Comparing Speech Recognition for Listeners With Normal and Impaired Hearing: Simulations for Controlling Differences in Speech Levels and Spectral Shape

Abstract

Purpose

This study investigated methods used to simulate factors associated with reduced audibility, increased speech levels, and spectral shaping for aided older adults with hearing loss. Simulations provided to younger normal-hearing adults were used to investigate the effect of sensation level, speech presentation level, and spectral shape in comparison to older adults with hearing loss.

Method

Measures were assessed in quiet, steady-state noise, and speech-modulated noise. Older adults with hearing loss listened to speech that was spectrally shaped according to their hearing thresholds. Younger adults with normal hearing listened to speech that simulated the hearing-impaired group's (a) reduced audibility, (b) increased speech levels, and (c) spectral shaping. Group comparisons were made based on speech recognition performance and masking release. Additionally, younger adults completed measures of listening effort and perceived speech quality to assess if differences across simulations in these outcome measures were similar to those for speech recognition.

Results

Across the various simulations employed, testing in the presence of a threshold matching noise best matched differences in speech recognition and masking release between younger and older adults. This result remained consistent across the other two outcome measures.

Conclusions

A combination of audibility, speech level, and spectral shape factors is required to simulate differences between listeners with normal and impaired hearing in recognition, listening effort, and perceived speech quality. The use of spectrally shaped and amplified speech in the presence of threshold matching noise best provided this simulated control.

Supplemental Material

https://doi.org/10.23641/asha.13224632

Investigating the effects of age-related hearing loss on speech recognition measures often involves comparison among several listener groups to parse out independent factors related to listener age and the underlying hearing impairment. One desired comparison has been to recruit four groups of listeners: older listeners with hearing impairment (OHI), older listeners with normal hearing (ONH), younger listeners with hearing impairment, and younger listeners with normal hearing (YNH). Comparison between younger and older normal-hearing (NH) listeners investigates the effect of age, while comparison between NH and hearing-impaired (HI) listeners investigates the effect of hearing loss. Therefore, this comparison accounts for differences in auditory processing due to the underlying cochlear pathology, and disentangles this from other factors related to aging, such as cognitive decline. However, while such a paradigm has been used successfully (e.g., Dubno et al., 1984; Gordon-Salant & Fitzgibbons, 1993), it is relatively rare because of the difficulty of recruiting young listeners with hearing loss or older adults with normal hearing matched to younger listeners. Further limitations entail differences between age groups in their cochlear pathology, audiometric configuration, auditory processing, and speech-language development.

Due to these challenges, careful consideration of listening conditions is important when comparing speech recognition between listeners with normal and impaired hearing. These may include simulations of the presumed effects of hearing loss, such as reduced speech audibility and changes in the speech spectrum due to unequal threshold elevations across frequency. For example, it has long been recognized that speech sensation level (SL) is a significant contributing factor for recognition of speech (Steinberg, 1929). Thus, NH and HI listeners have been tested at equal SLs (e.g., Stuart & Phillips, 1996). Testing at equal SLs allows for the natural filtering of the ear to equate the audibility of the signal between groups. However, as YNH and OHI have different thresholds across the spectrum, this method does not match SLs within individual frequency bands, which is another potential factor for simulation control. Furthermore, testing at equal SL introduces differences in the speech presentation levels between groups, with substantially higher presentation levels for HI listeners. As performance can decline at high speech presentation levels (Dirks et al., 1982; Studebaker et al., 1999), any group differences noted might be due to the presentation level differences rather than group differences in processing. To provide a means to equate SL and speech presentation level, NH listeners have been tested in the presence of a “threshold matching noise” (TMN, e.g., Dubno & Schaefer, 1992) that equates masked thresholds of NH listeners to quiet thresholds of HI listeners, which also matches their speech audibility. As such, the presence of a TMN results in a reduction of speech audibility. This reduced audibility may result in particular challenges in noisy environments, as the presence of noise has been shown to increase listening effort (Gosselin & Gagné, 2011) and reduced memory for words (e.g., Rabbitt, 1968). Therefore, it is important to consider how effects of SL and TMN presentations, among other manipulations, affect both speech recognition and listening effort, particularly in noisy backgrounds.

Further complicating the comparison between NH and HI listeners are factors related to spectral shaping, which is similar to amplification by well-fit hearing aids. Speech audibility is a primary predictor of speech recognition by HI individuals, as reflected by the Speech Intelligibility Index (American National Standards Institute, 1997). Indeed, providing amplification that differs across frequency as determined by elevated thresholds (i.e., spectral shaping) accounts for a large portion of the deficit in speech recognition experienced by individuals with hearing loss (Humes, 2007). In contrast, a limited number of other studies have shown continued deficits in speech recognition for HI listeners, relative to NH listeners, even with audibility restored using spectral shaping, particularly in more complex noise environments (reviewed by Humes, 2013). These equivocal results suggest that spectral shaping, while accounting for audibility, may introduce differences in speech presentation level, SL, and the spectral shape of speech between NH and HI listeners that contribute to differences in speech recognition.

