Spectrotemporal Modulation Discrimination in Infants with Normal Hearing

Abstract

Objectives:

Spectral resolution correlates with speech understanding in post-lingually deafened adults with cochlear implants (CIs) and is proposed as a non-linguistic measure of device efficacy in implanted infants. However, spectral resolution develops gradually through adolescence regardless of hearing status. Spectral resolution relies on two different factors that mature at markedly different rates: Resolution of ripple peaks (frequency resolution) matures during infancy whereas sensitivity to across-spectrum intensity modulation (spectral modulation sensitivity) matures by age 12. Investigation of spectral resolution as a clinical measure for implanted infants requires understanding how each factor develops and constrains speech understanding with a CI. This study addresses limitations to the current literature. First, the paucity of relevant data requires replication and generalization across measures of spectral resolution. Second, criticism that previously used measures of spectral resolution may reflect non-spectral cues needs to be addressed. Third, rigorous behavioral measurement of spectral resolution in individual infants is limited by attrition. To address these limitations, we measured discrimination of spectrally modulated, or rippled, sounds at two modulation depths in normal hearing (NH) infants and adults. Non-spectral cues were limited by constructing stimuli with spectral envelopes that change in phase across time. Pilot testing suggested that dynamic spectral envelope stimuli appeared to hold infants’ attention and lengthen habituation time relative to previously used static ripple stimuli. A post-hoc condition was added to ensure that stimulus noise-carrier was not obscuring age differences in spectral resolution. The degree of improvement in discrimination at higher ripple depth represents spectral frequency resolution independent of overall threshold. It was hypothesized that adults would have better thresholds than infants but both groups would show similar effects of modulation depth.

Design:

Participants were 53 6- to 7-month-old infants and 23 adults with NH with no risk factors for hearing loss who passed bilateral otoacoustic emissions screening. Stimuli were created from complexes with 33- or 100-tones per octave, amplitude-modulated across frequency and time with constant 5 Hz envelope phase-drift and spectral ripple density from 1 to 20 ripples per octave (RPO). An observer-based, single-interval procedure measured the highest RPO (1 – 19) a listener could discriminate from a 20 RPO stimulus. Age-group and stimulus pure-tone complex were between-subjects variables whereas modulation depth (10 or 20 dB) was within-subjects. A linear mixed model analysis was used to test for significance of main effects and interactions.

Results:

All adults and 94% of infants provided ripple density thresholds at both modulation depths. The upper range of threshold approached 17 RPO with the 100-tones/octave carrier and 20dB depth condition. As expected, mean threshold was significantly better with the 100-tones/octave compared to the 33-tones/octave complex, better in adults than in infants, and better at 20 dB than 10 dB modulation depth. None of the interactions reached significance, suggesting that the effect of modulation depth on threshold was not different for infants or adults.

Conclusions:

Spectral ripple discrimination can be measured in infants with minimal listener attrition using dynamic ripple stimuli. Results are consistent with previous findings that spectral resolution is immature in infancy due to immature spectral modulation sensitivity rather than frequency resolution.

Introduction

Cochlear implantation (CI) has become a common therapeutic option for children with severe to profound hearing loss, with earlier implantation prior to 2 years of age associated with improved vocabulary, comprehension and expression of language, and speech perception outcomes (Tomblin et al. 2005; Niparko et al. 2010; Wu et al. 2011; Dunn et al. 2014; Lyu et al. 2019). Emerging research suggests implantation within the first year of life may yield better speech perception, receptive language and speech production outcomes than implantation between 13 and 24 months of age in congenitally deaf children (Colletti et al. 2012; Leigh et al. 2013; Bruijnzeel et al. 2016; Dettman et al. 2016; Mitchell et al. 2019). However, even with early recognition of hearing loss and timely implantation, speech perception and language outcomes remain highly variable and unpredictable in children receiving implants before 2 years of age (Niparko et al. 2010; Geers et al. 2011) – highlighting the interplay of additional factors like a listener’s socioeconomic standing, family environment and access to habilitation programs (Kirkham et al. 2009; Holt et al. 2013; Noblitt et al. 2018). The fact that many implanted infants eventually achieve excellent speech perception and rates of spoken language development similar to normal hearing (NH) children suggests that a CI can provide sufficient auditory information to achieve such outcomes. However, some implanted infants do not achieve enough benefit from their device to develop spoken language on par with peers with NH and ultimately rely on alternative and/or adjunctive modalities of communication.

