ABSTRACT 
 
 
 
 
Title of Dissertation: EFFECTS OF INTERRUPTING NOISE AND 
SPEECH REPAIR MECHANISMS IN ADULT 
COCHLEAR-IMPLANT USERS  
  
 Brittany Nicole Jaekel, Doctor of Philosophy, 
2020 
  
Dissertation directed by: Professor Matthew J. Goupell, Department of 
Hearing and Speech Sciences 
 
Professor Rochelle S. Newman, Department of 
Hearing and Speech Sciences 
 
 
The long-term objective of this project is to help cochlear-implant (CI) users achieve better 
speech understanding in noisy, real-world listening environments. The specific objective 
of the proposed research is to evaluate why speech repair (?restoration?) mechanisms are 
often atypical or absent in this population. Restoration allows for improved speech 
understanding when signals are interrupted with noise, at least among normal-hearing 
listeners. These experiments measured how CI device factors like noise-reduction 
algorithms and compression and listener factors like peripheral auditory encoding and 
linguistic skills affected restoration mechanisms. We hypothesized that device factors 
reduce opportunities to restore speech; noise in the restoration paradigm must act as a 
plausible masker in order to prompt the illusion of intact speech, and CIs are designed to 
attenuate noise. We also hypothesized that CI users, when listening with an ear with better 
  
peripheral auditory encoding and provided with a semantic cue, would show improved 
restoration ability. The interaction of high-quality bottom-up acoustic information with 
top-down linguistic knowledge is integral to the restoration paradigm, and thus restoration 
could be possible if CI users listen to noise-interrupted speech with a ?better ear? and have 
opportunities to utilize their linguistic knowledge. We found that CI users generally failed 
to restore speech regardless of device factors, ear presentation, and semantic cue 
availability. For CI users, interrupting noise apparently serves as an interferer rather than a 
promoter of restoration. The most common concern among CI users is difficulty 
understanding speech in noisy listening conditions; our results indicate that one reason for 
this difficulty could be that CI users are unable to utilize tools like restoration to process 
noise-interrupted speech effectively.  
  
  
 
 
 
 
 
 
 
 
 
 
EFFECTS OF INTERRUPTING NOISE AND SPEECH REPAIR MECHANISMS 
IN ADULT COCHLEAR-IMPLANT USERS  
 
 
 
by 
 
 
Brittany Nicole Jaekel 
 
 
 
 
 
Dissertation submitted to the Faculty of the Graduate School of the  
University of Maryland, College Park, in partial fulfillment 
of the requirements for the degree of 
Doctor of Philosophy 
2020 
 
 
 
 
 
 
Advisory Committee: 
Professor Matthew J. Goupell, Chair 
Professor Rochelle S. Newman 
Professor Jan Edwards 
Professor Samira Anderson 
Professor Stefanie Kuchinsky 
Professor Catherine Carr 
 
  
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
? Copyright by 
Brittany Nicole Jaekel 
2020 
 
 
 
 
 
 
 
 
 
 
 
 
 
  
 
Acknowledgements 
Sarah Weinstein, Alyssa Giammetta, Maureen Shader, Kristina Milvae, Nicole Nguyen, 
Olga Stakhovskaya, Julie Cohen, Elizabeth Kolberg, Ginny Alexander, Will Bologna, 
Zilong Xie, Emily Shroads, Debbie Moon, Emma Peterson, and Kelly Miller.  
 
NIDCD F31DC017362 (BNJ) and NIDCD R01DC014948 (MJG). 
ii 
 
 
 
Table of Contents 
 
 
Acknowledgements ............................................................................................................. ii 
Table of Contents ............................................................................................................... iii 
List of Tables ...................................................................................................................... v 
List of Figures .................................................................................................................... vi 
Chapter 1: Introduction ....................................................................................................... 1 
1.1 Speech-in-noise perception with cochlear implants ................................................. 1 
1.1.1 Factors affecting speech-in-noise perception ..................................................... 4 
1.1.2 Front-end preprocessing in cochlear implants ................................................... 9 
1.1.2.1 Front-end preprocessing in Cochlear-brand CIs ....................................... 10 
1.2 Perceptual restoration.............................................................................................. 18 
1.2.1 Early perceptual restoration studies ................................................................. 20 
1.2.2 More recent perceptual restoration studies ...................................................... 27 
1.2.3 Perceptual restoration with degraded input ...................................................... 29 
1.2.3.1 Perceptual restoration in CI users ............................................................. 30 
1.2.3.2 Perceptual restoration in CI simulations (vocoding studies) .................... 34 
1.2.3.3 Perceptual restoration in other populations with hearing impairment ...... 39 
1.3 Summary and Hypotheses....................................................................................... 41 
Chapter 2: Experiment 1: The impact of front-end preprocessing algorithms on 
restoration ......................................................................................................................... 45 
2.1 Methods................................................................................................................... 47 
2.1.1 Participants ....................................................................................................... 47 
2.1.2 Stimuli .............................................................................................................. 49 
2.1.3 Equipment ........................................................................................................ 50 
2.1.4 Procedure ......................................................................................................... 53 
2.2 Results ..................................................................................................................... 54 
2.3 Discussion ............................................................................................................... 61 
Chapter 3: Experiment 2: Measuring effects of presentation level and peripheral auditory 
encoding on restoration ..................................................................................................... 66 
3.1 Methods................................................................................................................... 68 
3.1.1 Participants ....................................................................................................... 68 
3.1.2 Stimuli .............................................................................................................. 71 
3.1.3 Equipment ........................................................................................................ 73 
3.1.4 Procedure ......................................................................................................... 73 
3.2 Results ..................................................................................................................... 74 
3.2.1 Main Findings .................................................................................................. 74 
3.2.2 Control Conditions ........................................................................................... 80 
3.2.3 Poorer Ear Performance ................................................................................... 82 
3.3 Discussion ............................................................................................................... 82 
Chapter 4: Experiment 3: Measuring interactions of peripheral auditory encoding and 
prior linguistic knowledge on restoration ......................................................................... 86 
4.1 Methods................................................................................................................... 87 
4.1.1 Participants ....................................................................................................... 87 
4.1.2 Stimuli .............................................................................................................. 91 
iii 
 
 
 
4.1.3 Equipment ........................................................................................................ 93 
4.1.4 Procedure ......................................................................................................... 94 
4.2 Results ..................................................................................................................... 95 
4.2.1 Multilevel Model Analysis .............................................................................. 95 
4.2.2 Cluster Analysis ............................................................................................. 104 
4.3 Discussion ............................................................................................................. 110 
Chapter 5: Experiment 4: Replicating previous work with CI users and perceptual 
restoration ....................................................................................................................... 116 
5.1 Methods................................................................................................................. 117 
5.1.1 Participants ..................................................................................................... 117 
5.1.2 Stimuli ............................................................................................................ 120 
5.1.3 Procedure ....................................................................................................... 121 
5.2 Results ................................................................................................................... 123 
5.2.1 Baseline speech understanding and its correlation with restoration benefits 128 
5.2.2 Are the IEEE sentences ?restorable?? ............................................................ 129 
5.3 Discussion ............................................................................................................. 130 
Chapter 6: Conclusion..................................................................................................... 135 
6.1 Device factors affecting perceptual restoration in CI users .................................. 136 
6.2 Listener factors affecting perceptual restoration in CI users ................................ 138 
6.2.1 Restoration in prelingual vs. postlingual adult cochlear implant users ......... 139 
6.3 Replicating previous work in restoration .............................................................. 141 
6.4 Limitations ............................................................................................................ 142 
6.5 Future Directions .................................................................................................. 143 
6.6 Final Conclusion ................................................................................................... 144 
Bibliography ................................................................................................................... 146 
 
 
 
   
 
 
 
 
 
 
iv 
 
 
 
List of Tables 
 
Table 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 
Table 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 
Table 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 
Table 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 
Table 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 
Table 6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 
Table 7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 
Table 8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 
Table 9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 
Table 10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 
Table 11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 
Table 12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 
Table 13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128 
v 
 
 
 
List of Figures 
 
Figure 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 
Figure 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 
Figure 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 
Figure 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 
Figure 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 
Figure 6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 
Figure 7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98 
Figure 8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .100 
Figure 9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .101 
Figure 10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .102 
Figure 11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .103 
Figure 12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .106 
Figure 13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .110 
  
vi 
 
 
 
Chapter 1: Introduction 
To introduce the broad topic of this dissertation, I first discuss how cochlear implant 
(CI) users are generally impacted by noise and how this affects speech understanding in 
noisy listening environments. I then describe the ways in which CI manufacturers have 
attempted to combat the speech-in-noise problem by adding noise-reduction software to CI 
sound processors. Finally, I discuss a mechanism by which normal-hearing (NH) listeners 
often repair noise-interrupted speech, called ?perceptual restoration,? and follow this by 
reporting how signal degradations introduced by CI processing may decrease perceptual 
restoration. My goal for this project is to determine the extent to which we can increase 
speech repair in CI users to ultimately improve communication in everyday listening 
scenarios.  
 