To address this potential confound, groups of YNH have been tested under acoustically identical listening conditions as OHI listeners who are listening to spectrally shaped speech (e.g., Fogerty et al., 2016). In general, speech recognition by YNH was poorer when listening to spectrally shaped speech according to OHI thresholds than when listening without spectral shaping. In some cases, performance of YNH with spectral shaping matched the poorer performance of OHI listeners. Finally, while providing identical spectral shaping to both YNH and OHI might account for an altered spectrum and increased speech presentation level, this procedure fails to account for the higher SL provided to YNH than OHI.

This discussion by no means represents the only methods of controlling key factors that differ between younger and older individuals with normal and impaired hearing. Factors such as temporal jitter (e.g., Pichora-Fuller et al., 2007) and spectral smearing (Baer and Moore, 1993), among others, have also been investigated through the use of simulation or control groups to determine the role of other processes associated with hearing loss beyond speech audibility. However, no systematic comparison between these various control methods has been reported.

Current Study

Toward this end, the current study examined factors contributing to speech recognition that are related to SL, speech presentation level, and spectral shape for YNH in comparison to a group of OHI listeners. Speech recognition was tested in the background of steady-state noise (SSN) and a speech-modulated noise (SMN) to determine how performance might vary due to the listening background.

Additionally, a listener's experience is multidimensional (Pichora-Fuller et al., 2016), with speech manipulations affecting the perceived quality and listening effort (e.g., response times) to different (and potentially greater) degrees than speech recognition outcomes (e.g., Baer et al., 1993). Changes in perceived quality (and effort) can occur in the absence of changes in recognition (Boike & Souza, 2000). This can have effects on subsequent cognitive operations, such as memory (e.g., Rabbitt, 1968). Therefore, assessment of differences between NH and HI listeners' performance should also consider effects on these other measures, particularly as they may influence listener behavior (e.g., memory, decision making). In addition to measuring speech recognition, the younger listeners in this study also completed subjective ratings of listening effort and speech quality. Response times were also measured as an objective assessment of processing time associated with listening effort. Comparison of these additional measures among the younger participants is informative regarding how the various simulation methods might influence these additional outcomes as compared to influences on speech recognition.

The simulation methods examined here are intended to control for differences in audibility between NH and HI listeners. Controlling for these factors facilitates the isolation of other factors associated with aging and hearing loss (broader auditory filters, temporal resolution, cognition, etc.) that might further account for group differences beyond audibility, such as those listed here (see also Humes, 2007). As such, the simulation methods detailed here represent an important and often used methodological approach. Simple accounts of audibility, speech level, and spectral shape often drive differences between speech recognition for adults with normal and impaired hearing. These simulation methods are designed to control these differences so that additional processing differences associated with the age-related effects of hearing loss can be identified and studied.

Overview of Experiments

This study consisted of two sets of experiments as outlined in Table 1. Experiment (Exp.) 1 used a standard group-design approach to examine speech recognition across four groups of participants that varied in age, hearing status, and spectral shaping. In comparison to the OHI group tested in this first experiment, Exp. 2 was designed to test several different methods to simulate OHI performance in YNH. Three subexperiments in Exp. 2 assessed outcomes in speech recognition, listening effort, and perceived speech quality. Institutional review board approval was obtained from the University of South Carolina and the Medical University of South Carolina. All listeners provided written informed consent prior to participating.

Table 1.

Experimental summary.

Experiment	Measures	Listeners	N
Group comparison
Exp. 1	Recognition	YNH	20
		ONH	20
		OHI	21
		YNHOHI-shaped	21
Simulation comparison
Exp. 2a	Recognition	YNH	47
Exp. 2b	Listening Effort	YNHa	47
Exp. 2c	Speech Quality	YNHb	20

Note. The YNH_OHI-shaped listeners received spectral shaping matched to a random, individually assigned OHI listener. Exp. = Experiment; YNH = younger normal hearing; ONH = older normal hearing; OHI = older hearing impaired.

^a Same participants as Experiment 2a.

^b Ten participants previously completed Experiments 2a and 2b.

