Participants

A total of 40 older adults were included in this analysis: 20 older adults with normal hearing (ONH; 17F, 3M; mean 67 years, 60–74 years) and 20 older adults with hearing loss (OHI; 13F, 7M; mean 72 years, 60–85 years). Participants completed the Mini Mental Status Exam (MMSE) [33] with all participants scoring ≥ 26 (mean = 29 out of 30). All participants had corrected vision at or better than 20/40 on a Snellen chart, except for one participant in each group with scores of 20/50. An additional 64 younger adults with normal hearing (46 F, 18 M; mean 22 years, 18–38 years) also completed all tasks. Results on the individual speech recognition tasks from these younger listeners are detailed in previous reports [34–36] and are not discussed here given the focus on individual differences among older adults. However, so that variance in scores due to age and hearing loss are captured in the current report, factor analyses reported here were conducted with the full participant sample (N = 104) to increase power and better capture individual variance associated with age. The experimental protocol was approved by Institutional Review Boards of the University of South Carolina and the Medical University of South Carolina, with subsequent data use procedures approved by the University of Illinois. Written informed consent was obtained for all participants. Participants were reimbursed for their participation. Recruitment of the older participants reported here occurred between December 4^th, 2017 and December 14^th, 2018.

Auditory measures

Pure-tone hearing thresholds. All participants completed pure-tone threshold testing at standard audiometric frequencies between 250–8000 Hz. Audiometric thresholds for the two listener groups are displayed in Fig 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Mean audiograms for the two listener groups.

Error bars display ± 1 standard deviation.

https://doi.org/10.1371/journal.pone.0331487.g001

Psychoacoustic measures of modulation detection and modulation interference using speech-modulated noise. Speech temporal-envelope modulations, adaptively compressed using an exponential function, were imposed on target noise bands (i.e., 500 Hz and 3150 Hz) of a speech-spectrum 500-ms noise. Psychoacoustic tasks presented stimuli in a three-interval two-alternative forced choice task. Modulation detection required participants to indicate the interval containing the target band modulation compared to the interval with two unmodulated noise bands. Modulation interference required participants to indicate the interval containing the target band modulation with concurrent uncompressed speech modulation in the non-target band, which was present in the target and comparison intervals. Additional details regarding the methods and data reduction results for this task can be found in S1 Appendix.

Speech glimpsing refers to the ability to make use of partial acoustic information to understand speech. Glimpses are time-frequency regions in which the speech occurs at relatively favorable signal-to-noise ratios (SNRs). Participants’ open-set sentence recognition abilities were measured in two maskers (steady-state or unmodulated speech-shaped noise, SSN) and speech-modulated noise (SMN). Speech glimpsing was operationally defined as a fluctuating masker benefit (FMB), calculated as the difference in keyword accuracy for speech presented in SMN and SSN (that is, FMB = SMN – SSN). SSN was created to match the long-term average speech spectrum for concatenated IEEE sentences [37] prior to shaping. The SMN was created by amplitude modulating the SSN by the wideband temporal envelope of the concatenated sentences, extracted using half-wave rectification. Amplitude modulations were low-pass filtered using a 4^th order Butterworth filter with a cutoff modulation frequency of 16 Hz. A random selection of noise with a 100-ms onset/offset padding was added to the speech on every trial. To reduce individual variability, SSN was always tested prior to SMN because speech recognition in SMN is an easier task due to speech glimpsing (e.g., [38,39]).

Cognitive process-based measures

In order to index a broad range of cognitive processes, an in-person, computerized version of the Brief Test of Adult Cognition by Telephone (BTACT) [40] was created using speech recorded by a female speaker. This test was designed to be sensitive to cognitive changes in normal aging. Seven subtests were included: Short Recall, Delayed Recall, Category Fluency, Auditory Backward Digit Span, Backward Counting, and Number Series Completion.

Short and Delayed Recall subtests assessed immediate and delayed episodic memory. These subtests auditorily presented a list of 15 words, spoken at a rate of one word per second. Participants recalled as many words from the list as possible within 1 minute. Short Recall involved repeating the word list immediately after presentation. The Delayed Recall subtest prompted the participants to recall the same word list from the Short Recall task, but the prompt was provided after completing all other BTACT subtests.