1.1 Speech-in-noise perception with cochlear implants 
The CI is an auditory prosthesis that can restore a sensation of sound. The device 
works by transforming acoustic signals into electrical pulses, which directly stimulate the 
spiral ganglion cells and auditory nerve fibers in the inner ear. The Food and Drug 
Administration (FDA) has approved cochlear implantation in people with severe-to-
profound hearing loss for whom hearing aids are no longer useful. The main purpose of the 
CI is to restore the user?s ability to understand speech. 
CIs have been approved for implantation in people of nearly all ages, from children 
as young as one year old to elderly adults ("Fact Sheet: Cochlear Implants," 2010). 
Generally, the CI is successful in providing the user speech understanding in quiet listening 
environments (Fetterman & Domico, 2002), dramatically improving the user?s ability to 
1 
 
 
 
communicate through oral language. However, in noisy1 listening environments like caf?s 
and classrooms, CI users? speech understanding is often reduced. For example, in one 
study, CI users were only able to understand 64% of sentences in a noisy listening 
environment, even when the sentences were much louder than the background noise 
(Balkany et al., 2007). Similar findings have been found for speech recognition in both 
adult and child CI users (A. Caldwell & Nittrouer, 2013; Fetterman & Domico, 2002), and 
such performance is quite low compared to how a NH listener would perform under 
comparable conditions (A. Caldwell & Nittrouer, 2013; Kaandorp, Smits, Merkus, 
Goverts, & Festen, 2015; Loizou et al., 2009). In a study by Kaandorp et al. (2015), for 
example, NH adults achieved speech reception thresholds (SRTs) of ?4.2 in noisy 
backgrounds, much lower (i.e., better) than CI users? SRTs of +8.0.2 Therefore, it appears 
that CIs are highly susceptible to noise interference, reducing their utility in day-to-day 
listening environments (Fetterman & Domico, 2002; Fu & Nogaki, 2005; Sladen & 
Zappler, 2015; Zhao, Stephens, Sim, & Meredith, 1997). To underscore the importance of 
this problem, in an open-ended questionnaire about the hearing concerns of CI users, 
difficulty hearing speech in noise was the most reported problem (Zhao et al., 1997), and 
has continued to be reported as a major problem even with more recent technology 
(Faulkner & Pisoni, 2013; Gomersall, Baguley, & Carlyon, 2019; Lassaletta, Castro, 
Bastarrica, de Sarria, & Gavilan, 2006).  
                                                 
1 For the purposes of this dissertation, ?noise? will refer to any unwanted sound (Moore, 2012), including 
signals like competing speech.  
2 As a comparison with other groups experiencing non-normal hearing, listeners with bilateral hearing aids 
had SRTs of +0.5 for digits in noise, while NH listeners had SRTs of ?9.1 (Bost, Versfeld, & Goverts, 
2019). 
2 
 
 
 
Several potential factors can reduce a CI user?s ability to understand speech in noisy 
conditions. These factors will be discussed below. First, however, it is important to stress 
that a CI does not provide a perfect imitation of a functioning, healthy human cochlea, and 
thus aspects of CI software and hardware can introduce unique problems when users need 
to communicate in noise. To highlight this, a discussion of how a CI works is essential.  
CIs translate acoustic information to electrical information in the following way. 
Speech sounds are captured by the CI?s microphone(s) and then passed through various 
software algorithms, undergoing a process called ?pre-emphasis? (Loizou, 2006). 
Proprietary ?pre-emphasis? algorithms, specific to each CI manufacturer, work to alter and 
enhance the incoming acoustic signals in a way that is meant to provide the best, most 
perceptible signal possible to the CI user. A deeper discussion of these ?pre-emphasis? 
algorithms will appear below. Next, these pre-emphasized acoustic signals are sent through 
a bank of bandpass filters. Each bandpass filter captures acoustic information within a 
specific band of frequencies; this method is meant to imitate the frequency-based 
organization of the human cochlea, which also processes an incoming acoustic signal based 
on frequency. Next, the temporal envelopes of the bandpass filtered signals are extracted. 
The temporal envelope of a signal is the slow change in amplitude over time, and is one 
type of timing information available in an acoustic signal. Another type of timing 
information is called ?temporal fine structure.? Temporal fine structure encodes the fast-
moving amplitude and phase changes in a signal. CI processing typically discards this 
temporal fine structure and instead only preserves the temporal envelope (Loizou, 2006). 
Next, compression is applied to each extracted temporal envelope. The goal of compression 
is to reduce the changes in amplitude (i.e., intensity) present in the signal to a small specific 
3 
 
 
 
range. While speech signals have large changes in amplitude, a typical CI user can safely 
utilize only a small range of amplitude changes. Very quiet, just barely audible sound levels 
are called ?thresholds? and uncomfortably, almost painfully loud levels are called 
?uncomfortable loudness levels.? While this range in loudness from threshold to 
uncomfortably loud is quite large in NH listeners ? up to approximately 140 dB (Billings 
& Gray, 1972) ? this range can be as small as 5 dB in some CI users (Loizou, 2006). Thus, 
compressing the range of possible amplitudes into a CI-user-specific range is important in 
order to allow some perceptible variation in speech amplitude to be available to CI users. 
Having access to these changes in speech amplitude is important for discriminating among 
speech sounds and syllables (Greenberg, Carvey, Hitchcock, & Chang, 2003), especially 
consonants (Nie, Barco, & Zeng, 2006; Xu, Thompson, & Pfingst, 2005). Next, the 
compressed temporal envelopes modulate a series of electric pulse trains, and these trains 
are specific to each frequency band. Finally, these envelope-modulated electric pulse trains 
are transmitted to the CI user?s electrodes, and the pulse trains stimulate the surviving spiral 
ganglion cells and auditory nerve fibers of the cochlea (Loizou, 2006).  
 With the idea of how a CI typically works in mind, we can now consider how the 
processing of speech sounds to electrical impulses may be impacted by noise. The three 
main factors that I will cover are (1) poor spectral resolution, (2) duration of deafness, and 
(3) age. 
 
1.1.1 Factors affecting speech-in-noise perception 
  The CI?s poor spectral resolution is one potential cause of poor speech-in-noise 
understanding among CI users. Spectral resolution refers to how much frequency 
4 
 
 
 
information about the signal is available; since the signal processing algorithms in the CI 
discard fine spectrotemporal detail and there are limited channels available to present 
envelope information, resolution is typically low (Loizou, 2006). While this low resolution 
may not adversely affect speech understanding in quiet environments, it appears to have a 
dramatic negative effect on speech understanding in noisy environments. For example, 
Dorman, Loizou, Fitzke, and Tu (1998) showed that in NH listeners presented a CI 
simulation (that is, a vocoder), only four channels of spectral information were needed to 
obtain high sentence understanding in quiet. However, more channels were needed to 
understand speech in a background of constant speech-shaped noise. When speech was 
more intense than the background noise, 12 channels were needed; when background noise 
was more intense than speech, at least 20 channels were needed (Dorman et al., 1998). 
Thus, a lack of spectral resolution ? access to several channels of frequency information ? 
appears to reduce speech-in-noise understanding. This may be because with better spectral 
resolution, a listener can more easily distinguish speech from noise, as the two are less 
perceptually similar (Jin, Nie, & Nelson, 2013; O'Neill, Kreft, & Oxenham, 2018). 
While some CIs can provide almost two dozen electrodes to the user, research has 
shown that the number of ?effective? channels ? that is, the number of channels from which 
the user can actually derive benefit ? at any one time is approximately 8 channels (Friesen, 
Shannon, Ba?kent, & Wang, 2001)3. This is much lower than the 20 CI simulated-channels 
                                                 
3 However, see more recent reports by Croghan, Duran, and Smith (2017) and Berg et al. (2019) for a different 
result, where up to 22 or 16 channels of information (respectively) was useful for CI users perceiving speech 
in noise. These results may be influenced by CI participants having access to newer CI technology (e.g., pre-
curved electrode arrays that have better electrode/nerve interface due to more effective alignment with the 
modiolus) and better neural survival due to trends in earlier implantation and/or looser criteria for 
implantation in recent years. CI users with older technology and who were implanted less recently may not 
be able to achieve such a level of performance.  
5 
 
 
 
NH listeners needed in loud noise as reported by Dorman et al. (1998). The amount of 
spectral information that can be provided by the activated CI electrodes is therefore not 
necessarily the same as the amount of actual spectral detail a CI user can receive. The 
health and proximity of the spiral ganglion cells upon which electrodes are supposed to 
fire, and the amount of current spread from the electrodes, also affect one?s effective 
spectral resolution (Fu & Nogaki, 2005; Fu, Shannon, & Wang, 1998; Oxenham & Kreft, 
2014). Current spread occurs when electrical current from the electrodes unintentionally 
spreads to neurons beyond the target neuron population. When this happens, each electrode 
is contributing less independent sound information, and resolution suffers.   
Beyond affecting speech understanding in steady-state noise environments, poor 
spectral resolution can also reduce a CI user?s ability to obtain release from masking in 
modulated (i.e., fluctuating) noise (Oxenham & Kreft, 2014). Masking can be defined as a 
situation where a sound of interest is ?made inaudible by the presence of other sounds? 
(Moore, 2012). Release from masking, then, indicates situations where the sound of interest 
once again becomes audible, even in the presence of the masking sounds. NH listeners are 
able to ?listen in the dips? of a fluctuating noise masker, gleaning clear bits of speech 
information with which they can build up their perception of the speech stream in the brief 
moments when noise volume is reduced. Compared to NH listeners, CI users show less 
release from masking in a modulated noise masker (Nelson & Jin, 2004; Nelson, Jin, 
Carney, & Nelson, 2003). The fact that CI users cannot apparently ?listen in the dips? 
underscores the point that CI users generally struggle to access speech information in noisy 
listening situations to the same degree as NH listeners. Oxenham and Kreft (2014) reported 
that CI users? poor spectral resolution and the effects of current spread contribute to noise 
6 
 