Method

Participants

Exp. 1 consisted of four groups of participants. Two groups listened to unshaped speech at 70 dB SPL: 20 YNH (19–28 years, M = 23 years) and 20 ONH (60–74 years, M = 67 years). Another two groups received spectrally shaped speech: 21 OHI listeners (62–84 years, M = 72 years) and 21 YNH who listened to shaped speech at the SPL set according to individually matched OHI listeners (YNH_OHI-shaped; 19–26 years, M = 22 years). Both groups of younger listeners had pure-tone thresholds ≤ 20 dB HL at octave frequencies from 0.25 to 8 kHz (American National Standards Institute, 2004). ONH participants had thresholds ≤ 25 dB HL at 4 kHz and below. Audiometric thresholds for the ONH and OHI listeners in the test ear are displayed in Figure 1. Five out of the 21 OHI adults reported using hearing aids. An overview of the experimental groups is provided in Table 1. Younger listeners were recruited from the University of South Carolina, and older listeners were recruited at the Medical University of South Carolina. All participants in this and subsequent experiments were native speakers of American English.

Figure 1.

Older listeners with normal hearing (ONH; left) and older listeners with hearing impairment (OHI; right) audiometric thresholds (dB HL) as a function of frequency (thin solid lines). The thick solid line indicates the average hearing thresholds for all ONH or OHI listeners. Spectral shaping provided to younger listeners with normal hearing (YNH) in Experiment 1 was matched to randomly assigned individual OHI listeners. For Experiment 2, spectral shaping provided to YNH was defined by the mean audiogram of 14 OHI listeners (OHI₁), displayed as the dashed line in the right panel.

Stimuli and Design

Sentences from the Harvard/IEEE corpus (IEEE, 1969) with five keywords per sentence were used as the experimental stimuli (e.g., The birch canoe slid on the smooth planks). Sentences were presented in quiet, SMN, and SSN at –3 dB signal-to-noise ratio (SNR). ¹ Stimuli were scaled to 70 dB SPL and then spectrally shaped according to the individual thresholds of the YNH, ONH, or OHI to ensure at least 15 dB SL through at least 4 kHz. In practice, this resulted in minimal changes in the spectrum for YNH and ONH. The average speech level following OHI spectral shaping was 82 dB SPL. The YNH_OHI-shaped group listened to spectrally shaped speech that was identical to the shaped speech provided to individual OHI participants to which they were randomly paired. Examples of the unshaped (UN_SL/SPL) and shaped (S_SPL) speech spectra are displayed in Figures 2A and 2B, respectively. In this experiment, the unshaped speech was presented at a conversational level, while the exact shaped spectrum varied on an individual listener basis determined by their hearing thresholds.

Figure 2.

Speech spectra are displayed for the six simulation methods used in Experiment 2. Speech spectra (solid lines) from left to right for (A) unshaped SL/speech level (UN_SL/SPL), (B) shaped speech level without or with threshold-matching noise (S_SPL, TMN_SPL), (C) shaped SL (S_SL), and (D) matched-band SL (MB_SL) conditions. In B, the TMN (thick gray line) was shaped to the OHI₁ thresholds (dotted line) and presented continuously with the shaped S_SPL speech to form the TMN_SPL condition. Gray dashed lines in all panels are the mean thresholds for the group of YNH in Experiment 1, which were used to determine levels for the SL conditions. Experiment 1 presented OHI and YNH_OHI-shaped groups with speech shaped according to the S_SPL condition in B. SL = speech sensation level; SPL = sound pressure level; YNH = younger listeners with normal hearing; OHI = older listeners with hearing impairment.

Stimulus Processing

Speech stimuli were first scaled in matlab to the same root-mean-square (RMS) level and then passed through a low-pass, linear phase, finite-impulse-response, 128th-order filter with a cutoff of 5.623 kHz. The filter was designed such that the level of each 1/3 octave band from 0.1 to 8 kHz was adjusted to achieve the original overall speech level of 70 dB SPL, or to achieve a higher band level needed for each OHI listener (or for YNH, paired listener) based on the targeted band SL.

For listening in noise, SSN was created to match the long-term average speech spectrum for a concatenation of the IEEE sentences 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 Butterworth filter with a cutoff modulation frequency of 16 Hz. SSN and SMN underwent the same signal processing as the speech for matching the spectral shape. A random selection of noise with a 100-ms onset/offset padding was added to the speech on every trial. These processing procedures meant that spectral shaping occurred for the speech and noise independently prior to mixing them for presentation. This was done to ensure full audibility of the speech materials following amplification.

Procedure

Participants listened in a sound-attenuating booth to stimuli at a sampling rate of 48.828 kHz 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 SLs were better obtained using the left ear (three ONH, 14 OHI). Participants completed the quiet condition first in a separate session. To reduce individual variability, SSN was tested prior to SMN as it is already well known that SMN leads to potential advantages in speech recognition due to speech glimpsing (e.g., Bernstein & Grant, 2009; Fogerty et al., 2015).