Given the gradual development of speech and language, suboptimal outcomes may go unrecognized for several years following cochlear implantation (McConkey Robbins et al. 2004; Ganek et al. 2012). Outcome measures commonly used to study device efficacy include test batteries addressing speech perception, speech production, and language. These require a level of speech and language understanding and are inaccessible to very young listeners; therefore, optimization of CI programming and language habilitation cannot be identified for individual patients for many years (Geers et al. 2003; Geers 2004; Geers et al. 2011). Other measures to track progress in young, implanted listeners include caregiver elicited measures, observational reports and/or measures of sound detection – which do not directly address how well speech sounds are discriminated. Significant progress has been made to develop objective, quantitative methods to measure discrimination of English speech sounds in infants (Martinez et al. 2008; Uhler et al. 2017). However, these measures have several limitations. First, they only assess a subset of speech contrasts and are limited in use with non-English learning infants. Additionally, they only can provide a result on whether or not the child discriminates a given speech sound, rather than a continuous measure of a listeners’ auditory acuity. Finally, a listener’s spectral and temporal auditory acuity can be conflated by standard clinical tests of speech perception.

In part due to these limitations, there has been significant interest in use of non-linguistic measures of auditory acuity to predict speech discrimination. In adults with CIs, performance on non-linguistic tests of auditory discrimination has been shown to correlate with vowel, consonant, and spondee identification in quiet and in noise (Henry and Turner 2003; Henry et al. 2005; Litvak et al. 2007; Won et al. 2007; Won et al. 2011). Such measures – if found to be sufficiently predictive of clinical outcomes – could serve as proxy measures for device efficacy in infants and toddlers before the development of speech and language.

Spectral resolution, one non-linguistic predictor of speech understanding in adult listeners with CIs, refers to the ability to perceive the spectral distribution of sound energy in complex auditory signals. In speech, for example, distinct patterns of acoustic energy across the spectrum indicate differences in vowel or consonant place-of-articulation. Various tasks have been developed to probe spectral resolution of listeners with NH and CI. Noise- and tone-excited vocoders have been used to demonstrate reliance on spectral and temporal cues in speech discrimination in infants and children (Newman and Chatterjee 2013; Cabrera et al. 2015; Cabrera and Werner 2017). These methods have also been employed to illustrate the difficulty infants and young children have with recognizing spectrally degraded speech compared to older children and adults (Eisenberg et al. 2000; Warner-Czyz et al. 2014). More direct measures of spectral resolution are based on the perception of noises with periodic variation in level across the spectrum, referred to as spectrally rippled sounds. Such stimuli are characterized in terms of the density of spectral ripples, in ripples per octave (RPO), and modulation depth (measured in dB). These tasks typically involve detection of modulation from un-rippled noise or discriminating ripple stimuli based on differences in the spectral envelope. The latter method, spectral-ripple discrimination (SRD), has been used to measure development of spectral resolution in listeners with NH and CI (Horn et al. 2017a; DiNino and Arenberg 2018a; Landsberger et al. 2018).

Regardless of the method used, findings are comparable across studies that use SRD to assess spectral resolution. First, thresholds are consistently poorer in pre-lingually implanted children who use CIs (1 – 3 RPO) when compared to chronologically aged-matched NH listeners (7 – 8 RPO) when tested at 20 dB modulation depth (Horn et al. 2017a; DiNino and Arenberg 2018a; Landsberger et al. 2018). This pattern is noted in post-lingually implanted adults as well with adults with NH outperforming adults with CIs (8 – 9 RPO versus 2 – 4 RPO) at 20 dB modulation depth (Landsberger et al. 2018; Moberly et al. 2018). These findings are expected given implanted listeners’ coarse place code as a result of current spread, the limited number of electrodes across the cochlea and incomplete neural survival (Henry et al. 2005; Won et al. 2007; Anderson et al. 2011; Won et al. 2011; Drennan et al. 2014; Jeon et al. 2015; Winn et al. 2016; Horn et al. 2017a). Mapping strategies designed to enhance cochlear implant spectral resolution have been found to yield improved performance on tasks of SRD (Drennan et al. 2010; Smith et al. 2013). Better SRD thresholds also predict better vowel, consonant and speech recognition by listeners with CIs; this association is most well-studied in post-lingually implanted adults (Henry and Turner 2003; Henry et al. 2005; Won et al. 2007; Holden et al. 2016; Winn et al. 2016; Horn et al. 2017a; Lawler et al. 2017; Zhou 2017). Data on SRD in prelingually-implanted adults and children, on the other hand, are relatively lacking. SRD performance has been shown to correlate with recognition of closed-set spondees in noise in school-aged listeners with CIs (Jung et al. 2012; Horn et al. 2017a) and closed-set vowel recognition in school-aged listeners with CIs (DiNino and Arenberg 2018b), but not with identification of open-set monosyllabic words in quiet (Jung et al. 2012). Horn et al. (2017a) showed that the association of spondee identification with SRD performance is significantly weaker in school-aged children with CIs than in adults (Horn et al. 2017a). This is consistent with findings from Gifford et al. (2018) using a related ripple detection task in school-aged and adult listeners with CIs (Gifford et al. 2018).