Category Fluency is a test of verbal fluency and executive function that requires participants to name as many animals as possible within one minute. Scores indicate the number of unique responses within the allowed time limit.

Auditory Backward Digit Span assessed working memory, requiring listeners to recall a list of digits in backward order. Digits were presented at the rate of one digit per second in sequences ranging from two to eight. A set of two trials were presented at each sequence length. The test was discontinued when the participant missed both trials in a set and the score was the number of digits in the longest sequence correctly reproduced.

Number Series Completion assessed inductive reasoning. Participants controlled the rate of presentation for a series of five numbers. Following the presentation of the fifth number, participants were prompted to indicate a sixth number that would best complete the sequence. Scores indicated the number of sequences completed correctly out of five.

Backward Counting (30-SACT subtest) was a measure of processing speed that involved counting backward from 100 as quickly as possible. The score was the number of correctly repeated numbers within 30 seconds.

In addition to these auditory measures from the BTACT, three additional cognitive measures were included in the visual modality.

Visual Backward Digit Span replicated the Auditory Backward Digit Span procedures using a visual presentation of digits one at a time on the screen, each digit visible for 1 second.

Connections Test [41] is a version of the Trail Making Test in which participants draw lines to connect numbers and/or letters in sequence. Participants first completed drawing lines between ascending numbers and alphabetically sequenced letters in two different trials. They then completed two trials where they were required to alternate between ascending sequences of numbers and letters, beginning with either a number or a letter, respectively. The number of circles properly sequenced within 20 seconds was counted. The final score was the average of the two alternating sequence trials.

Stroop Color-Word Mismatch [42] assessed inhibitory control. Participants completed three training trials: (1) reading color words printed in black ink; (2) naming the color of ink for a “xxxx” text sequence, and (3) reading color words printed in colored ink that matched the target word. Following these three trials, participants were presented with a page with color words printed in color ink that mismatched the written words. Participants were required to name the color of the ink while ignoring the written words. The score was the number of correctly identified colors within 45 seconds.

Cognitive-linguistic processing

In addition to the process-based measures of cognitive function, two additional measures were associated with product-based measures of cognitive-linguistic processing: vocabulary knowledge and linguistic closure.

Vocabulary Knowledge was assessed using a computerized version of the Peabody Picture Vocabulary Test Fourth Edition (PPVT-IV) [43]. Participants heard words recorded by a female speaker and selected a corresponding picture using a touch screen interface. Standard test procedures were used, and raw scores were saved for analysis.

Linguistic Closure was assessed using text recognition threshold testing (e.g., [44]. In the current implementation, the text of IEEE sentences was presented with periodic portions of text deleted. Four proportions of text preservation were tested at 60, 70, 80, and 100%. Logistic functions were fit to the data to derive the text proportion threshold for 50% correct keyword recognition (SRT50), which was the outcome variable of interest for the current investigation. Results from this test were previously reported [45]. All participants scored ≥ 94% on the 100% condition, indicating sufficient visual acuity for the visual tasks.

Speech recognition outcome variables

Speech recognition scores from three previously reported studies were used as the outcome variables of interest. To examine reproducibility, these studies were examined independently to determine if the same pattern of predictors would be observed across studies. All three studies examined the recognition of speech filtered in the temporal modulation domain to examine the contribution of different temporal speech modulation rates to the recognition of degraded speech in noise. As a final measure of generalizability, more standard measures of speech recognition without filtering in the temporal modulation domain were examined. While this analysis provides an indication of whether the same predictors consistently emerge across studies, it is important to note that these studies are not completely independent observations as the same listeners participated in all studies. Therefore, the degree of replication across datasets should be considered as a measure of the reliability of individual differences observed in our sample population.

Experimental datasets for the recognition of temporal-envelope filtered speech

Dataset #1: Spectrally reduced speech (D1_SR). Fifteen acoustic conditions of spectrally reduced speech were analyzed from [35]. TIMIT sentences [46] were bandpass filtered into 18 one-third octave bands. The Hilbert envelopes were extracted from each band. Envelopes were bandpass filtered into two modulation bands: 0–8 Hz and 8–16 Hz. Filtered envelopes were combined with the original spectral components and summed across bands to re-synthesize the original speech sample with reduced temporal modulation cues. Contribution of the modulation depth was assessed by scaling the consonant or vowel levels by a factor of 0.5, 1.0, or 2.0. This resulted in amplitude compression/expansion of the consonant–vowel intensity ratio using the TIMIT phonetic markings. Two modulation bands, two manipulated segments (consonants/vowels), and three segment level settings (0.5, 1.0, 2.0), plus three control conditions of the full sentence limited with temporal modulations filtered at 0–8, 8–16, or 0–16 Hz were tested. This resulted in 15 acoustic conditions tested in the presence of a 2 dB SNR signal-correlated noise, which primarily reduced the spectral content of the speech.