Open-set responses were live scored. Participants were encouraged to guess. No feedback was provided. A response was scored as correct if the participant repeated each keyword exactly (e.g., without missing or extra phonemes). Percent keyword correct was transformed to rationalized arcsine units to stabilize the error variance (Studebaker, 1985).

Results and Discussion

Figure 3A displays the keyword recognition results for the four experimental listening groups. Overall, scores were highest for both YNH groups, followed by ONH, and then OHI. Results for the reference quiet condition were near ceiling, and therefore, analysis was only conducted over the two noise conditions (see Supplemental Material S1 for speech recognition scores in quiet).

Figure 3.

Experiment (Exp.) 1 speech recognition (A) and masking release (B) scores for the four listener groups. Exp. 2 speech recognition (C) and masking release (D) scores for younger normal hearing (YNH) listeners simulated conditions compared relative to older hearing-impaired (OHI) listener scores. To facilitate this comparison, solid and dashed lines across A and C indicate performance for OHI listeners in speech-shaped noise (SSN) and speech-modulated noise (SMN), respectively. The dotted line across B and D provides the reference OHI masking release score. Error bars indicate the standard error of the mean. RAU = rationalized arcsine unit; ONH = older listeners with normal hearing; UN_SL/SPL = unshaped SL/speech level; S_SPL = shaped speech level; TMN_SPL = threshold-matching noise; S_SL = shaped SL; MB_SL = matched-band SL.

A mixed-model analysis of variance with background noise (SSN and SMN) as repeated factors and group (YNH, YNH_OHI-shaped, ONH, OHI) as a between factor was conducted. Significant main effects were found for background noise, F(1, 78) = 259.8, p < .001, η_p ²= .77, and group, F(3, 78) = 3.1, p = .002, η_p ²= .11, along with a significant interaction, F(3, 78) = 5.4, p = .002, η_p ²= .17. All groups demonstrated a significant difference between the two noise backgrounds, but the interaction and group effect were mainly driven by reduced performance in SMN for the OHI group in comparison to the NH groups. Post hoc results demonstrated that the OHI group performed significantly worse than the other groups in SMN (p < .05), but not SSN. Previous studies found similar YNH_OHI-shaped and OHI performance (e.g., Fogerty et al., 2015). However, YNH_OHI-shaped subjects listening to spectrally shaped speech at HI-equivalent levels did not successfully model OHI performance in the current case. Specifically, this observation is related to performance in SMN where OHI appear to have reduced ability to benefit from improved SNRs that occur when the noise level momentarily dips. This is indicated by the smaller masking release (i.e., SMN–SSN) scores displayed in Figure 3B for OHI versus YNH groups.

Individual performance and pairings between YNH_OHI-shaped and OHI listeners are provided in Supplemental Material S2. While the YNH_OHI-shaped group was provided an acoustic match to the spectral shaping and overall levels provided to OHI listeners, variations in individual performance were not matched. These variations were likely due to differences in other suprathreshold and cognitive factors. However, by controlling for individual differences in audibility, future work may be better able to define the contribution of these additional individual factors.

Finally, high consistency was seen in individual listeners between scores for SSN and SMN. Figure 4 displays the correlation for these two conditions for the four listener groups. The strong association between conditions indicates that a large portion of individual variability (62% of the variance across all four groups) may be related to general speech-in-noise processing abilities.

Figure 4.

Individual speech recognition scores in speech-shaped noise (SSN) versus speech-modulated noise (SMN) for the four listener groups tested in Experiment 1. The correlation collapsed across all groups is displayed by the dotted line. Points above the solid reference line indicate participants who had higher speech recognition scores in SMN. OHI = older listeners with hearing impairment; ONH = older listeners with normal hearing; YNH = younger listeners with normal hearing; RAU = rationalized arcsine unit.

Simulation Methods

Table 1 summarizes Exp. 2, which included three subexperiments that tested YNH on measures of speech recognition (2a), listening effort (2b), and perceived speech quality (2c). Each of these experiments used these outcome measures to compare results with the various simulation methods outlined in Table 2.

Table 2.

Definition of the six experimental conditions.

Condition	Spectral shape	Speech level (dB SPL)	Parameter controlled
			SL	SPL	Shape
UNSPL	Unshaped	82		X
SSPL	OHI1 shape	82		X	X
TMNSPL	OHI1 shape + noise at OHI1 thresholds	82	Xa,b	X	X
UNSL	Unshaped	55	Xa
SSL	OHI1 shape	55	Xa		X
MBSL	OHI1 band-SL shape	58	Xb

Note. X denotes the parameter(s) controlled in the test condition. SL = sensation level; SPL = sound pressure level; OHI₁ = older listeners with hearing impairment from Exp. 1; UN_SL/SPL = unshaped SL/speech level; S_SPL = shaped speech level; TMN_SPL = threshold-matching noise; S_SL = shaped SL; MB_SL = matched-band SL.