That children with CIs would be less reliant on spectral resolution for speech perception than adults is somewhat unexpected, given a stronger reliance on spectral speech cues by pediatric listeners with NH (Eisenberg et al. 2000; Nittrouer 2007). An alternative explanation is evident when considering that performance on SRD or any other test of spectral modulation perception relies not only on resolution of the position of spectral modulation peaks – referred to as “frequency resolution”. Rather, SRD relies on both the listener’s frequency resolution and their sensitivity to differences in sound intensity across peaks and troughs – referred to as “spectral modulation sensitivity”. Both factors determine SRD performance independently and rely on different aspects of auditory processing (Supin et al. 1994, 1999; Saoji et al. 2009; Anderson et al. 2012; Isarangura et al. 2019). Supin and others have used the spectral modulation transfer function, which demonstrates the relationship of modulation depth and modulation density to SRD.

Independence of frequency resolution and spectral modulation sensitivity is evident from the fact that SRD remains immature in listeners with NH until around 9–12 years of age (Wightman et al. 1989; Peter et al. 2014; Kirby et al. 2015). This finding is in contrast with the literature on development of frequency resolution during infancy, as derived from psychophysical tuning curves, showing that 6-month old infants have adult psychophysical tuning curve widths (Spetner and Olsho 1990). Markedly different trajectories of frequency resolution and SRD suggest that in listeners with NH, maturation of SRD around age 9–12 years old is primarily due to maturation of spectral modulation sensitivity.

Immature spectral modulation sensitivity may limit the utility of SRD as a measure of spectral resolution in infants. To address this, Horn et al. (2017a) measured SRD thresholds at five different modulation depths in 7–14-year-old children implanted during infancy and in post-lingually implanted adults. Age-group and individual data were used to fit logarithmic spectral modulation transfer functions. With these functions, frequency resolution could be measured independently of spectral modulation sensitivity. The authors found that x-intercept, but not slope, correlated positively and significantly with increased chronological age in children with CIs whereas the slopes were nearly identical for children and adults. In a second study, Horn et al. (2017b) obtained SRD thresholds in 3- and 7-month old infants and adults with NH at two modulation depths (Horn et al. 2017b). While infants had significantly poorer thresholds than adults, mean SRD performance improved with increased modulation depth to a similar degree in infants and adults. The findings from both studies were consistent with the hypothesis that frequency resolution matures during infancy and that prolonged development of SRD is due to maturation of modulation sensitivity.

Taken together, the studies reviewed above are compelling preliminary evidence that spectral resolution is influenced independently by both frequency resolution and spectral modulation sensitivity. The present study was aimed at strengthening confidence in this interpretation of the data through realization of several goals. The first goal was to replicate the Horn et al. (2017b) infant study and generalize the findings to another SRD task. The second was to address criticism that previously used measures of spectral resolution reflect non-spectral cues – such as within-channel temporal cues. This criticism is particularly important in testing infants, who appear to be more sensitive to these cues than adults. For example, in the Horn et al. (2017b) study, infants with NH showed mature SRD at 20 dB depth when the ripple phase was not randomized but immature SRD when phase was randomized. In contrast, neither adults with NH nor those with CIs demonstrated a difference in SRD performance with phase randomization (Won et al. 2011; Horn et al. 2017b). A stronger reliance on within-channel cues by infants is somewhat unexpected given poor sensitivity to intensity modulation (Horn et al. 2017b) and susceptibility to informational masking (Werner and Bargones 1991; Newman et al. 2013). Nevertheless, these findings suggest that differences in performance on static SRD tasks attributed to age may actually represent differences in sensitivity to within-channel temporal cues. To address this, the present study modified the Spectrally and Temporally Modulated Ripple Test (Aronoff and Landsberger 2013) in which the phase of the spectral modulation envelope moves slowly across the spectrum in order to minimize within-channel, non-spectral, intensity cues.

A third goal was to establish a method to obtain multiple SRD thresholds from an individual infant. Pilot testing suggested that the dynamic spectral envelope stimuli appeared to hold infants’ attention and lengthen habituation time relative to the previously-used static ripple stimuli (Horn et al. 2017b). Finally, a fourth goal was to address a limitation discovered after testing had begun. The SRD stimuli initially used were created using the same 33-tones/octave complex previously used by other groups (Aronoff and Landsberger 2013; Landsberger et al. 2018; Moberly et al. 2018). This complex is not sufficient to represent NH-listeners’ highest-perceivable ripple densities and thus could create an artificial ceiling to performance that obscured age differences in spectral resolution at the higher ripple densities (Resnick et al. 2020). Thus, another condition was created using stimuli with a 100-tones/octave complex.