 
 
perception differences between CI users and NH listeners, with CI users experiencing 
smoother temporal envelopes of noise and fewer fluctuations. Having smoother temporal 
envelopes means CI users have less access to envelope modulations, which are critical for 
conveying speech information. 
 Attributes of the CI user him- or herself could also be influencing how well he or 
she can understand speech in noise. Fetterman and Domico (2002) found that duration of 
profound deafness was predictive of how well the CI users were able to understand speech 
in noise. ?Durations of deafness? refer to periods during which CI users received almost 
no audible acoustic input. Fetterman and Domico (2002) concluded that long durations of 
reduced audible input might result in anatomical and/or physiological changes to the 
auditory system, which in turn reduce one?s ability to understand speech in noise, perhaps 
through the deterioration of auditory neurons. Duration of deafness has been shown to be 
associated with reduced speech understanding in quiet as well, indicating that it is an 
important factor for hearing-related outcomes in CI users in general (Blamey et al., 2013; 
Green et al., 2007; Holden et al., 2013).  
 Another attribute of CI users that could be affecting their ability to understand 
speech in noise is age. Sladen and Zappler (2015) found that CI users older than 60 years 
had significantly worse speech understanding in noise and quiet compared to younger CI 
users. Both age groups were matched for duration of deafness and length of CI use; 
therefore, age itself appeared to have a unique influence on speech perception. Aging could 
affect factors like cognition, particularly processing speed, which is important for quickly 
and accurately processing speech and which slows with age (Lin et al., 2012). Aging is 
also associated with central auditory pathway changes. For example, aging can reduce 
7 
 
 
 
accurate and efficient encoding of incoming auditory signals in the central auditory nervous 
system (Anderson, Parbery-Clark, White-Schwoch, & Kraus, 2012; Gordon-Salant, 
Fitzgibbons, & Yeni-Komshian, 2011; Ouda, Profant, & Syka, 2015).  
 To summarize, difficulty understanding speech in noise is the number one concern 
among CI users (Zhao et al., 1997). The main causes of this difficulty appear to be related 
to poor spectral resolution, or a lack of access to clear frequency information. The CI?s 
signal processing removes fine spectrotemporal detail (Loizou, 2006), and the effective 
number of frequency channels available to CI users at any one time is approximately eight 
(Friesen et al., 2001), much lower than what was needed for NH listeners presented a CI 
simulation to successfully perceive speech in noise (Dorman et al., 1998). The number and 
health of neurons in the auditory nerve, their proximity to the stimulating electrode, and 
the extent of current spread can also impact spectral resolution. In situations where release 
from masking should be possible, that is, in fluctuating noise (during which listeners should 
be able to glean speech information during ?dips? in the noise), reduced spectral resolution 
prevents CI users from showing speech understanding improvements (Jin et al., 2013; 
Nelson & Jin, 2004; Nelson et al., 2003; Oxenham & Kreft, 2014).  
Beyond spectral resolution, duration of deafness (Fetterman & Domico, 2002) and 
age (Sladen & Zappler, 2015) could also be reducing speech-in-noise understanding in CI 
users. Longer durations of deafness and advanced age could result in physiological and 
anatomical changes along the auditory pathway, decreasing its ability to clearly encode 
speech signals in noise, and aging could additionally impair the cognitive skills crucial for 
processing speech.  
 
8 
 
 
 
1.1.2 Front-end preprocessing in cochlear implants 
In the US, three CI manufacturers are currently approved by the FDA for 
implantation in humans: Cochlear Ltd., MED-EL, and Advanced Bionics (Loizou, 2006). 
Choice of CI brand is up to the patient, though the patient may be directed toward a certain 
brand based on surgeon or hospital preference. Each of the three CI manufacturers offer 
models with similar components, though the specific workings of these components may 
differ. For example, while all CIs contain an array of intracochlear electrodes for encoding 
sound information with amplitude-modulated electrical pulse trains, the number of 
electrodes can vary based on manufacturer. Similarly, while all CIs provide microphones, 
the number of microphones and their location on the head may differ across brands 
(Chaikof, 2016). In addition, while all CIs utilize proprietary software strategies in their 
speech processors, the exact details of these strategies are not public and are specific to 
each brand.  
All three manufacturers have attempted to mitigate the ?speech-in-noise problem? 
in CIs. One way they have done so is by implementing various sound pre-processing 
algorithms. These algorithms may attempt to remove noise from incoming speech signals 
prior to sending signals to the electrodes; they can also adjust aspects of the CI microphones 
and other settings in an attempt to make target speech information more salient. In this 
paper, these various algorithms will be referred to as ?front-end preprocessing.?  
As mentioned previously, each CI manufacturer has their own proprietary front-
end preprocessing strategy. Cochlear Ltd. advertises front-end preprocessing strategies like 
SmartSound iQ, Wind Noise Reduction (WNR), and Signal-to-Noise-Ratio Noise-
Reduction (SNR-NR), all of which are meant to improve speech understanding in noisy, 
9 
 
 
 
non-optimal listening environments ("The Cochlear? Nucleus? Implant System," 2018). 
Similarly, MED-EL advertises Automatic Sound Management and Wind Noise Reduction 
("Cochlear Implants," 2017) and Advanced Bionics advertises ClearVoice Speech 
Enhancement Technology and AutoSound ("AB?s Exclusive Sound Processing 
Technologies," 2018), all of which are also meant to boost and improve speech signals in 
background noise. Again, the point of all these front-end preprocessing strategies is to 
partially ?solve? the speech-in-noise problem in CIs.  
While speech understanding in noise generally improves with front-end 
preprocessing (Davidson, Geers, & Brenner, 2010; Gifford & Revit, 2010; Rakszawski, 
Wright, Cadieux, Davidson, & Brenner, 2016; Wolfe et al., 2015), it is possible that 
individualized algorithm adjustments could improve speech-in-noise perception further. 
This section will detail front-end preprocessing in Cochlear-brand CIs, as well as research 
on how these algorithms affect speech understanding in noise. We focus on Cochlear-brand 
CIs because they are the most common device among the CI population to which we have 
access. Detailed descriptions of front-end preprocessing that is likely to be engaged during 
a speech repair paradigm (which involves listening in noise) will be the focus of this 
section, though other programs will be mentioned briefly. 
 
1.1.2.1 Front-end preprocessing in Cochlear-brand CIs 
The first strategy I will discuss is Adaptive Dynamic Range Optimization, or 
ADRO. ADRO aims to place incoming speech signals into the CI user?s dynamic range 
(Wolfe et al., 2015). The dynamic range is a range of loudness, stretching from the CI 
user?s threshold, or the quietest sound they can hear, to the CI user?s maximum loudness 
10 
 
 
 
level, or the loudest sound they can hear and still feel it is not uncomfortably loud (Gifford 
& Revit, 2010). Because CI users typically have very small dynamic ranges compared to 
NH listeners (Blamey, 2005), without ADRO, they could miss important speech 
information. This is because signals less intense than the user?s threshold and more intense 
than their maximum loudness level would fail to be transmitted.  
With ADRO, if a sound in the environment is quieter than the CI user?s threshold 
level, ADRO will boost the volume of that sound, and conversely will reduce the intensity 
of a sound that exceeds the CI user?s maximum loudness level (Blamey, 2005). ADRO 
adjusts signals within each individual channel, providing a fine-grained adjustment in 
intensity levels (Gilden, Lewis, Grant, & Crosson, 2015).  
Another strategy that responds to loudness is called Automatic Sensitivity Control, 
or ASC. This strategy adjusts the sensitivity of the microphone(s) of the CI, depending on 
the level of the ambient noise in the listening environment (Gifford & Revit, 2010). With 
ASC turned off, sounds are encoded linearly between 25- and 65-dB SPL, after which 
?infinite? compression turns on. Compression distorts speech signals and makes them more 
difficult to understand (Rakszawski et al., 2016). With ASC turned on, at very low values 
of microphone sensitivity, sounds below the CI user?s threshold level are not encoded, and 
sounds above the CI user?s loudest comfortable sound level are compressed (Wolfe et al., 
2015). As microphone sensitivity increases, the sound level at which compression is 
engaged decreases, as does the sound level at which signals are not encoded. These 
microphone sensitivity changes are slow acting, and thus not abrupt (Mauger, Warren, 
Knight, Goorevich, & Nel, 2014; Wolfe et al., 2015). ASC acts on all channels; it does not 
make adjustments on individual channels like ADRO (Yathiraj & Rao, 2013).  
11 
 
 
 
Another strategy that is related to sound levels is called Whisper. While ASC is 
slow acting, Whisper is fast to respond to incoming signals (Mauger et al., 2014), boosting 
the level of quiet or distant signals by 10 dB. This strategy is useful in quiet listening 
environments, and has been shown to result in significantly better speech understanding in 
quiet compared to when this program is turned off (Mauger et al., 2014). 
ADRO, ASC, and Whisper are all types of Automatic Gain Control (AGC). As 
stated above, CI users have narrow dynamic ranges, approximately stretching 10 to 20 dB 
(Khing, Swanson, & Ambikairajah, 2013), which makes encoding the dynamic range of a 
typical talker, 40 to 50 dB (Zeng et al., 2002), difficult. AGC aims to compress sounds in 
the environment into CI users? narrow dynamic ranges (Spencer, Tillery, & Brown, 2018). 
ADRO and ASC are ?slow-acting? AGCs, in that they respond over the course of several 
seconds to changes in level in the listening environment; Whisper is ?fast-acting,? meaning 
it can react to changes in the environment on the order of milliseconds (Khing et al., 2013). 
Fast-acting compression can help shield CI users from loud, brief noises, but may introduce 
discontinuities or distortions in the amplitude of the incoming signals (Ba?kent, Eiler, & 
Edwards, 2009; Khing et al., 2013). 
A strategy targeting noise directly is Signal-to-Noise-Ratio Noise Reduction (SNR-
NR), which is part of the SmartSound iQ system. SNR is defined as the ratio of the level 
of the target signal (usually speech) to the level of the noise, with positive SNRs meaning 
target sounds are more intense than noise, and negative SNRs meaning target sounds are 
less intense than noise. SNR-NR works by flagging CI channels that seem to be encoding 
steady-state noise (Wolfe et al., 2015). Working HVAC systems and car engines are 
examples of steady-state noise producers. The gain of noisy channels is then reduced 
12 
 
 
 