^a Overall SL was matched.

^b SL was matched within individual frequency bands.

The same speech materials from Exp. 1 were used in Exp. 2. The mean audiogram from the first 14 OHI listeners (OHI₁) from Exp. 1 (see Figure 1, right panel, dashed line) was used to define the spectral shaping for the six simulation conditions listed in Table 2 and displayed in Figure 2. The first three conditions (outlined in Table 2) presented speech scaled to the overall SPL of the OHI₁ group. Speech was either unshaped (UN_SPL; see Figure 2A), shaped identical to the spectral shaping provided to the OHI and YNH participants in Exp. 1 (S_SPL; see Figure 2B), or shaped identically to S_SPL in the presence of TMN (TMN_SPL; see Figure 2B). The next three conditions presented speech scaled to the overall SL of the OHI₁ group. Speech was either unshaped (UN_SL; see Figure 2A), shaped identical to S_SPL, and scaled to the overall SL of OHI₁ (S_SL; see Figure 2C) or shaped so that the speech spectrum matched SLs within individual 1/3 octave bands (matched-band [MB_SL]; see Figure 2D). In the matched-band condition, individual band attenuation created SLs equal to the SLs for OHI₁ in each spectral band, relative to the shaped spectrum. These procedures resulted in average overall speech levels of 82 dB SPL for UN_SPL/S_SPL and TMN_SPL, 55 dB SPL for UN_SL/S_SL, and 58 dB SPL for MB_SL. ²

These six conditions were presented in three backgrounds: quiet, SMN, and SSN. SSN and SMN ³ were analogous to Exp. 1. This resulted in a total of 18 conditions (six spectral conditions × 3 backgrounds). UN_SL, UN_SPL, S_SL, and S_SPL speech recognition results for YNH have been previously reported (see Madorskiy, 2019). The focus of the current report is to present and discuss YNH performance across a broader range of conditions, including MB_SL and TMN_SPL, as compared to ONH and OHI performance.

Experimental setup and testing procedures were identical to Exp. 1 with the following exception. The additional quiet conditions were presented first to increase familiarity with the spectrally shaped conditions. This was followed by the two background noise conditions counterbalanced across participants. Spectral conditions were presented in randomly ordered blocks of 10 sentences. Each background condition was preceded by a set of 10 demonstration trials.

Experiment 2a: Speech Recognition

Method

Exp. 2a consisted of 47 YNH (18–29 years, M = 22 years) who completed open-set speech recognition testing, analogous to Exp. 1, on all six simulation conditions. Responses were live scored and were recorded time-locked with the stimulus file for subsequent analysis in Exp. 2b. No feedback was provided.

Data Analysis

It was of interest in Exp. 2a to determine the simulation conditions in which performance of the OHI from Exp. 1 differed significantly from the YNH simulation conditions tested in Exp. 2. Significant differences between the groups suggest that factors other than the parameter experimentally controlled in the simulation condition may contribute to group differences. Toward this end, a series of mixed-model analyses of variance, one for each of the six simulation methods, was conducted with background noise (SSN and SMN) as the repeated measure and group (YNH simulation and OHI) as between groups. Again, as performance was near ceiling in quiet, this is provided only as a reference condition to indicate maximum performance (see Supplemental Material S1).

In addition, a second (and primary) question of Exp. 2a was whether the simulation methods provided to YNH resulted in equivalent performance to the OHI group. Toward this end, a series of independent-samples equivalence tests was conducted between YNH scores under the six simulation conditions and OHI scores. These comparisons were conducted for SSN, SMN, and masking release (i.e., SMN–SSN). Following the procedures of Lakens (2017), a sensitivity analysis in G*Power (Faul et al., 2007) was first conducted to determine the equivalence bounds based on the examined sample size, 80% power, and an alpha of .05. These bounds were then used to conduct the independent-samples equivalence test.

Results and Discussion

Figure 3C displays speech recognition performance for YNH subjects for the six simulation conditions tested in Exp. 2a. As expected, scores were higher in SMN (gray bars) compared to SSN (black bars) across all conditions. To aid in comparing performance to the OHI listeners in Exp. 1., horizontal reference lines display speech recognition scores for OHI in SSN (solid) and SMN (dotted).

Tables 3 and 4 provide the results for examining statistical difference and equivalence, respectively, for speech recognition by YNH and OHI. Regarding equivalence testing, a value of p < .05 indicates that the two groups performed statistically equivalent to each other (see bold values in Table 4).