Like previous work, the threshold was measured as the highest SRD density that could be perceived at two different depths of modulation. A similar effect of modulation depth on SRD in the two age groups would suggest that frequency resolution of infants is mature. An additional condition was introduced post-hoc based on realization that the number of sine-wave components per octave of the noise carrier limited accuracy of SRD thresholds due to aliasing of the modulation peaks (Resnick et al. 2020). Based on these findings, it was hypothesized that if infant frequency resolution is adult-like, infant and adults should show a similar degree of improvement in SRD thresholds when the number of components per octave is higher.

EXPERIMENT

This study was conducted in the Department of Speech and Hearing Sciences at the University of Washington. All procedures followed National Institutes of Health regulations and were approved by the Human Subject Institutional Review Board of the University of Washington.

Materials and Methods

Participants –

Adults and infants with NH were recruited for the study from the Communication Studies Participant Pool, a database of individuals interested in research through the University of Washington. Infant inclusion criteria included full term birth (≥ 38-week gestational age), no medical or developmental concerns per parental report, no history of otitis media within 3 weeks of the test date, no more than 2 previous episodes of acute otitis media, no risk factors for hearing loss (Joint Commission on Infant Hearing, 2007), and newborn hearing screening passed bilaterally. Adults reported no history of hearing loss or unusual noise exposure. All subjects were required to pass 4-frequency bilateral distortion product otoacoustic emissions screening (Otodynamics Ltd., Otocheck Screener) in order to participate.

Stimuli –

Stimuli were generated using The MathWorks, Inc. MATLAB software (R2018b) based on the “spectral-temporally modulated ripple test” described by Aronoff and Landsberger (2013). MATLAB scripts provided by that group were used with alterations to modify ripple density, stimulus length, and carrier density (Aronoff and Landsberger 2013). Phase-randomized sinusoidal carriers, P(i,t), were produced according to Equation 1:

$$P\left(i,t\right)=\frac{\text{sin}\left(2\mathrm{*}\pi \mathrm{*}\left(f\left(i\right)\mathrm{*}t+U\left(\mathrm{0,1}\right)\right)\right)}{\sqrt{n_i}}$$ Equation 1

Where n_i represents the number of carriers, i is carrier index, f(i) is carrier frequency and U(0,1) represents a random number drawn from a uniform distribution within the interval (0,1). Carrier frequencies were uniformly spaced logarithmically within the frequency interval 100 to 6440 Hz at the selected carrier densities (C_density). Each carrier’s amplitudes were spectrotemporally modulated with a time-variant full wave rectified sinusoid according to Equation 2:

$$AM\left(i,t\right)=R_{depth}\mathrm{*}\left|\sin \left(\left(\frac{i\mathrm{*}{R}_{density}\mathrm{*}\pi }{C_{density}}\right)+\left(R_{rate}\mathrm{*}\pi \mathrm{*}t\right)+\mathrm{\Phi }\right)\right|$$ Equation 2

Where R_depth symbolizes modulation depth in dB (10 dB or 20 dB), R_density is ripple density in RPO, R_rate is temporal ripple modulation rate in cycles per second (Hz), C_density is carrier density in number per octave (33-tones/octave or 100-tones/octave) and Φ is the starting ripple phase (0, 90, 180 or 270 degrees). Note that the density of the carrier refers to the noise used to generate stimuli whereas ripple density refers to the spectrotemporally modulated envelope. Onset and offset ramps, 100 ms each, were applied to the modulated stimuli. Stimuli were created with two different noise carrier densities: 33 or 100 sine wave components per octave. Lower carrier density stimuli (33-tones/octave) have been used in prior studies (Aronoff and Landsberger 2013; Holden et al. 2016; Lawler et al. 2017; Zhou 2017; DiNino and Arenberg 2018b; Landsberger et al. 2018; Moberly et al. 2018). As previously reported, 33-tones/octave stimuli were found to be insufficient to represent spectral ripple densities beyond 10–11 RPO, whereas 100-tones/octave stimuli overcome this limitation (Resnick et al. 2020).

Two types of stimuli were generated and categorized based on ripple density as “target” (1 RPO – 19.027 RPO) or “no-target” (= 20 RPO). With increasing ripple density, target stimuli became harder to discriminate from the 20 RPO no-target referent. Figure 1 shows the time-waveforms and spectrograms for example stimuli.

Figure 1.

Time-waveforms and spectrograms of high carrier density (100-tones/octave) 2 RPO (a, b) and 8 RPO (c, d) “target” stimuli and 20 RPO (e, f) “non-target” stimulus.

Procedure.