Dataset #2: Noise-masked speech (D2_NM). Sixteen acoustic conditions of temporal-envelope filtered speech (IEEE sentences) and noise were analyzed from Experiment 1 of [36]. As in D1_SR, sentences were passed through 18 one-third octave bands. Hilbert envelopes were then extracted, temporally filtered into a specified speech modulation band, recombined with the original spectral components, and summed across bands. This processed speech was then presented in a steady-state noise (SSN) that matched the long-term average spectrum of the target speech, or a speech-modulated noise (SMN) that modulated the SSN by the temporally filtered Hilbert envelope of a different sentence spoken by the target talker. The temporal envelope for SMN modulation was exponentially modified, E(t)^K, (K = 0.5, 1.0, or 2.0; with 1.0 equal to the original speech modulation) to compress or expand the modulation depth (as in [34,47,48]. Four baseline conditions tested included 0–16 Hz temporally filtered speech in unmodulated SSN and in 0–16 Hz SMN at K = 0.5, 1.0, and 2.0. The remaining 12 conditions tested two speech modulation bands (0–8 Hz, 8–16 Hz), in SMN with two noise modulation bands (0–8 Hz, 8–16 Hz), at three noise modulation depths (K = 0.5, 1.0, and 2.0). This resulted in a total of 16 conditions of temporally reduced speech in modulated noise.

Dataset #3: Spectrally reduced and noise masked speech (D3_SR+NM). A final set of 16 acoustic conditions was selected from Experiment 2 of [36]. This experiment was identical to the speech and noise conditions just described in D2_NM (i.e., speech modulation (0–8 Hz, 8–16 Hz) × SMN modulation (0–8 Hz, 8–16 Hz) × noise modulation depths (K = 0.5, 1.0, and 2.0) + 0–16 Hz speech in SSN, or 0–16 Hz SMN at K = 0.5, 1.0, 2.0), with one exception. During speech processing, instead of recombining filtered Hilbert envelopes with the original spectral components of the speech, it was combined with the spectral components of the SSN. This resulted in an 18-channel noise-vocoded speech signal that was used as the target.

Generalized measures of speech recognition

Three additional speech recognition tasks were also measured to determine the generalizability across measures of speech without temporal-envelope modulation filtering.

Interrupted speech. Speech recognition thresholds for interrupted IEEE sentences were selected from [45]. Speech was interrupted at 2 Hz, which represents the rate most affected by advancing age [49,50]. Speech was tested at fixed proportions of speech preservation (i.e., duty cycles) and fit with a logistic function to determine the duty cycle threshold for 50% and 80% correct recognition.

Competing talker. Following procedures from [51], the Coordinate Response Measure (CRM) [52] sentences were presented in the presence of a single competing talker at 0 dB SNR under selective and divided attention conditions. CRM sentences are in the form: “Ready {call sign} go to {color} {number} now.” Three male and three female speakers were selected. One of the six talkers was randomly selected on each trial as the target talker. The masker talker was selected from the three talkers of the different sex. The call sign was used to cue the target talker and was presented as text onscreen prior to the start of the trial for selective attention or after the presentation of the trial for divided attention. Participants responded using a touchscreen interface to indicate the number and color of the sentence. Selective and divided attention conditions were tested in separate blocks of 32 trials each.

Speech in babble. Fifty sentences from the Revised Speech Perception in Noise (R-SPIN) [53,54] test were tested in 12-talker babble at 0 dB SNR. Participants responded by repeating the final word in the sentence. Half of the sentences had low contextual predictability and half of the sentences had high predictability for the final word. Two scores were recorded, corresponding to the two sentence types.