Table 3.

Results for 2 (noise: steady-state noise, speech-modulated noise) × 2 (group: younger listeners with normal hearing [YNH], older listeners with hearing impairment [OHI]) mixed-model analyses of variance for comparing speech recognition scores for OHI subjects to scores for YNH in Experiment 2a in each of the simulation conditions.

Condition	Effect	F(1, 66)	p	ηp 2
UNSPL	Noise	315.5	< .001	.827
	Group	32.9	< .001	.333
	Noise × Group	97.7	< .001	.597
SSPL	Noise	237.1	< .001	.782
	Group	9.7	.003	.128
	Noise × Group	49.8	< .001	.430
TMNSPL	Noise	84.5	< .001	.561
	Group	1.1	.30	.017
	Noise × Group	0.05	.82	.001
UNSL	Noise	173.6	< .001	.725
	Group	0.1	.76	.001
	Noise × Group	36.9	< .001	.359
SSL	Noise	144.4	< .001	.686
	Group	13.1	< .001	.166
	Noise × Group	9.5	.003	.126
MBSL	Noise	285.0	< .001	.812
	Group	31.1	< .001	.321
	Noise × Group	64.0	< .001	.492

Note. UN_SL/SPL = unshaped SL/speech level; S_SPL = shaped speech level; TMN_SPL = threshold-matching noise; S_SL = shaped SL; MB_SL = matched-band SL.

Table 4.

Results from the independent-samples two, one-sided tests of equivalence comparing speech recognition scores for OHI subjects to speech recognition scores for YNH subjects in each of the simulation conditions.

Condition	SSN		SMN		MR
	t(66)	p	t(66)	p	t(66)	p
UNSPL	–5.932	1.000	1.326	.095	7.049	1.000
SSPL	–2.562	.006	2.926	.998	4.211	1.000
TMNSPL	–1.873	.033	–1.792	.039	2.597	.006
UNSL	–0.133	.553	–0.501	.309	3.258	.999
SSL	–0.722	.236	1.960	.973	0.237	.593
MBSL	–5.022	1.000	0.404	.344	5.185	1.000

Note. Significantly equivalent comparisons, p < .05, are in bold. SSN = steady-state noise; SMN = speech-modulated noise; MR = masking release; UN_SL/SPL = unshaped SL/speech level; S_SPL = shaped speech level; TMN_SPL = threshold-matching noise; S_SL = shaped SL; MB_SL = matched-band SL.

SSN

Comparing the solid OHI horizontal reference line in Figure 3C to the black bars for YNH simulation in SSN, only two conditions (S_SPL and TMN_SPL) result in statistically equivalent performance (see Table 4). Both conditions provide the same speech level and spectral shape as provided to OHI. Notably, the other condition presented at the same speech level (UN_SPL) had remarkably poorer performance in SSN. Furthermore, S_SL, which provided the same spectral shape, resulted in significantly better performance in SSN. Therefore, it appears that both speech level and spectral shape are important for accounting for OHI performance in SSN.

SMN

Performance in SMN can be assessed by comparing the dotted horizontal OHI reference line in Figure 3C to the gray bars for YNH simulation. Unlike in SSN, YNH in the S_SPL condition performed significantly better than OHI in SMN. This is consistent with the results from Exp. 1 for comparisons with the YNH_OHI-shaped group and with reports of reduced glimpsing and modulation masking release in OHI compared to YNH (e.g., George et al., 2006). Listening in the dips may be further reduced for OHI listeners due to other factors, such as reduced temporal resolution (e.g., Dubno et al., 2002; Festen, 1993; Glasberg et al., 1987). In the current study, these conditions and listener groups were all tested at the same SNR and audibility was ensured for the OHI group. Therefore, explanations of reduced glimpsing (or benefit from masker modulations) due to effective differences in glimpses across SNRs (Bernstein & Grant, 2009) cannot explain the group differences observed here. Instead, group differences in SL during the dips in the SMN masker may be important.

Examining Table 2, there were four simulation conditions that could have controlled for SL during the masker dips: the three SL conditions (UN_SL, S_SL, and MB_SL) and TMN_SPL. Of these conditions, only TMN_SPL resulted in statistically equivalent performance to OHI listeners in SMN (see Table 4). This indicates that, in addition to accounting for reduced SL during the masker dips, the other factors of speech level and spectral shape observed as important in SSN still apply in SMN.