A single-interval target / no-target observer-based psychoacoustic procedure (Horn et al. 2017b) was used to determine the highest ripple density that could be discriminated from 20 RPO in infants. Testing was conducted in a single-walled sound-attenuated booth with the infant seated on a caregiver’s lap facing a loudspeaker 1.6m away at roughly zero azimuth. A test assistant seated to the right of the infant used toys and age-appropriate play to direct infant attention forward. Adjacent to the loudspeaker was a reinforcement tower containing two mechanical animal toys obscured from view by dark plexiglass and illuminated on activation of the tower. A 15-inch monitor atop the tower played a four-second DVD clip once activated. An observer seated outside of the booth watched the infant through a glass window. All adults (test assistant, caregiver, and observer) were fitted with circumaural noise-reducing headphones playing masking sounds (repeated background no-target stimuli). Thus, the observer, test assistant and caregiver were all blinded to trial type. A one-way microphone-to-headphone transmission allowed the observer to talk to the assistant. Stimuli were presented via loudspeaker at 70 dBA.

The listeners’ task was to respond to target trials but not to no-target trials. The reason for including no-target trials in this single-interval procedure is to control for response bias, or the tendency of the listener-observer pair to respond. For instance, responses to every trial would achieve 100% target detection regardless of whether the listener could discriminate target and no-target trials. Correct and incorrect responses for target and no-target trials determine “hit” and “false alarm” rates which are used to compute a bias-free estimate of sensitivity. Between trials, repeating no-target stimuli were presented with an inter-stimulus interval of one second with spectral phase selected randomly for each presentation. Therefore, the no-target trials could not be distinguished from the stimuli presented between trials. Once the listener was calm and facing forward, the observer initiated a trial by clicking a button on the computer. On each trial, there was a 50% chance of a target. Listener behavior was studied by the observer to determine whether a target trial had occurred. Behavioral changes included head turn, eyebrow raise, startle response or brief pause in respiration and often varied by listener and throughout the session. The observer pressed a key to indicate a “target” decision. If correct on a target trial (a “hit”), the observer was notified on the screen and the listener was reinforced with activation of the reinforcement tower (four-second illumination of mechanical toys and/or DVD clip) along with social reinforcement by the assistant. If incorrect on a target trial (a “miss”), the observer was notified, and no reinforcement was given to the listener. For correct no-target (“correct rejection”) or incorrect no-target (“false alarm”) trials, the observer was notified but the reinforcement tower was not activated.

The procedure began with a “training phase” during which the easiest ripple density (1 RPO) was presented on target trials and reinforcement was given after every target trial regardless of listener response. This training phase was meant to teach the listener to associate target trials with the reinforcer and provided the observer with the opportunity to learn behavioral cues of the infant. Target trials were presented on 75% of trials and stimulus type varied pseudo-randomly in blocks of 8. Training concluded when the observer reached an 80% hit rate on the last 5 target trials or at least 20% correct rejection on the last 5 no-target trials. If a listener did not meet these criteria after a maximum of 20 training trials in up to two consecutive sessions, the infant was excluded and no further testing was performed.

After completion of training, listeners progressed to the “criterion” phase in which the easiest ripple density (1 RPO) was used for target trials, the probability of a target or no-target trial was 50%, and the listener was reinforced only for hits. Stimulus type varied pseudo-randomly in blocks of 30. For each block, ripple density of the target trials was fixed, and testing continued until an 80% hit rate over the last 5 target trials and at least an 80% correct rejection rate on the last 5 no-target trials was achieved. If the criterion was not met after a maximum of 30 trials in up to two consecutive test sessions, the infant was excluded from further testing and no threshold data was obtained.

In the test phase, blocks of up to 30 trials were administered. For each block, ripple density of the target trials was fixed, and testing continued until an 80% hit rate over the last 5 target trials and at least an 80% correct rejection rate on the last 5 no-target trials was achieved. The listener then moved to the next block where ripple density of the target was fixed at a higher level. The target density of each block increased by a factor of 2 up to 8 RPO. For the 33-tones/octave carrier condition, target density of each block increased by a factor of 1.414 beyond 8 RPO (8 RPO, 11.314 RPO, 16 RPO). For the 100-tones/octave carrier condition, target density of each block above 8 RPO increased by a factor of 1.189 (8 RPO, 9.514 RPO, 11.314 RPO, 13.454 RPO) and by a factor of 1.044 above 16 RPO (16 RPO, 16.708 RPO, 17.448 RPO, 18.221 RPO, 19.027 RPO). If 30 trials were administered at any ripple density without reaching the criterion, testing was ended. Blocks of 12 “reminder trials” were administered whenever four consecutive misses occurred. For reminder trials, the probability of a target trial was 50% and the ripple density on target trials was 1 RPO. Testing resumed at the highest ripple density that had been previously reached only if the observer correctly responded on 5/6 consecutive reminder trials. Otherwise, testing was aborted. Up to two sets of reminder trials were allowed for each ripple density. If testing was aborted, infants were retested starting with the highest ripple density they had previously reached either after a break or on a subsequent test visit. If the listener did not reach criteria after 30 trials and time permitted, additional sets of trials beginning one step easier than the previously highest ripple density tested were administered. This was done to obtain more data points near the listener’s presumed threshold. Infants were tested over three to four visits.