?instantaneously? (Gilden et al., 2015), as it is assumed that these channels do not contain 
important acoustic information. SRTs improved in CI users using an early version of SNR-
NR, compared to using no SNR-NR strategy, in a study by Dawson, Mauger, and Hersbach 
(2011). In that study, SRTs improved across all three types of noise tested: speech-
weighted noise, party noise, and city noise. These latter two noises were real recordings 
made at a cocktail party and beside a city street, respectively. However, Dawson et al. 
(2011) noted a large amount of variability among their participants in terms of SNR-NR 
benefit, in that some participants showed poorer SRTs with SNR-NR turned on, especially 
in the more ?dynamic? noises ? party noise and city noise. Mauger et al. (2014) concluded 
that SNR-NR was most effective in listening environments where background noise is 
continuous rather than modulated.  
Another strategy targeting noise is Wind Noise Reduction (WNR), which is also 
part of the SmartSound iQ system. This feature attempts to reduce the harmful effects of 
wind noise on incoming signals, and could be useful during bicycle riding or when working 
outdoors (Gilden et al., 2015). If wind noise is detected, WNR changes microphone settings 
and reduces the intensity of low-frequency information (that is, frequency information 
below 400 Hz, which is associated with wind noise) that is provided to the CI user (Mauger 
et al., 2014; Wolfe et al., 2015). Note that some speech energy occurs in this frequency 
range as well, and thus could be affected by the WNR algorithm. Wind noise is detected 
by comparing microphone inputs on the CI; if these signals are decorrelated, wind is likely 
present. Typically, slowly modulated signals like speech should be correlated at the CI 
microphones, since the microphones are situated close together on the device.  
13 
 
 
 
Different microphone settings in general can be helpful in different listening 
environments, and automatic changes to microphone settings have already been shown to 
occur with ASC and WNR. They also occur with SCAN, which is discussed next. With the 
Fixed directional microphone setting (also known as ?Zoom?), sounds at specific locations 
away from the direction in which the CI user is looking are attenuated by 15-20 dB (Gilden 
et al., 2015). This reduces the intensity of sounds that could be distracting to the CI user, 
and which are not in his or her direct field of view. With the Adaptive directional 
microphone setting (also known as ?Beam?), the point of attenuation is no longer fixed, 
and can adapt to new locations, meaning it can adapt to movements of a distracting noise 
source (Gilden et al., 2015). Finally, with the Standard directional microphone setting, no 
major attenuation in sounds occurs?attenuation mimics the attenuation experienced by 
NH listeners, in that sounds toward the side and back of the CI user are reduced slightly in 
level (Gilden et al., 2015).  
The last strategy I will discuss is SCAN, which is a scene analyzer algorithm and 
part of the SmartSound iQ system. SCAN attempts to detect and classify the type of 
listening environment the CI user is presently in, and based on this classification alters 
aspects of sound processing to provide the best listening experience for the CI user (Biever, 
Gilden, Zwolan, Mears, & Beiter, 2018; Wolfe et al., 2015). Scene types included in SCAN 
are Speech, Speech in Noise, Quiet, Music, Wind, and Noise. According to Wolfe et al. 
(2015), SCAN classifies environments by analyzing intensity (loudness), spectral 
characteristics (e.g., voice pitch), temporal characteristics (e.g., timing information and 
rhythm), and modulation (changes in intensity across time). Based on the scene 
classification, microphone settings are changed. For example, when the Speech in Noise 
14 
 
 
 
scene is detected, the Adaptive microphone (Zoom) is turned on; when Speech or Music 
scenes are detected, the Standard microphone setting is turned on (Mauger et al., 2014).  
Gilden et al. (2015) tested whether SCAN improved speech understanding in 40 
unilateral CI users, aged 13.2 to 81.2 years. These CI users were tested on whether their 
speech understanding in noise improved with the Nucleus 6 sound processor (which 
includes SCAN) compared to their regular Nucleus 5 sound processor (which does not 
include SCAN). Speech understanding was tested with sentences presented at 60 dB-A 
from a loudspeaker directly in front of the participant. Speech-weighted noise was 
presented at 55 dB-A from a loudspeaker at 90 degrees from the front of the participant, 
on the side of the participant?s CI. The study involved two visits. At the first visit, CI users 
achieved an average 21.2% correct speech understanding in noise with their usual Nucleus 
5 settings, and an average 53.1% correct speech understanding with Nucleus 6/SCAN, a 
large improvement. After this first visit, the participants took the Nucleus 6 home and used 
it as their new default CI for four weeks. At the end of the four weeks, the participants 
completed their second visit, and their speech-in-noise understanding was tested again. 
With the Nucleus 6/SCAN, participants achieved an average 56.4% correct speech 
understanding, indicating a stable effect of the benefit of the new processing strategy and 
device for speech understanding. Subjectively, nearly all of the participants reported that 
they felt that the Nucleus 6/SCAN improved their speech understanding in quiet and noise, 
improved their ability to localize sound, and improved sound quality and ease of listening 
compared to the Nucleus 5.  
Mauger et al. (2014) tested whether several of the new features in the Nucleus 6 
improved speech understanding in 21 CI users who wore Nucleus 5 as their default sound 
15 
 
 
 
processor. All participants were post-lingually deafened (i.e., experienced hearing 
impairment after learning oral language), aged 49 to 90 years, and had worn their CI for 1 
to 10 years. Prior to testing, all participants were asked to wear a Nucleus 6 sound processor 
for a minimum of two weeks. Participants were then tested five separate times, with each 
experimental session occurring one week apart, on various speech materials in various 
noise environments. Participants were tested with the default and their preferred Nucleus 
5 settings and with six different Nucleus 6 settings. Speech materials included 
monosyllabic words in quiet and the Australian sentence test in noise. From this latter test, 
SRTs were calculated by presenting speech at 65 dB SPL and adjusting the background 
noise level until a level was reached at which participants achieved 50% correct sentence 
understanding. Lower SRTs were indicative of better speech understanding. The Australian 
sentence test was presented with two types of noise (speech-weighted noise or 4-talker 
babble noise) at two possible noise locations (noise at the same location as the target speech 
[co-located at 0 degrees] or noise at a different location from the target speech [not co-
located, presented simultaneously at 90, 180, and 270 degrees from the target, which 
remained at 0 degrees]).  
For speech in quiet, the best performance was with the Nucleus 6/Whisper setting. 
Since the Whisper program boosts quiet and distant sounds, it makes sense that Nucleus 
6/Whisper provided the best speech-in-quiet performance. For SRTs in co-located 
speech-weighted noise, Nucleus 6/SCAN elicited the highest speech understanding 
scores, performing significantly better than all other settings. For SRTs in co-located 4-
talker babble, no program worked significantly better than any other; this condition 
appeared to be the most difficult of the noise conditions for CI users. For SRTs in non-
16 
 
 
 
co-located speech-weighted noise and 4-talker babble, the best performance was found 
with settings that allowed microphone directionality changes (i.e., Nucleus 6/SCAN, 
Nucleus 6/Beam, and Nucleus 6/Zoom). Thus, microphone directionality appears to be 
key for improving speech understanding in these types of noisy environments. As a 
reminder, SCAN implements microphone-setting changes based on scene classification. 
Finally, the researchers reviewed the logged scene classifications provided by 
SCAN for the speech-in-noise conditions where noise was co-located with the target. After 
presenting four sentences with co-located speech-weighted noise, SCAN classified the 
scene as Quiet for 1/3 of the participants, Noise for 1/3 of participants, and Speech in Noise 
for 1/4 of participants. By the 20th sentence, almost all participants? SCAN had classified 
the scene as Noise, despite the fact that the scene was not simply noise but contained 
speech, too. After presenting four sentences of co-located 4-talker babble, SCAN classified 
the scene as Speech in Noise for the majority of participants (87%), and Quiet for all other 
participants; by the 20th sentence, almost all participants? SCAN had classified the scene 
as Speech in Noise. These logs show that SCAN is slow to classify scenes across 
participants and is not always a reliable classifier. 
To summarize, several front-end preprocessing strategies are introduced by the 
Nucleus 6, and key strategies that will likely be important to consider with respect to the 
speech repair paradigm (which involves listening to speech containing either noise-burst 
or silent-gap interruptions) include ADRO, ASC, SNR-NR, SCAN, and perhaps Whisper. 
Note that the default settings of the Nucleus 6 sound processor, which is one of the most 
recent CI sound processors available from Cochlear, make all of the above front-end 
preprocessing strategies available for adult users (Gilden et al., 2015; Wolfe et al., 2015). 
17 
 
 
 
However, poor classification of sound scenes could reduce the chance that the best set of 
strategies are being utilized by the device at any one time. 
 