General procedures

All participants completed the battery of tests described above (summarized in Table 1) over at least five test sessions. Participants completed all testing in a sound-attenuating booth and listened to stimuli at a sampling rate of 48,828 Hz via one of a pair of Sennheiser HDA 200 headphones following a TDT System III digital-to-analog processor (RP2/RX6) and headphone buffer (HB7/HB5). Presentation was monaural to the right ear, unless target sensation levels (SLs) were better obtained using the left ear (3 ONH, 14 OHI). To ensure audibility of the speech materials (i.e., > 15 dB SL) through at least 4000 Hz, including all speech presented during cognitive testing (e.g., task instructions, number or word stimuli, etc.), all listeners received frequency-specific gain based on individual hearing thresholds (i.e., spectral shaping). To limit the contribution of reduced audibility in the higher frequencies, all stimuli were subsequently passed through a low-pass, linear phase, finite-impulse-response, 128th-order filter with a cutoff of 5,623 Hz. All auditory stimuli were calibrated to be presented at 70 dB SPL, with a mean presentation level of 82 dB SPL for OHI listeners following spectral shaping. Thus, all test materials – auditory and cognitive – were presented at SLs that ensured audibility for each participant.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Summary table of experimental measures, including domains, tasks, and variables.

https://doi.org/10.1371/journal.pone.0331487.t001

All measures of speech recognition involved participants repeating sentences, unless otherwise stated (e.g., CRM closed-set responses). These open-set responses were live-scored and recorded. Participants were encouraged to guess and no feedback was provided. A response was scored as correct if the participant repeated each keyword exactly (e.g., without missing or extra phonemes). Finally, while some measures included stimuli from the same speech corpus, no sentence was ever repeated across the complete set of predictor and outcome measures.

Dimension reduction of predictor variables

Significant redundancy was expected among the various sets of variables (see Table 1 for a summary of auditory, cognitive process, and cognitive-linguistic variables). To examine the contribution of the underlying constructs corresponding to these auditory and cognitive variables, it was necessary to first reduce the dimensionality of the data.

Cognitive process-based measures.

A similar dimension reduction step was employed on the nine cognitive process-based measures. Raw scores were first transformed to z-scores and then subjected to PCA to extract factors with eigenvalues greater than 1. A three-factor solution emerged that explained 77% of the variance. All communalities were greater than 0.58 (7 out of 9 were > 0.78, KMO = .79). The pattern matrix following varimax rotation with kaiser normalization is provided in Table 2. Principal components were interpreted as indexing episodic memory, working memory, and interference.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Component weights for each cognitive process-based measure for the rotated PCA factor solution.

https://doi.org/10.1371/journal.pone.0331487.t002

Preparatory analyses.

The first step in this analysis was to derive the psychometric functions for each of the three speech recognition tasks. The psychometric functions were previously calculated for the two noise masked datasets (D2_NM and D3_SR+NM) in [36]. Using this method, psychometric functions were also calculated for D1_SR. As all datasets used different and multiple forms of distortion, it was important to base psychometric functions on a single objective measure of speech distortion. For this reason, all stimuli were analyzed using the Extended Short-Time Objective Intelligibility metric (eSTOI) [55]. The eSTOI compares the spectro-temporal modulation envelopes of the clean and degraded speech signals over short-time segments to produce a similarity measure, with values less than 1.0 indicating the degree of acoustic distortion. The 15–16 experimental conditions were defined according to eSTOI units and logistic functions were fit to the individual data (see Fig 2), which explained 82% − 93% of the total variance on each dataset for the two groups. Speech recognition thresholds (SRT) were calculated at 20%, 50%, and 80% correct (i.e., SRT₂₀, SRT₅₀, SRT₈₀, respectively). Overall, Fig 2 demonstrates that OHI listeners required greater speech preservation (i.e., higher eSTOI values) than ONH listeners to obtain comparable speech recognition, as indicated by the rightward shift of the psychometric functions in panel (b) for the three datasets. Full analysis details can be obtained from the prior dataset reports [35,36].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Speech recognition psychometric functions.