Adult listeners were tested using the same procedure as the infants with the modification of having no assistant in the booth and the instruction to raise a hand when they heard a sound that triggered the reinforcement tower. The observer – blind to trial type – recorded each hand-raise response. Adults completed testing in a single visit.

Infants and adults were tested using either the lower carrier density (33-tones/octave) or higher carrier density (100-tones/octave) stimuli. In both carrier density cohorts, listeners were tested at 10 dB and 20 dB modulation depth, in counterbalanced order. For each modulation depth, testing was performed until a threshold was obtained; listeners were then tested at the second modulation depth. Hit rate and false alarm rates were recorded for each ripple density tested using .txt files and transferred to Microsoft Excel (Version 16.23) files designed to compute and analyze psychometric functions. To control for observer bias, target and no-target trials were included in analysis to obtain a percent correct (p(C)_max) – an unbiased estimate of sensitivity defined as the probability of obtaining a Gaussian deviate that is less than d’/2 (Bargones et al. 1995; Werner and Boike 2001). For each participant and ripple depth condition, plots of p(C) as a function of RPO were constructed and fit with a logistic function. Threshold was computed as the stimulus level of difficulty (in RPO) at p(C) = 0.7. Functions with r² < 0.5 were treated as if no threshold had been obtained.

All statistical analyses were conducted using a full-factorial linear mixed model with compound symmetry covariance matrix in SPSS (Heck et al. 2011). One random effect (participant) and four fixed effects (test order, modulation depth, carrier density, and age group) were tested. Modulation depth was a repeated measures variable while test order, carrier density, and age group were between-subject variables. Estimated marginal means and pairwise comparisons were calculated for the two-way interactions relevant to the hypotheses (age × modulation depth and age × carrier density). For all main effects and interactions, findings were considered statistically significant if p < 0.05. Eta squared (ηp2) was calculated to reflect effect size for all significant effects.

Results

Twenty-three adults and 53 infants with NH were tested in this study. Thirteen adults and 36 infants were included in the low carrier density cohort; all adults and 33 infants (92%) in this cohort provided thresholds at both modulation depths. For 3 infants, thresholds were not obtained at one of the two modulation depths (1 in 10 dB, 2 in 20 dB condition) due to poorly fitting functions with r² < 0.5. Data from these infants were not included in analyses. Ten adults and 17 infants were included in the high carrier density cohort; all listeners in this cohort provided thresholds in 10 dB and 20 dB conditions.

In the low carrier density cohort (33-tones/octave), adult 10 dB thresholds ranged from 3.9 RPO to 10.5 RPO (mean = 6.3, SD = 1.9) and 20 dB thresholds ranged from 6.2 RPO to 11.3 RPO (mean = 8.8 RPO, SD = 1.8). Infants tested with the low carrier density stimuli had thresholds ranging from 1.4 RPO to 11.3 RPO at 10 dB (mean = 4.3, SD = 2.7); 20 dB thresholds ranged from 2.2 RPO to 15.5 RPO (mean = 7.2, SD = 3.5; Figure 2a).

Figure 2a.

Boxplot dynamic SRD thresholds for adults and infants at 10 dB (■) and 20 dB (☐) depths tested with 33-tones/octave stimuli. Mean thresholds indicated by “✕”. Significant effects indicated by “*”. SRD = Spectral ripple discrimination.

Figure 2b. Boxplot dynamic SRD thresholds for adults and infants at 10 dB (■) and 20 dB (☐) depths tested with 100-tones/octave stimuli. Mean thresholds indicated by “✕”. Significant effects indicated by “*”. SRD = Spectral ripple discrimination.

When tested with the higher carrier density stimuli (100-tones/octave), adult 10 dB thresholds ranged from 6.5 RPO to 16.1 RPO (mean = 11.8 RPO, SD = 2.8) and 20 dB thresholds ranged from 13.5 RPO to 17.2 RPO (mean = 15.5 RPO, SD = 1.5). In the infant cohort, 10 dB thresholds ranged from 2.8 RPO to 16.3 RPO (mean = 8.2 RPO, SD = 4.6) and 20 dB thresholds ranged from 3.8 RPO to 16 RPO (mean = 10.7 RPO, SD = 3.3; Figure 2b).