1.2 Perceptual restoration 
In everyday listening environments, important speech information may be 
interrupted by noise or competing talkers. A speech repair strategy called perceptual 
restoration can help a listener improve understanding (Bashford, Riener, & Warren, 1992; 
Ba?kent, 2012; Verschuure & Brocaar, 1983). The perceptual restoration effect can be 
quantified as the boost in speech understanding a listener achieves when listening to noise-
burst interrupted speech compared to silent-gap interrupted speech. If the sudden disruption 
and loss of speech information were all that affected speech understanding in realistic 
environments, we would expect that performance with noise-burst interrupted and silent-
gap interrupted speech would be similar. Since performance with noise-burst interrupted 
speech is actually better in most cases, additional factors must be at work during the 
processing of noisy speech. 
One reason that noise-burst interrupted speech can be easier to understand than 
silent-gap interrupted speech is that the presence of noise in a signal interruption promotes 
an illusion of an intact speech stream. The noise acts as a plausible masker of the speech, 
and the speech seems natural and as if it is continuing through the noisy interruption. This 
?naturalness? and illusory ?intactness? improves speech understanding despite the fact that 
the noise bursts do not actually add any phonetic information to the sentence. The illusion 
of intactness is so strong that it can be observed in the auditory cortex ? the brain has been 
shown to represent intact versions of a word (presented in quiet) similarly to noise-burst 
18 
 
 
 
interrupted versions of a word (Leonard, Baud, Sjerps, & Chang, 2016). Silent-gap 
interrupted speech is comparatively more difficult to understand than noise-burst 
interrupted speech, even if the same amount of speech is available between interruptions. 
Silent-gap interrupted speech seems ?hoarse and raucous? (Miller & Licklider, 1950), and 
similar to speech heard through a cellular phone with poor reception. The listener does not 
perceive an illusorily intact speech signal continuing ?through? the interrupting silences.  
In general, it is thought that perceptual restoration involves top-down mechanisms 
interacting with bottom-up acoustic information (Ba?kent, 2012; Ba?kent et al., 2016; 
Shinn-Cunningham & Wang, 2008). The top-down factors that can interact with this 
bottom-up information include: context usage, or applying one?s knowledge about the 
context in which the speech is occurring; expectations about the speaker and topic of 
conversation; and linguistic knowledge like vocabulary and grammatical constraints 
(Ba?kent et al., 2016; Ishida & Arai, 2016; Patro & Mendel, 2020; Samuel, 1987; Shinn-
Cunningham & Wang, 2008). Furthermore, the presence of noise might provide enough 
ambiguity to stimulate potential word candidates (Ba?kent et al., 2016; Bhargava, 
Gaudrain, & Ba?kent, 2014), thus leading to a higher chance of successfully identifying 
the interrupted word. Finally, the presence of noise may also be (1) masking misleading 
cues generated by the silent-gap condition (e.g., a silent gap being misperceived as a 
voiceless consonantal stop), (2) masking distortions of the temporal speech envelope 
caused by the abrupt changes in the signal when speech is turned off prior to an interruption 
(Bologna, Vaden, Ahlstrom, & Dubno, 2018), and/or (3) maintaining the temporal 
patterning of speech, leading to easier processing (Powers & Wilcox, 1977).  
19 
 
 
 
Adult NH listeners can typically obtain excellent speech understanding in the 
presence of noise by utilizing restoration (Ba?kent, 2012; Newman, 2004; Samuel, 1981; 
Warren, 1970). I will first focus on perceptual restoration in these types of listeners. In later 
sections, I will discuss the extent to which CI users utilize perceptual restoration.  
 
1.2.1 Early perceptual restoration studies 
In this section, I will describe four early studies in the area of perceptual restoration. 
In the subsequent section, I will describe more recent work on perceptual restoration. 
The term ?perceptual restoration? was introduced by Warren (1970), who described 
a study in which young adults restored a missing speech sound in a sentence context when 
it was replaced by a noisy interruption (a cough) but not when it was replaced by a silent 
gap. The presented sentence was ?The state governors met with their respective legislatures 
convening in the capital city? and the missing speech sound was the first /s/ in 
?legislatures.? When a cough replaced the /s/, the majority of participants reported that they 
perceived no missing speech sounds ? that is, the majority of participants restored the 
missing /s/ and heard the cough as an extraneous sound. In contrast, when the /s/ was simply 
deleted from the sentence, creating a silent gap, all participants were able to detect which 
speech sound was missing and did not perceive a restored word. In a separate paper, Warren 
and Sherman (1974) concluded that restoration was a normal, necessary process: without 
restoration, understanding speech in any realistic, noisy listening environment might be 
impossible.  
Miller and Licklider (1950) reported a study about interrupted speech a few decades 
before Warren (1970), but did not refer to the phenomenon as perceptual restoration. They 
20 
 
 
 
were interested in how intelligibility was affected by deletions of speech across time. If 
speech could be deleted at certain intervals without affecting intelligibility, this method 
could potentially conserve resources in a speech-transmission system. To test this, they 
applied a periodic rectangular wave to a speech signal, which had the effect of creating 
periodic silent gaps, and varied factors that are still important variables in restoration 
paradigms involving multiple interruptions today. These factors were the frequency of 
interruption (usually denoted in Hz), the proportion of speech available within each 
interruption segment (usually referred to as duty cycle), and the temporal regularity of 
interruptions.  
Miller and Licklider (1950) initially tested five NH young adults on their speech 
understanding of an auditorily presented list of 50 monosyllabic words. Across conditions, 
the frequency of interruptions was varied, but the duty cycle remained at 50%. This means 
that if the frequency of interruption was 10 Hz, the first 50-ms of each 100-ms segment 
contained speech information, and the second 50-ms of each 100-ms segment was 
interrupted (here, replaced with a silent gap). Participants struggled with interruption 
frequencies of 1 Hz, after which performance improved as frequencies increased to 100 
Hz. Miller and Licklider (1950) concluded that participants struggled with interruption 
frequencies of 1 Hz because at this rate, large portions of the individual words were 
missing, for example the entire initial or final consonant. Higher performance at faster 
frequencies was believed to be due to the fact that participants were provided with several 
?looks? at each word or even each phoneme. Thus, even though information was deleted, 
enough ?looks? or ?glimpses? at the signal were sufficient for participants to interpret the 
overall phoneme or word.  
21 
 
 
 
Another condition in that study varied duty cycle, or the proportion of speech 
available compared to the proportion of silent-gap interruption within each segment. 
Having greater proportions of the speech signal available, regardless of interruption 
frequency, greatly improved speech understanding scores; for example, a duty cycle of 
80% resulted in speech understanding of over 80% correct for all tested frequency 
conditions containing regular interruptions.  
While the previous conditions from Miller and Licklider (1950) evaluated word 
recognition when the word lists were interrupted with silent gaps, they also tested how well 
participants perceived speech that alternated with noise (i.e., noise-burst interrupted 
speech). At slower than 5-Hz interruption rates, speech understanding was similar in both 
noise-burst interrupted and silent-gap interrupted conditions; at faster interruption rates, 
speech understanding with noise-burst interruptions declined. This is the opposite of the 
expected restoration effect. Miller and Licklider (1950) did not provide information about 
how speech and noise were alternated, whether any amplitude envelope ramping between 
speech and noise was used, or whether there was large variability among their five 
participants, which could potentially help us interpret their atypical result.  
Miller and Licklider (1950) reported that although speech understanding with 
noise-burst interruptions became poorer than speech understanding with silent-gap 
interruptions at frequencies above 5 Hz, participants described the noise-burst interrupted 
speech as ?more natural and probably more intelligible.? This sentiment was repeated 
among participants in another restoration study by Verschuure and Brocaar (1983), who 
reported noise-burst interrupted stimuli to be ?easier to follow? and ?continuous.? Miller 
and Licklider (1950) described this sentiment as the ?picket fence? effect. When noise 
22 
 
 
 
alternates with speech, the speech begins to be perceived as continuous, like scenery 
observed through the slats of a picket fence. Though the scenery is obviously truly 
continuous behind the fence, the observer can only see slices of scenery between slats, and 
must put the image of the scene back together in her mind?s eye. The noise interruptions 
thus served as slats, with enough speech information between that could be used to restore 
the signal, despite the fact that this notion was not supported by the experimental data 
(Miller & Licklider, 1950). It is possible that by presenting monosyllabic words, which 
have no context clues and do not exist in a sentential framework, few top-down 
mechanisms (context, expectations, etc.) were able to interact with the bottom-up acoustic 
information to produce a typical restoration effect.  
Samuel (1981) investigated perceptual restoration using a paradigm informed by 
signal detection theory. Previous work in the field had focused mostly on presenting speech 
with silent-gap or noise-burst interruptions, and then measuring either percent words 
reported correctly or self-report as to whether the interrupted phoneme was present 
(Warren, 1970; Warren & Obusek, 1971; Warren & Sherman, 1974). In contrast to these 
approaches, Samuel (1981) believed that measuring restoration with a signal detection task, 
and applying signal detection theory to analyze the results, provided a more ?direct? 
measure of the phenomenon.  
The signal detection theory framework allows the researcher to separate listener 
bias from the results using calculations of hit, false alarm, correct rejection, and miss rates 
(Samuel, 1981). Presenting two types of noise-burst interrupted stimuli ? one type where 
the noise is added to the speech, and one type where the noise replaces the speech ? and 
asking participants to report whether the presented speech signal was intact, allows the 
23 
 
 
 
researcher to measure the strength of the restoration illusion (Samuel, 1981). If a listener 
fails to detect the deletion of the speech signal, as occurs in the noise-replaced condition, 
then it means that the listener is experiencing the restoration illusion.  
Samuel (1981) predicted that if a listener is provided with context, they will have 
strong expectations about the incoming speech signal, and the listener will need very little 
bottom-up acoustic information to confirm these expectations4. In the first experiment, 
participants were presented with high-frequency and low-frequency words. A high-
frequency word is a word to which listeners are frequently exposed, and thus listeners may 
have more practice retrieving its representation from their lexicon. A high-frequency word 
was predicted to be highly restorable, since the brain has more expectations for the word 
to appear than it does for a low-frequency word. Another variable in this first experiment 
was the number of syllables in the presented words. Longer-syllable words were predicted 
to be highly restorable, as more context is provided (and more constraints are made) by 
words with more syllables. Another variable was phoneme class: it was hypothesized that 
phonemes that were most acoustically similar to the white noise burst interruption (i.e., 
fricatives and stops) would be more likely to be restored, due to bottom-up confirmations. 
Finally, the last variable was the location of the interruption in the word: at the beginning, 
middle, or end. It was hypothesized that the most restoration would occur for an interrupted 
phoneme at the end of a word, as the brain would be working with more context prior to 
the interruption.  
                                                 
4 Samuel (1981) used the terms ?expectation? and ?confirmation? to refer to top-down mechanisms and 
bottom-up acoustic information, respectively.  
24 
 
 
 