Lines indicate each of the three datasets for (a) ONH and (b) OHI groups.

https://doi.org/10.1371/journal.pone.0331487.g002

Next, an initial exploration of the correlational matrix for each group was examined for the seven potential predictor variables (Table 3). This included the principal components derived during factor analysis along with an additional auditory measure and two cognitive-linguistic measures for a total of eight predictors (i.e., Audiogram PCA, psychoacoustic speech modulation PCA, three cognitive PCAs, speech glimpsing, vocabulary knowledge, and linguistic closure). Only four measures consistently demonstrated significant correlations with any of the outcome variables for either group. These involved two auditory measures: Audiogram PCA and glimpsing; and two cognitive measures: working memory PCA and vocabulary knowledge. Each of these measures had significant correlations on 13–16 comparisons, with average absolute r-values ranging from 0.43 to 0.82 (p < .05). These four predictor variables were selected for the full analysis. Two other measures, linguistic closure and interference, were significant for one or two comparisons, respectively, with absolute r-values from 0.45 to 0.51 (p < .05). These variables were excluded due to their relatively low associations across the range of measures, particularly considering the multiple comparisons.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Correlation matrix of predictor variables with the three speech recognition outcome datasets.

https://doi.org/10.1371/journal.pone.0331487.t003

The distributions for these four selected predictor variables are shown in Fig 3 for the two listener groups. Independent samples t-tests confirmed that OHI, compared to ONH, had poorer thresholds (t(39) = −8.4, p < .001, d = −2.6) and reduced speech glimpsing (t(39) = 3.3, p = .002, d = 1.0). Comparisons for the two cognitive measures (right two panels) indicated no significant group differences for either working memory (t(39) = 1.2, p = .23) or vocabulary knowledge (t(39) = −1.4, p = .16). This indicates that while the groups were defined by differences on the auditory predictor variables, they were matched on the cognitive predictor variables.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Violin plots for the four predictor variables.

Center box plots display interquartile range and median (dot). * p < .01.

https://doi.org/10.1371/journal.pone.0331487.g003

Explanatory analysis.

Next, a dominance analysis [56,57] was conducted to determine the relative contribution of these four predictor variables to speech recognition. This follows a similar procedure previously conducted by Humes et al. [7] to model global speech recognition in older adults. Dominance analysis measures the importance of each predictor variable as defined by the amount of variance accounted for that variable alone and in all possible combinations of the other predictor variables from the linear regression model. This dominance analysis was conducted for each of the nine outcome variables (i.e., three SRTs for three datasets). Results of the full dominance analysis (i.e., all comparisons) are provided in S1 Supporting Information.

Fig 4 plots the general dominance (i.e., total variance in speech recognition) obtained for each dataset (i.e., bars) and listener group (i.e., columns) for each of the four predictors from the dominance analysis (i.e., rows). Each row in the figure is associated with one of the four predictors, with the left panels displaying results for ONH and the right panels for OHI. Bars within each plot display the general dominance for that predictor variable for each dataset at each SRT. The mean dominance, averaged across the datasets, is plotted as the solid line.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Results of the dominance analysis.

Plotted is the proportion of total variance explained for the three datasets at each of the three SRTs for ONH and OHI groups. Rows display results for each of the four predictor variables and columns display results for the two groups. Bold lines connect the average dominance of three datasets across SRTs.

https://doi.org/10.1371/journal.pone.0331487.g004

Some general variance is observed across datasets and at the different SRTs. However, the most predominant differences observed are between groups and across the different predictor variables (i.e., individual panels in Fig 5). Notable are the group differences for glimpsing, which provided minimal contribution for ONH compared to OHI speech recognition. An opposite pattern was observed for vocabulary knowledge, with greater contributions for ONH across datasets. Azen and Budescu [56] describe a hierarchy of dominance from complete dominance (dominant for all comparisons) to conditional dominance (dominant for all model sizes, i.e., number of predictors entered) to general dominance (dominant based on the overall average, as plotted in Fig 4). Most notable is that for OHI listeners, complete dominance (i.e., dominance for all comparisons) was observed for glimpsing in D1_SR involving spectral distortion across all SRTs, which switched to complete dominance of the Audiogram PCA in D2_NM and D3_SR+NM involving noise masking (except for one comparison at SRT₅₀ in D2_NM in which the Audiogram PCA demonstrated conditional dependence). No other comparisons demonstrated complete dominance across SRTs.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Overall and generalized dominance.

Percent of total variance explained for ONH and OHI groups on the principal components derived from (a) the three experimental speech datasets, and (b) the generalization speech measures.

https://doi.org/10.1371/journal.pone.0331487.g005