The linear mixed-model resulted in significant main effects for age [F(1, 61.23) = 21.48, p < 0.001, ηp2 = 0.260], noise carrier density [F(1, 61.23) = 56.21, p < 0.001, ηp2 = 0.479], and modulation depth [F(1, 58.96) = 32.63, p < 0.001, ηp2 = 0.348]. Test order was not a significant effect [F(1, 61.23) = 0.093, p = 0.671]. None of the interactions reached significance. Consistent with the non-significant age × noise carrier density interaction and non-significant age × modulation depth interaction, there was no evidence that effect of carrier density or modulation depth was different for infants and adults [p > 0.05].

Discussion

Previously reported data has suggested that infants’ spectral resolution is limited by immature spectral modulation sensitivity despite mature frequency resolution. The present study was motivated to strengthen confidence in this interpretation of the data by replicating earlier findings using a dynamic SRD task that minimizes within-channel, non-spectral, intensity cues (Aronoff and Landsberger 2013). As found with static SRD stimuli, infant performance was significantly poorer than adults, and both infants and adults showed better SRD thresholds at higher versus lower ripple depth (Horn et al. 2017b). Moreover, there was no significant difference in the strength of the effect of ripple depth on SRD in infants versus adults.

As described previously, the relationship between spectral modulation depth and ripple density define the spectral modulation transfer function (Supin et al. 1999). The axes and shape of this function depend on the way in which SRD is measured. If threshold on the y-axis is the smallest ripple depth a listener can discriminate at fixed ripple densities, the function is exponential with a y-intercept reflective of spectral modulation sensitivity and the slope independently reflective of frequency resolution (Supin et al. 1994, 1999). If threshold on the y-axis is the highest ripple density a listener can discriminate at fixed modulation depth (x-axis), the function is logarithmic with an x-intercept reflecting spectral modulation sensitivity and the slope reflective of frequency resolution (Horn et al. 2017a; Horn et al. 2017b). Any interaction between age and slope of the spectral modulation transfer function would suggest differences in frequency resolution between infants and adults. To the extent that spectral modulation transfer function shape can be extrapolated from two datapoints, the present data support the hypothesis that frequency resolution is mature in 6- to 7-month old infants and that the effect of age group on SRD is due to immature modulation depth sensitivity in infants (Horn et al. 2017b).

The third goal was to demonstrate the feasibility of obtaining two SRD thresholds in infants: Thresholds at 10 dB and 20 dB depths were obtained in 94% of infants tested in this study, showing a yield comparable to or higher than similar studies using the observer-based psychoacoustic procedure to measure thresholds in young listeners with NH and CIs (Olsho et al. 1987; Werner and Gillenwater 1990; Dasika et al. 2009; Horn et al. 2017b). Improved threshold yield with the dynamic SRD stimuli was immediately evident during pilot testing which led to adaptation of the dynamic SRD paradigm for other concurrent studies. However, neither direct testing of this hypothesis, nor exploring reasons for improved threshold yield with dynamic SRD stimuli were within the scope of the present study and this observation should be interpreted with caution. Future studies will work to determine test-retest reliability and validity of these measures with infants who use cochlear implants.

As discussed earlier, a post-hoc goal was to address the possibility that the age effect in SRD threshold was not obscured by the fact that 33-tones/octave carrier stimuli used by others (Landsberger et al. 2018; Moberly et al. 2018) does not sufficiently represent spectral ripple densities above 10–11 RPO. The present findings replicate and confirm the recent results of Resnick et al., in that listeners’ SRD thresholds were significantly higher with the 100-tones/octave carrier. The effect of carrier was observed for both infants and adult listeners and the lack of a significant age × carrier interaction suggest that infants’ spectral resolution extends up to and beyond the ceiling introduced by use of a low-density carrier and that their degree of immaturity is not underestimated by the ceiling affect.

The present findings are relevant to the development of methods to assess CI efficacy in young, deaf children using measures of spectral resolution. In adult listeners with CIs, various spectral resolution tasks have been shown to be predictive of speech understanding with a CI including detection of spectral modulation (Saoji et al. 2009; Anderson et al. 2012; Gifford et al. 2018) and discrimination of spectral modulation density or phase (Henry and Turner 2003; Henry et al. 2005; Won et al. 2007; Drennan et al. 2010; Anderson et al. 2011; Drennan et al. 2014; Drennan et al. 2015; Jeon et al. 2015; Horn et al. 2017a; Lawler et al. 2017; DiNino and Arenberg 2018b). These tests can be used to assess the efficacy of mapping or processing strategy alterations aimed at improving spectral resolution (Drennan et al. 2010; Smith et al. 2013; Goehring et al. 2019). However, the dichotomy in development of frequency resolution and sensitivity to spectral modulation may limit the usefulness of standard SRD tests for young listeners. There are two main reasons for this. First, until around age 9 to 12 years, sensitivity to spectral modulation will markedly limit performance on these tests (Peter et al. 2014; Kirby et al. 2015; DiNino and Arenberg 2018b; Landsberger et al. 2018). Thus, it would be impossible to attribute performance on an SRD task to a fundamental property of the listeners’ resolution of spectral peaks as opposed to immature spectral modulation sensitivity.