Participants were asked to identify whether the presented word had a phoneme 
replaced by noise or mixed with the noise. D-prime values were calculated for the various 
conditions, a d-prime value being an index of how discriminable the added versus replaced 
versions of the words were. High d-prime values meant the participant was able to 
discriminate between the added and replaced versions, and thus no restoration occurred. 
Low d-prime values meant discriminability was poor, and the participant likely 
experienced restoration. 
The results were as follows. The restoration effect (i.e., low d-prime values) was 
strongest for high-frequency words and for longer, four-syllable words, which revealed an 
effect of context and expectation on restoration. The restoration effect was also strongest 
for fricatives and stops, revealing an effect of bottom-up acoustic information on 
restoration. The predicted effect of context was not observed for the variable of interruption 
location: restored phonemes were most likely to occur in any position of a word, be it the 
beginning, middle, or end.   
The second experiment by Samuel (1981) compared restoration of non-words and 
real words, to measure how lexical knowledge affected restoration. The intact (non-
interrupted) version of the word or non-word was presented first, and then either a noise-
added or noise-replaced version of the same stimulus. Participants were asked to compare 
the two stimuli and decide which type of interruption had been used in the second stimulus 
(noise-added or noise-replaced). Participants showed high discriminability for non-words 
compared to real words and much lower discriminability for real words. Lexical knowledge 
thus appears to be a top-down mechanism that can affect restoration. In summary, Samuel 
25 
 
 
 
(1981) helped demonstrate through a signal detection task that top-down and bottom-up 
streams of information interact during restoration.  
Other early work in restoration includes a study by Bashford et al. (1992). Their 
study was composed of two experiments, the first investigating the bottom-up acoustic 
spectral information requirements for restoration, and the second investigating how 
contextual information affects restoration. For the first experiment, sentences were 
multiply interrupted (i.e., the same sentence contained several interruptions) with either 
silent gaps or bursts of five different types of pink noise, each with a different noise-band 
center frequency (375, 750, 1500, 3000, or 6000 Hz). The narrow-band noise-bands were 
each 1/3-octave wide. The sentences were filtered so that the maximum speech information 
was available at 1500 Hz, with the slopes of the low-pass and high-pass filter on either side 
of 1500 Hz dropping 48 dB per octave. The interruptions were presented at ~2.86 Hz with 
a 50% duty cycle, and the signal-to-noise ratio (SNR) of speech to noise was ?10 dB. The 
researchers found that speech understanding was maximized with the pink noise-burst 
interruptions centered on 1500 Hz, and the boost in speech understanding from the silent-
gap condition to the 1500-Hz-centered noise condition was 9%. This result, according to 
the researchers, demonstrated the importance of plausible maskers for restoration: a 
narrowband speech signal (centered on 1500 Hz) required a narrowband noise burst 
(centered on the same frequency) to maximize the restoration illusion. For the second 
experiment, the researchers varied the context available in the interrupted speech. At the 
?low-context? end of the continuum, they presented monosyllabic word lists. At the ?high-
context? end of the continuum, they presented highly predictive sentences. They also 
presented low-predictive sentences, which provide a word in a neutral carrier sentence. The 
26 
 
 
 
most accurate speech understanding and the greatest restoration effects were found with 
the highly predictive sentences, followed by the low-predictive sentences. Participants 
experienced no restoration with monosyllabic words. The researchers concluded that 
context is important for restoration to work, even if that context does not provide a plethora 
of semantic information (as occurs in the low-predictive sentences). Syntactic constraints, 
coarticulation, intonation, and stress could all be useful for speech repair, even if semantic 
clues are unavailable (Bashford et al., 1992). 
In summary, these early studies discovered the phenomena of perceptual restoration 
and began to probe how and why it occurs through various methodologies like signal 
detection theory, single word interruptions, and multiply interrupted sentences. More 
recent work in this area will be discussed in the next section. 
 
1.2.2 More recent perceptual restoration studies 
Additional novel methodologies for studying perceptual restoration have been used 
in recent work in this area. For example, some researchers have investigated the role of the 
auditory brainstem in restoration, with some advocating a strong role for bottom-up 
acoustic information in the process, and less for top-down mechanisms (Bidelman & Patro, 
2016). Other researchers have advocated for a deeper analysis of speech understanding 
scores in restoration paradigms. For example, Smith and Fogerty (2017) examined the 
types of errors participants made during a noise-burst interrupted speech understanding 
task, rather than a more holistic measure of ?percent words correct.? They found that at 
low duty cycles, where less of the speech information is available between interruptions, 
participants often omitted whole words from their answer (i.e., provided no response) or 
27 
 
 
 
substituted incorrect words. At high duty cycles, where more of the speech information is 
available between interruptions, participants most often substituted an incorrect phoneme. 
While Smith and Fogerty (2017) did not analyze error types in a silent-gap condition, they 
posit that their methodology could be useful for restoration research, as this could help 
determine individual differences in restoration benefit. 
Benard, Mensink, and Ba?kent (2014) noted large individual variability in 
perceptual restoration in NH listeners, even among younger adult listeners, and asked 
whether linguistic and cognitive skills could account for this variability. Linguistic and 
cognitive skills are posited as top-down mechanisms that interact with bottom-up acoustic 
information. Benard et al. (2014) found that NH participants? performance in several of the 
interrupted (both silent-gap and noise-burst) speech conditions were positively correlated 
with vocabulary scores (PPVT III), but not with cognition scores (WAIS IV). While Benard 
et al. (2014) correlated linguistic and cognitive scores with interrupted speech performance, 
they did not directly correlate these scores with restoration benefits ? thus, while it 
appeared that vocabulary was an important factor for interrupted speech understanding, its 
exact role in restoration remained unclear. 
Jaekel, Newman, and Goupell (2018) also included linguistic and cognitive 
measures in a restoration experiment, in the hopes that these measures could explain 
individual differences in restoration benefit. Linguistic skill, quantified as participants? 
average phonemic and semantic lexical access ability, or how quickly they could generate 
words with a specific initial speech sound or belonging to a specific semantic category 
(e.g., ?animals?), respectively, was positively correlated with restoration benefits, but only 
when speech signals were degraded to simulate CI processing (Jaekel et al., 2018). 
28 
 
 
 
Cognitive skill, measured with a general working memory task, was not associated with 
restoration benefit in any speech condition. In contrast, Nagaraj and Magimairaj (2017) 
found that both receptive vocabulary and verbal working memory were predictive of higher 
restoration of low-context sentences. There are thus conflicting results about how top-down 
mechanisms like linguistic knowledge and cognitive abilities moderate restoration and 
whether they can successfully explain individual variability in restoration benefits. 
Finally, another line of recent work has investigated how restoration works in other 
listener groups besides young NH adults. Interest in how people with hearing loss restore 
speech is the focus of the following section. The impacts of aging on restoration have been 
studied also (Bologna et al., 2018; Jaekel et al., 2018; Saija, Aky?rek, Andringa, & 
Ba?kent, 2014). A study by Saija et al. (2014) showed that older NH listeners experience 
greater restoration benefits than younger NH listeners, perhaps revealing the important role 
of top-down mechanisms for the restoration effect. Older NH listeners tend to have larger 
vocabularies and show better use of context and expectations due to a lifetime of practice 
with dealing with noisy backgrounds and compensating for missing speech information 
(Pichora-Fuller, 2008; Sheldon, Pichora-Fuller, & Schneider, 2008), all of which are 
believed to be important top-down mechanisms. In contrast, neither Bologna et al. (2018) 
nor Jaekel et al. (2018) found evidence that greater restoration of unprocessed, interrupted 
speech in older NH adults was correlated with language or cognitive abilities. 
 
1.2.3 Perceptual restoration with degraded input 
Perceptual restoration has been shown to be a useful tool for NH adult listeners 
hearing speech in noisy environments. The extent to which perceptual restoration is 
29 
 
 
 
accessible to adults without normal hearing, particularly those who use hearing devices 
like CIs, is the focus of this section. Much of the research in this area has been conducted 
with NH listeners presented simulations of CI processing rather than with CI users 
themselves. As discussed previously, difficulties understanding speech in noise is one of 
the chief concerns reported by CI users; thus, ensuring that perceptual restoration is 
accessible to this group is important.  
 
1.2.3.1 Perceptual restoration in CI users 
CI users struggle to comprehend silent-gap interrupted speech (Chatterjee, Peredo, 
Nelson, & Ba?kent, 2010; Gnansia, Pressnitzer, Pean, Meyer, & Lorenzi, 2010). They 
experience less ?release from masking,? or boosts in speech understanding when noise 
interferers are intermittent rather than constant, than NH listeners presented normal or CI-
simulated speech (Jin et al., 2013; Nelson & Jin, 2004; Nelson et al., 2003; Perry & Kwon, 
2015). While such ?release from masking? paradigms are not perfect correlates of the 
noise-burst interrupted sentences in a speech repair task (the former overlays speech with 
noise, while the latter replaces speech with noise), results from these studies support the 
idea that CI users likely struggle with speech repair. 
Bhargava et al. (2014) measured perceptual restoration in CI users directly. They 
predicted that CI users would struggle to restore speech because they would have less 
access to high-quality bottom-up acoustic information during speech repair. First, 
limitations in CI processing would lead to lower-quality, more degraded signals, since CIs 
encode only limited spectral information and no temporal fine structure. In the same vein, 
front-end preprocessing in CIs such as compression can distort the shape of temporal 
30 
 