Masking Release (SMN–SSN)

As discussed and observed in Figure 3C, TMN_SPL provided a close match to overall OHI performance in both noise backgrounds. Indeed, as indicated in Table 4, the TMN_SPL condition resulted in statistically equivalent speech recognition to OHI. However, did TMN_SPL sufficiently simulate OHIs' reduced masking release for modulated compared to unmodulated noise backgrounds? Figure 3D provides the masking release scores for each simulation condition in reference to the group masking release scores obtained in Exp. 1 (see Figure 3B). As Table 4 indicates, only TMN_SPL resulted in statistically equivalent masking release compared to OHI listeners (11.1 vs. 11.7 rationalized arcsine unit difference for TMN_SPL and OHI, respectively; 95% CI difference [−5.5, 4.4]). In all other conditions, YNH had significantly greater masking release (see significant interactions in Table 3). Thus, the speech recognition benefit in SMN compared to SSN (i.e., masking release) for the OHI group was closely matched by the YNH performance in the TMN_SPL condition (see also George et al., 2006). It is notable that this condition had the smallest masking release for the YNH group of all spectral conditions. These results provide further support for the assumption that reduced masking release by OHI is related to the reduced SL within masker dips (see also Festen, 1993).

Listening Experience Measures

Experiment 2b: Listening Effort Methods

The measures reported for Exp. 2b were collected within the same experimental session as Exp. 2a with the same participants. At the end of each spectral condition, NASA Task Load Index ratings (Hart & Staveland, 1988) were obtained for mental demand, effort, frustration, and performance. The internal consistency of these rating measures was examined using Cronbach's alpha, which was maximized by removing the performance scale (α = .92). Rating scores for the remaining three scales were combined as a proportion of the total score and square root transformed to better meet assumptions of normality. Results for the Performance subscale can be found in Supplemental Material S4.

Response time, the delay between the end of the stimulus and the beginning of the response, was calculated from dual-channel recordings of the stimulus and participant response from Exp. 2a as an objective measure of listening effort (Baer et al., 1993; Houben et al., 2013). Skipped trials were removed from the analysis.

Experiment 2c: Speech Quality Methods

Twenty YNH (18–26 years, M = 22 years) completed speech quality ratings. Ten of these participants had previously participated in Exp. 2a and 2b. Exp. 2c consisted of identical experimental presentation and testing to Exp. 2a and 2b with one exception. Instead of participants repeating each sentence, participants provided speech quality ratings using a 1–10 scale for each sentence in terms of overall impression, loudness, pleasantness, and intelligibility (method adapted from Davies-Venn et al., 2007). Listeners were specifically directed to ignore the noise and rate the quality of the speech that could be detected within the noise background. Internal consistency of these rating measures was again examined using Cronbach's alpha, which was maximized by removing the loudness scale (α = .88). Rating scores for the remaining three scales were averaged and scaled to 1 for a summary measure of speech quality. Results for the Loudness subscale can be found in Supplemental Material S4.

Results for Listening Effort and Speech Quality

Figure 5 displays the results for YNH ratings of listening effort (see Figure 5A), response time (see Figure 5B), and speech quality on individual sentences (see Figure 5C). In Figures 5A and 5B, listening effort ratings and response times were reversed. This change orients the graph to be consistent with the other scores (i.e., shorter bars indicate more effort and longer response times). Results were generally consistent with the speech recognition measures across the three backgrounds. The quiet condition results are provided in Supplemental Material S5, as a reference. In quiet, MB_SL and TMN_SPL had notably reduced quality, likely due to the distorted spectrum and presence of noise, respectively. In SSN (black bars), ratings were generally consistent with speech recognition, although better scores were observed for MB_SL and UN_SPL than predicted by recognition (although see UN_SPL response times). Measures of effort and quality in SMN best matched patterns observed for recognition across the simulation methods. Measures of effort and quality were most consistent with speech recognition for TMN_SPL, both relative to the other simulation conditions and in demonstrating minimal difference between SSN and SMN.

Figure 5.

(A) Listening effort ratings, (B) response times, and (C) speech quality ratings for young normal-hearing listeners. Effort and response time measures were collected in Experiment 2b and quality in Experiment 2c. To facilitate comparisons, scores are plotted so that up indicates better perception (i.e., less effort, faster response times, or higher quality). Error bars indicate the standard error of the mean. UN_SL/SPL = unshaped SL/speech level; S_SL = shaped SL; MB_SL = matched-band SL; S_SPL = shaped speech level; TMN_SPL = threshold-matching noise.