A second, and perhaps, more fundamental limitation to the usefulness of spectral resolution measures in young listeners is the lack of a clear understanding of whether frequency resolution or spectral modulation sensitivity are independent predictors of speech understanding in listeners with CIs. Horn et al. (2017a) found that spectral modulation transfer function slopes, but not x-intercepts, were significantly correlated with word recognition in speech shaped noise for both prelingually-implanted school-aged children and post-lingually implanted adults (Horn et al. 2017a). Importantly, the relationship between slope and word recognition was significantly weaker for children than adults. While this could be explained by the fact that the sample of children with CIs in that study all performed excellently on the word recognition task relative to the adults, it raises the question of whether frequency resolution could be less important for speech understanding with a CI in children than in adults (Landsberger et al. 2018).

In contrast to Horn et al. (2017a), other studies have found indirect evidence that modulation sensitivity at low modulation depths is most predictive of speech perception with a CI. Importantly, these studies were all conducted in adults with, presumedly, mature spectral modulation sensitivity. Litvak et al. (2007) varied modulation depths when testing adult listeners with CIs at fixed ripple densities (0.25 and 0.5 RPO) and found that vowel and consonant correlation was strongest at low modulation depth (Litvak et al. 2007). Saoji et al. (2009) expanded on these findings by testing at two additional fixed ripple densities (1 RPO and 2 RPO) and similarly found that detection thresholds for low modulation frequencies account for a significant proportion of the variance in phoneme recognition among adult listeners with CIs (Saoji et al. 2009). Additionally, Anderson et al., (2012) found evidence for the importance of low-spectral-density resolution for speech perception in adults with CIs (Anderson et al. 2012). To address these discrepancies, studies to independently measure frequency resolution and spectral modulation sensitivity and to determine the relative contribution of each to CI outcomes are currently underway in our lab.

There are some important limitations of the present study to address. Most importantly, as discussed above, extrapolating a logarithmic function from two datapoints must be done cautiously. While measuring the spectral modulation transfer function in older children is fairly straightforward (Peter et al. 2014; Horn et al. 2017a), reliable methods for obtaining functions in individual infants are more elusive (Bargones et al. 1995). This is mainly a result of difficulty in obtaining sufficient trials to measure several thresholds from an individual infant. This issue could be addressed by future studies where ripple depth is a between-subjects variable, and five or more groups of infants are tested at different ripple depths. Due to the authors’ ultimate interest in characterizing spectral resolution in individual infants, methods to obtain direct estimates of the x-intercept and slope of the spectral modulation transfer function in individual infants are currently being explored including non-behavioral methodologies such as electroencephalography or functional near-infrared spectroscopy (Spahr et al. 2014; Wiggins et al. 2016; Anderson et al. 2019; Bortfeld 2019).

Several other limitations should be emphasized. With the present study design, the hypothesis that frequency resolution is mature is supported by a lack of a significant effect (age × modulation depth interaction) and therefore, a small degree of immaturity could be missed due to insufficient power. In addition, it is important to note that the present data do not illuminate the reasons for slow maturation of spectral modulation sensitivity. Although this study was not designed to address the source of immaturity of spectral resolution in infants with NH, it is prudent to consider the possible interplay of multiple sensory and non-sensory factors including immature intensity coding, internal noise, and inattention (Werner and Boike 2001; Werner 2007; Werner et al. 2012). In standard SRD tasks, sensitivity can be confounded by listening efficiency. Psychoacoustic methods to control for immature listening efficiency, such as defining the slope of the spectral modulation transfer function, are important to understand how acoustic spectral resolution constrains speech perception abilities of listeners with CIs (Saoji et al. 2009; Anderson et al. 2012; Horn et al. 2017a; Isarangura et al. 2019). Because SRD testing requires some degree of sustained attention, it is plausible that poorer performance in infants and young children is due to inattentiveness rather than true age differences in spectral resolution (Allen and Wightman 1992; Bargones et al. 1995; Werner and Boike 2001). Reassuringly, modeling of the impact of inattention on psychometric functions has shown that it alone cannot entirely account for differences in detection thresholds between age groups (Allen and Wightman 1992; Werner and Boike 2001).

Conclusion

The present study demonstrates that spectral resolution in 6- to 7-month-old infants with NH is immature using a modified version of the spectral-temporally modulated ripple test. The pattern of results is consistent with the hypothesis that age differences in performance on this task in listeners with NH is due to immature spectral modulation sensitivity rather than frequency resolution.