 
 
envelopes (Ba?kent et al., 2009). Second, limitations in peripheral auditory encoding could 
lead to less well-represented sound information, since some CI electrodes may have poor 
interface with surviving auditory neurons. Poor interface can be due to the way the 
electrode array was inserted into the cochlea, and because some areas of the cochlea may 
have no surviving auditory neurons. Bhargava et al. (2014) predicted that it may be more 
difficult to integrate top-down knowledge with low-quality bottom-up acoustic 
information, and thus restoration in CI users may be more likely to fail.  
For the experiment, Bhargava et al. (2014) tested 13 Dutch CI users, aged 22 to 65 
years. All CI users in the study wore unilateral CIs and had at least 1 year of experience 
with their devices. The majority of CI users in the study wore Cochlear-brand implants, 
and all participants had to be high performing with regards to speech understanding (i.e., 
have at least 70% monosyllabic word recognition in quiet). The researchers also tested 14 
Dutch NH listeners, aged 19 to 28 years. The CI users in the study were therefore older on 
average than the NH listeners. Older age is associated with more restoration in NH adults 
(Saija et al., 2014) and NH adults presented vocoded speech (Jaekel et al., 2018). 
Experimental stimuli were everyday Dutch sentences, which were presented at 60 dB-A. 
Sentences were multiply interrupted with silent gaps or noise bursts at a frequency of 1.5 
Hz. Duty cycles were either 50% or 75%, which removed one-half or one-fourth of the 
speech information in each 666-ms segment of the sentence, respectively. SNRs were 
either ?10-, ?5-, 0-, or +5-dB SNR. As a reminder, a negative SNR means that speech is 
less intense than noise. CI users listened to speech with their regular default sound 
processor settings. NH listeners were presented stimuli in two conditions: either 
31 
 
 
 
unprocessed (normal) or 8-channel noise-vocoded. All stimuli were presented in the free 
field through one loudspeaker directly in front of the listener.  
 At the 50% duty cycle, CI users showed no significant restoration effect (i.e., no 
change in speech understanding between silent-gap and noise-burst interruption 
conditions). In contrast, NH listeners presented normal speech achieved a significant 
restoration benefit of +8.6 RAUs. RAUs stand for ?rationalized arcsine units? and are 
similar to percentages but have properties making them meet certain criteria for inferential 
statistics (Studebaker, 1985)5. For the same NH listeners presented vocoded speech, no 
significant restoration benefits were observed. Thus, when only half of the speech 
information was available in each segment, neither CI users nor NH listeners presented a 
CI simulation could restore speech. This supported the authors? prediction that poorer 
bottom-up information impairs restoration, particularly when only short durations of 
speech information are available.  
 A different story emerged when more speech information was available between 
interruptions. At the 75% duty cycle, CI users achieved a significant restoration benefit of 
+5.6 RAUs. In contrast to the 50% duty cycle (where a restoration benefit was observed), 
NH listeners presented normal speech showed no consistent restoration effect with the 75% 
duty cycle, likely due to ceiling performance; for the same NH listeners presented a CI 
simulation, no consistent restoration effect was observed, though performance was not at 
                                                 
5 Arcsine transforms were developed to deal with biases in statistics analyzing percent correct. When scores 
are constricted by a range, like percent correct (0-100), the possible range of variabilities around mean scores 
are dependent on the mean. A mean score of 50% affords a large range of score variability around the mean, 
while a mean score of 5% or 95% results in a compressed range of score variability around the mean. 
However, arcsine transforms on their own are difficult to interpret. Thus, the rationalized arcsine transform 
was developed to improve interpretation of transformed scores. RAUs can be interpreted similarly to 
percentages and extend from approximately ?20 to +120. See Studebaker (1985) for a more detailed 
explanation.  
32 
 
 
 
ceiling in this condition. Thus, with a greater amount of speech information available in 
each segment, CI users can achieve restoration, but NH listeners presented a CI simulation 
cannot.  
 Bhargava et al. (2014) next looked into the CI users? individual data. CI users were 
more likely to get a restoration benefit in the 50% duty-cycle condition if they had better 
baseline speech understanding scores (that is, better speech understanding with intact, non-
interrupted sentences) or longer durations of CI use. In the 75% duty-cycle condition, no 
demographic information nor baseline speech understanding scores significantly correlated 
with restoration benefit. The researchers concluded that the high-performing CI users 
obtaining a restoration benefit at the 50% duty cycle were better able to perceive and 
encode speech information and/or to utilize the speech information they had access to. The 
researchers also posited that perhaps CI users did better overall in the 75% duty cycle 
because a greater proportion of temporal envelopes were left intact, which could have led 
to more accurate lexical activation. Temporal envelopes are important for CI users as they 
are one of the few cues available for speech understanding, as temporal fine structure and 
spectral information are missing or degraded with CI processing.  
 In summary, Bhargava et al. (2014) showed that (1) when interruptions obscured 
greater portions of the speech signal, neither CI users nor NH listeners presented a CI 
simulation could obtain restoration similar to NH listeners presented normal, unprocessed 
speech, on average. However, some CI users with longer use of their CIs and better overall 
speech understanding could obtain restoration in this difficult speech condition, while the 
rest of the CI users showed no restoration or a negative restoration effect ? that is, better 
speech understanding with silent gaps than with noise bursts. Bhargava et al. (2014) also 
33 
 
 
 
showed that (2) when interruptions obscured a smaller portion of the speech signal, CI 
users could restore speech on average, while NH listeners presented a CI simulation could 
not. Overall, CI users failed to show typical restoration in scenarios where restoration was 
possible for NH listeners presented non-vocoded speech. Possible reasons for this reduced 
restoration in CI users include the following. (1) Noise bursts may initiate compression 
algorithms, leading to distortions in the speech envelopes prior to and/or after the noise 
burst (Ba?kent et al., 2009). Since noise was more frequent in the 50% duty cycle condition, 
perhaps compression algorithm effects were more likely in that condition. (2) While results 
from NH listeners with CI simulations could lead one to conclude that CI processing is to 
blame for a lack of restoration, as it degrades the bottom-up acoustic information, some CI 
users could restore speech at the 50% duty cycle. Perhaps these CI users had better 
peripheral auditory encoding, which is necessary for restoration in the most difficult 
listening conditions (Bhargava et al., 2014). 
 
1.2.3.2 Perceptual restoration in CI simulations (vocoding studies) 
Investigating how CI processing may affect restoration has been a focus of research 
for the Ba?kent lab group. To pursue this line of work, they have often utilized vocoders to 
simulate aspects of CI processing. Above, findings from Bhargava et al. (2014) were 
discussed, which included results from a young NH group of listeners presented an 8-
channel noise-vocoder CI simulation. In that study, no consistent restoration was observed 
across the eight duty cycle ? SNR conditions with vocoded speech ? the only significant 
restoration effect was seen with the 75% duty cycle at 0 dB SNR. One possible explanation 
for the results from Bhargava et al. (2014) is that NH listeners did not have adequate 
34 
 
 
 
experience or training with vocoded speech, which made them less effective at utilizing 
top-down strategies to repair it. To test whether training with vocoded speech improves 
restoration, Benard and Ba?kent (2014) trained young NH listeners to complete a 
restoration task with interrupted 8-channel noise-vocoded sentences over the course of five 
testing sessions in the same one-week period (the exact duration of training in this 
experiment was not reported by the authors). While both silent-gap and noise-burst 
interrupted speech understanding improved with training, the size of the restoration benefit 
did not change. Thus, greater exposure to vocoded speech does not seem to induce greater 
restoration in NH listeners.  
 Ba?kent (2012) tested several spectral resolutions and also simulated electric-
acoustic hearing to determine if a certain level of bottom-up acoustic information quality 
was necessary for restoration of vocoded, interrupted speech. Electric-acoustic hearing 
refers to a situation where a CI user experiences preserved residual, low-frequency hearing 
by being implanted with a short electrode array. Participants were young Dutch NH 
listeners (average age = 22 years). Stimuli were everyday Dutch sentences interrupted with 
silent gaps or noise bursts at a frequency of 1.5 Hz with a 50% duty cycle. Interrupted 
sentences were then noise-vocoded with either 4, 8, 16, or 32 channels, with the 4 channel 
stimuli having very poor spectral resolution, and the 32 channel stimuli having 
comparatively high spectral resolution. Note that 8 channels of spectral information has 
been posited as the maximum amount received at any one time by real CI users (Friesen et 
al., 2001)6. For the electric-acoustic hearing conditions, the lowest one-fourth of the 
                                                 
6 Although see more recent evidence supporting the use of more than 8 channels in CI users with newer 
devices (Berg et al., 2019; Croghan et al., 2017). 
35 
 
 
 
channels in each spectral resolution condition were replaced with low-pass-filtered 
acoustic speech. Participants also completed an unprocessed interrupted speech condition, 
in which the interrupted speech signals were not vocoded. 
 With unprocessed interrupted speech, participants showed a 9% restoration benefit. 
When speech was degraded spectrally, the restoration benefits decreased, and in the most 
degraded conditions, disappeared altogether. For the regular CI simulation, significant 
restoration benefit was seen at the highest spectral resolution only, at 32 channels. For the 
electric-acoustic CI simulation, significant restoration benefits were seen at the two highest 
spectral resolutions, 16 and 32 channels. Ba?kent (2012) concluded that the quality of 
bottom-up acoustic information is an important aspect of restoration with CI-simulated 
speech: without quality signals, top-down mechanisms are less able to interact with the 
bottom-up information. Furthermore, the availability of voicing and pitch cues in the 
electric-acoustic CI condition could have contributed to better restoration.   
 To determine the extent to which pitch is important for restoration of degraded 
speech, Clarke, Ba?kent, and Gaudrain (2016) used a vocoder that would allow the 
researchers to vary the strength of a pitch percept in the vocoded speech signals. CI 
processing provides CI users with temporal pitch cues only, whereas NH listeners have 
access to temporal and spectral information about pitch (Clarke et al., 2016). This lack of 
clear pitch information in conjunction with poor spectral resolution in CI users could mean 
that the speech information between interruptions is less intelligible, and that without pitch, 
it could be more difficult to link speech information across interruptions7. In their 
                                                 
7 Pitch may only be useful at linking degraded speech signals across noise bursts, not across silent gaps. 
Ardoint, Green, and Rosen (2014) found no difference in silent-gap interrupted speech understanding with 
8-channel noise-vocoded speech (containing no voicing information) compared to 8-channel noise-vocoded 
speech containing fundamental frequency contour information. 
36 
 