Summary and Conclusions

This study investigated the ability of several simulation methods to account for speech recognition differences between YNH and OHI. In order to control for poorer speech recognition performance due to reduced speech audibility, OHI listeners require amplification that can adequately present speech at positive SLs across frequency (e.g., Humes, 2007). This method results in spectrally shaped speech that differs in several different ways from unshaped speech presented to NH listeners. Spectral shaping not only alters the shape of the speech spectrum but also results in overall higher speech presentation levels and differences in overall SL. In this set of studies, six methods were examined that control for different aspects of these potential confounds. UN_SL and UN_SPL presented unshaped speech at the overall SL or SPL, respectively, of the OHI. However, these methods only control for overall differences in SLs or SPLs, and do not consider differences in the speech spectrum. S_SL presented the speech spectrum shaped to that appropriate for OHI at a speech level that equated the overall SL between YNH and OHI. However, as speech SLs across frequency are different for YNH and OHI, this method results in very different SLs within individual frequency bands for the two groups. Thus, the MB_SL was designed to match SLs within individual frequency bands for YNH and OHI. However, this method also significantly alters the speech spectrum presented to YNH from that presented to OHI. In contrast, S_SPL presented the same shaped spectrum as in S_SL, but at the same speech level presented to the OHI. This is a precise acoustic control for listening by OHI individuals as it presents an identical signal to the YNH. However, SL in this condition is clearly different. Finally, TMN_SPL is a method that includes a higher overall speech level, equivalently shaped spectrum, and reduced SL (both average and within individual frequency bands) experienced by OHI.

In comparison to the results for OHI obtained in Exp. 1, measures of speech recognition in Exp. 2a demonstrate statistically equivalent performance for YNH in the TMN_SPL and OHI. These results suggest that the three factors of reduced SL, shaped spectrum, and overall higher speech level all need to be controlled (or simulated) in order to provide an appropriate comparison of YNH and OHI speech recognition performance. It was this condition that also replicated the reduced masking release in YNH that was observed for the OHI. This result suggests that reduced SL in the masker dips may underlie reduced performance for OHI. Thus, using a noise to match masked thresholds of YNH to quiet thresholds of OHI (i.e., TMN) serves to not only lower YNH performance overall to a range within that of OHI, but it also replicates the effects of poorer SL within the dips of the masker. This argues that TMN is a potentially important and effective method for modeling reduced masking release by OHI listeners.

While the current study only investigated performance at one SNR, Dubno et al. (2002) have shown that hearing thresholds in the presence of TMN for younger and older adults are matched in the presence of TMN alone (our “quiet” condition), as well as in steady-state and interrupted noise when TMN is also provided. Thus, the effect of TMN is consistent with its intended purpose: It elevates hearing thresholds and thereby equates SLs to those of OHI listeners. The effect of differences in SL becomes more important at more favorable SNRs when the background noise falls below hearing thresholds, both long term and during speech glimpses.

Additional measures of listening effort (Exp. 2b) and quality ratings (Exp. 2c) also conformed to the general patterns of speech recognition results across conditions. Thus, YNH are sensitive to changes in spectral conditions in terms of effort and quality. Further research is needed to determine the extent to which these additional measures of the YNH's experience conform to the effort of speech recognition and perceived quality of speech by OHI.

The methods used here only ensured audibility up to 4 kHz for a reduced speech bandwidth (low-pass filtered at 5.623 kHz). While common hearing aid prescriptions (e.g., NAL-RP) do not fully restore audibility in the high frequencies, it is likely that the loss of this high-frequency information also affected speech recognition and listening effort in addition to the manipulations performed here. However, this method was preferred for the current experiments because it restricted the speech spectrum to within the audible range for the OHI.

Overall, this study suggests that, to effectively control for differences between unshaped YNH listening and shaped OHI listening, factors of increased speech level, reduced SL, and altered spectrum all need to be considered. As tested here, shaped speech at the speech level for OHI presented along with a noise (i.e., TMN) to match elevated thresholds across the speech spectrum provides the best control for these multiple and interrelated factors. The application of this simulation method may be particularly important for controlling SL differences during the dips of fluctuating maskers.

Simulation methods will never replace testing real OHI listeners. Instead, their utility is in explaining some aspect of the listening experience for OHI listeners. The simulation methods investigated here demonstrate that careful experimental control of audibility, SL, and spectral shape can explain differences in speech recognition between YNH and OHI. By minimizing these effects, these simulation methods will facilitate the ability to reveal the effects of suprathreshold auditory factors and nonauditory cognitive factors that may explain further differences among the groups and, potentially, individual listeners.

Funding Statement

This work was supported, in part, by the National Institutes of Health, National Institute on Deafness and Other Communication Disorders (Grant R01 DC015465, awarded to DF, and R01 DC000184, awarded to JRD), and the National Center for Advancing Translational Sciences of the National Institutes of Health (Grant UL1 TR001450, awarded to MUSC). Some of the research were conducted in a facility constructed with support from Research Facilities Improvement Program (Grant C06 RR 014516, awarded to MUSC) from the National Institutes of Health/National Center for Research Resources.