on your	LAP	HC	1
3. Bob discussed the	SPLASH	LC	1
4. The dog chewed on a	BONE	HC	1
5. Bill might discuss the	PILE	LC	1
6. The witness took a solemn	OATH	HC	1
7. You heard Jane called about the	VAN	LC	1
8. Tom has been discussing the	BEADS	LC	5
9. The nurse gave him first	AID	HC	3
10. Ann works in the bank as a	CLERK	HC	5
Practice 2 (quiet, with gating)
11. Stir your coffee with a	SPOON	HC	1
12. Miss White won't think about the	CRACK	LC	1
13. He would think about the	RAG	LC	1
14. The plow was pulled by an	OX	HC	1
15. We could consider the	FEAST	LC	1
Practice 3 (noise, no gating)
1. The king wore a golden	CROWN	HC	3
2. The pond was full of croaking	FROGS	HC	3
3. That accident gave me a	SCARE	HC	3
4. She's spoken about the	BOMB	LC	3
5. We can't consider the	WHEAT	LC	3
6. Ruth hopes he heard about the	HIPS	LC	1
7. The war was fought with armored	TANKS	HC	1
8. She wants to talk about the	CREW	LC	1
9. They had a problem with the	CLIFF	LC	1
10. They drank a whole bottle of	GIN	HC	1
Practice 4 (noise, with gating)
1. The old train was powered by	STEAM	HC	1
2. The old man talked about the	LUNGS	LC	1
3. I was considering the	CROOK	LC	1
4. Let's decide by tossing a	COIN	HC	1
5. Bill heard we asked about the	HOST	LC	1
List 1A
1. The glass had a chip on the	RIM	HC	3
2. The boat sailed along the	COAST	HC	5
3. The fireman heard her frightened	SCREAM	HC	5
4. A bicycle has two	WHEELS	HC	3
5. Bob wore a watch on his	WRIST	HC	5
6. Cut a piece of meat from the	ROAST	HC	5
7. Her cigarette had a long	ASH	HC	3
8. The doctor prescribed the	DRUG	HC	1
9. He tossed the drowning man a	ROPE	HC	3
10. The fur coat was made of	MINK	HC	5
11. She wants to speak about the	ANT	LC	5
12. Mr. Brown can't discuss the	SLOT	LC	5
13. Bob could consider the	POLE	LC	3
14. Jane hopes Ruth asked about the	STRIPES	LC	3
15. We're discussing the	SHEETS	LC	3
16. Mary can't consider the	TIDE	LC	5
17. Mr. Smith thinks about the	CAP	LC	3
18. Jane didn't think about the	BROOK	LC	5
19. Ruth has a problem with	JOINTS	LC	3
20. Bill heard Tom called about the	COACH	LC	3
List 1B
1. They tracked the lion to his	DEN	HC	1
2. The swimmer dove into the	POOL	HC	5
3. The airplane went into a	DIVE	HC	5
4. The storm broke the sailboat's	MAST	HC	3
5. Eve was made from Adam's	RIB	HC	5
6. The house was robbed by a	THIEF	HC	5
7. The rancher rounded up his	HERD	HC	5
8. The landlord raised the	RENT	HC	5
9. The chicks followed the mother	HEN	HC	5
10. A chimpanzee is an	APE	HC	5
11. Tom had spoken about the	PILL	LC	5
12. Tom is talking about the	FEE	LC	5
13. He hasn't considered the	DART	LC	5
14. We are speaking about the	PRIZE	LC	3
15. Jane has spoken about the	CHEST	LC	3
16. The girl should consider the	FLAME	LC	5
17. He is considering the	THROAT	LC	3
18. Paul spoke about the	PORK	LC	3
19. He hopes Tom asked about the	BAR	LC	5
20. Jane did not speak about the	SLICE	LC	5
List 2A
1. He hit me with a clenched	FIST	HC	3
2. The bandits escaped from	JAIL	HC	5
3. The judge is sitting on the	BENCH	HC	5
4. My TV has a twelve-inch	SCREEN	HC	3
5. The sailor swabbed the	DECK	HC	3
6. The boy gave the football a	KICK	HC	3
7. She faced them with a foolish	GRIN	HC	3
8. It's getting dark, so light the	LAMP	HC	5
9. The heavy rains caused a	FLOOD	HC	5
10. We shipped the furniture by	TRUCK	HC	3
11. Mr. White spoke about the	FIRM	LC	3
12. Harry will consider the	TRAIL	LC	5
13. Paul can't discuss the	WAX	LC	5
14. We could discuss the	DUST	LC	5
15. He is thinking about the	ROAR	LC	3
16. Betty knew about the	NAP	LC	5
17. We've been thinking about the	FAN	LC	5
18. We've been discussing the	CRATES	LC	5
19. They were interested in the	STRAP	LC	3
20. Betty can't consider the	GRIEF	LC	5
List 2B
1. Use this spray to kill the	BUGS	HC	3
2. Let's invite the whole	GANG	HC	5
3. His plans meant taking a big	RISK	HC	1
4. Watermelons have lots of	SEEDS	HC	3
5. The soup was served in a	BOWL	HC	3
6. Please wipe your feet on the	MAT	HC	3
7. The girl swept the floor with a	BROOM	HC	3
8. How much can I buy for a	DIME	HC	3
9. The boy took shelter in a	CAVE	HC	5
10. Mr. Brown carved the roast	BEEF	HC	3
11. David might consider the	FUN	LC	5
12. Tom could have thought about the	SPORT	LC	5
13. He was interested in the	HEDGE	LC	5
14. Mary knows about the	RUG	LC	5
15. Betty has considered the	BARK	LC	3
16. You'd been considering the	GEESE	LC	3
17. Tom will discuss the	SWAN	LC	3
18. Harry had thought about the	LOGS	LC	3
19. The man spoke about the	CLUE	LC	3
20. He could discuss the	BREAD	LC	3

Funding Statement

This work was supported by Veterans Affairs Rehabilitation Research & Development Service Grants C7450R, C7113N, C6116W, C4963W, and the National Center for Rehabilitative Auditory Research.