 
 
experiment, they varied spectral resolution and pitch availability. They tested 19 young 
Dutch NH listeners (average age = 23 years) with everyday Dutch sentences. Sentences 
were vocoded using a non-traditional vocoder, which allowed the researchers to tightly 
control temporal pitch cues. Speech understanding in five spectral resolutions (unprocessed 
speech, or 4, 6, 8, or 16 channels of noise-vocoded speech) and two pitch conditions (the 
signal?s fundamental frequency [F0] information was retained or removed) was tested. 
Interruptions were added to the signals after vocoding, rather than before, as was done in 
previous studies (Ba?kent, 2012; Bhargava et al., 2014), meaning speech information was 
degraded spectrally, but noise bursts were not. Interruptions had a frequency of 2.2 Hz and 
a 50% duty cycle, and were either silent gaps or bursts of white noise. Clarke et al. (2016) 
found significant restoration benefits at all spectral resolutions but the poorest resolution 
(4 channels) when pitch information was available. Significant restoration benefits were 
only observed in the unprocessed and 16-channel condition when pitch information was 
not available. Thus, access to pitch information seemed to be a useful cue when speech was 
degraded, leading to better speech repair. The pitch information in this experiment included 
information about voicing, which could help with phoneme identification, as well as pitch 
contours, which can provide information about word stress and prosody (Clarke et al., 
2016). Overall, the authors concluded that with enough complementary (though degraded) 
speech cues, and especially with cues about pitch, restoration should be possible in noisy 
speech situations processed by a CI.  
 Note that when Clarke et al. (2016) added interruptions after signals were vocoded, 
which created non-noise-vocoded noise bursts, participants were able to achieve 
restoration (without pitch cues) at 16 channels. In comparison, Ba?kent (2012), who added 
37 
 
 
 
interruptions before signals were vocoded, which created noise-vocoded noise bursts, 
found that participants could only achieve restoration at 32 channels. Jaekel et al. (2018) 
asked if spectral differences between noise interruptions and speech were important for 
restoration to occur with degraded signals. Perhaps the non-noise-vocoded noise bursts in 
the study by Clarke et al. (2016) led to better restoration at lower spectral resolutions 
because there was more of a perceptible difference, and thus better segregation, of noise 
bursts from the noise-vocoded speech. Perceptual restoration involves an important trade-
off when it comes to the noise-burst interruption conditions: first, noise-burst interruptions 
that are similar to the missing speech information can act as powerful, plausible maskers 
(Clarke et al., 2016; Samuel, 1981; Warren & Obusek, 1971); second, noise interruptions 
and speech need to be perceptually separable, in that the brain needs to be able to detect 
which portions of the incoming signal are speech segments (Clarke et al., 2016). This latter 
point may be violated when speech and noise interruptions are too perceptually similar (as 
they would be with noise-vocoded speech and noise-vocoded noise bursts), reducing 
restoration. 
 Jaekel et al. (2018) found that young NH listeners presented CI simulations 
benefitted from greater spectral differences between speech and noise bursts, while older 
NH listeners (i.e., aged 60 and older) did not. In that study, young and older NH listeners 
were presented 16- or 32-channel noise-vocoded sentences, which were interrupted at a 
2.5-Hz rate with a 50% duty-cycle with silent gaps, noise-vocoded noise bursts (vocoded 
with the same number of channels as the speech), or non-noise-vocoded noise bursts 
(unprocessed white noise). Sentences were in English and were either from the BKB corpus 
(Bench, Kowal, & Bamford, 1979) or the IEEE corpus (Rothauser et al., 1969). With the 
38 
 
 
 
BKB corpus, which contains shorter sentences appropriate for children, young NH listeners 
could only obtain a restoration benefit with non-noise-vocoded noise bursts. Older NH 
listeners, in contrast, could obtain restoration benefits with both types of noise.  
In the Jaekel et al. (2018) study, older NH listeners obtained significantly greater 
restoration than young NH listeners, demonstrating that older NH listeners not only 
perform greater speech repair with unprocessed speech (Saija et al., 2014) but also with 
degraded speech. While this ?aging benefit? for restoration, especially in degraded 
conditions, seems like a hopeful sign for CI users, many of whom are older and lost their 
hearing later in life, the Bhargava et al. (2014) study revealed that the older CI users (aged 
52 to 65 years) showed negligible restoration. Thus, more work in this area is needed to 
determine how aging and degraded speech interact during speech repair.  
 
1.2.3.3 Perceptual restoration in other populations with hearing impairment 
 Ba?kent, Eiler, and Edwards (2010) were interested in studying perceptual 
restoration in listeners with mild-to-moderate hearing loss, who also report difficulty 
understanding speech in noise. Due to damaged inner hair cells, these listeners may 
experience higher thresholds, increased masking, and reduced spectral and temporal 
resolution; all of these factors could result in poor-quality bottom-up acoustic information, 
which could impact the extent to which these listeners can repair noisy speech (Ba?kent et 
al., 2010). The researchers measured restoration in 9 NH listeners (aged 23 to 57 years), 9 
listeners with mild hearing loss (aged 47 to 83 years), and 9 listeners with moderate hearing 
loss (aged 64 to 81 years years). Participants in the NH group were required to have pure-
tone audiometric thresholds of <20 dB HL at 500, 1000, and 2000 Hz. Participants in the 
39 
 
 
 
mild hearing loss group were required to have thresholds between 26- and 40-dB HL, and 
participants in the moderate hearing loss group were required to have thresholds between 
41- and 55-dB HL, at those same frequencies. Stimuli were sentences interrupted with 
silence or noise bursts in several configurations of interruption rate, SNR, and duty cycle. 
To ensure that hearing-impaired listeners had access to audible stimuli, linear amplification 
was provided. Listeners with mild hearing loss showed significant restoration benefits in 
several of the configurations, particularly those with low SNRs (where noise was more 
intense than speech) and 50% duty cycles, which mirrored the results in NH listeners. 
Listeners with moderate hearing loss showed no significant restoration in any 
configuration. The amount of restoration benefit was not dependent on percent correct in 
the silent-gap interrupted conditions for either hearing-impaired group (Ba?kent, 2010), 
meaning the amount of speech information that could be gleaned by the listeners between 
interruptions was not driving the observed restoration effects. Thus, poor restoration 
outcomes with degraded speech are not specific to CI processing, and speech repair 
mechanisms appear to be affected by a range of signal quality degradations.  
 Finally, one study has analyzed how front-end preprocessing in devices for the 
hearing impaired (hearing aids and CIs) could potentially affect restoration. These devices 
use compression algorithms to keep signals within the user?s dynamic range, which both 
provides audible signals and protects the wearer from uncomfortably loud sounds. 
Dynamic ranges are narrower in listeners with hearing impairment than in NH listeners, 
thus necessitating some kind of compression that can translate a larger range of intensities 
to a smaller range. Ba?kent et al. (2009) investigated whether compression algorithms in 
these kinds of devices could negatively affect restoration. Compression algorithms cannot 
40 
 
 
 
turn on and off instantaneously in the presence of very quiet or very loud sounds; they thus 
might affect the temporal envelopes of speech sounds surrounding a noise burst. NH 
participants were presented speech interrupted with noise bursts, with envelope 
discontinuities of various durations added before and/or after noise bursts. These types of 
discontinuities would occur with a compression algorithm with various durations of attack 
times (onsets) and release times (offsets), as these factors vary in real compression 
algorithms in CIs and hearing aids. Ba?kent et al. (2009) found that very brief 
discontinuities (i.e., amplitude ramps of 10-ms) did not impact restoration, while longer 
discontinuities reduced restoration and affected the extent to which participants thought 
speech and noise was continuous. Thus, compression algorithms with long release times in 
response to sudden noise could be hindering speech repair in hearing-impaired listeners 
utilizing hearing devices, both in the restoration test paradigm and in realistic, noisy 
listening environments, and could help explain the atypical restoration observed in CI 
users.  
 
1.3 Summary and Hypotheses 
Cochlear-implant (CI) users struggle to understand speech in noisy, real-world 
listening environments, reducing their ability to participate in classrooms, restaurants, and 
playgrounds where noise is common. According to the data logs of 2.4 million listening 
hours in 1501 CI users of all ages, approximately four hours per day (of approximately 11 
total listening hours) were spent in noisy conditions (Busch, Vanpoucke, & van Wieringen, 
2017). These conditions can affect children?s ability to acquire language and impact adults? 
professional and personal lives (Busch et al., 2017).  
41 
 
 
 
Listening to speech in noise is difficult with a CI because its processing schemes 
can convey only degraded spectral (frequency) information and only some aspects of 
temporal (timing, intensity) information ? that is, the rich acoustic detail of speech is not 
available to CI users, making it more difficult to separate out important speech information 
from noise (Jin et al., 2013; O'Neill et al., 2018). CI manufacturers have attempted to 
combat the speech-in-noise problem by introducing noise reduction algorithms in the CI 
front-end preprocessing software. These algorithms were crafted to reduce the impact of 
noise on the incoming speech signals by evaluating soundscapes, changing microphone 
sensitivity, boosting quiet speech signals, and other measures. The use of these algorithms 
generally improves speech-in-noise understanding, but there is still room for improvement 
(M. T. Caldwell, Jiam, & Limb, 2017). Perhaps tweaks to these algorithms could provide 
CI users access to mechanisms that allow them to further repair noisy speech signals.    
NH listeners may improve their speech understanding in noise by relying on a 
speech repair mechanism called ?restoration.? Restoration works via an interaction of 
incoming bottom-up acoustic information with top-down linguistic and contextual 
knowledge. Noise-burst interrupted signals may be ?repaired? by applying one?s 
knowledge of the conversational context, vocabulary, grammatical constraints, and 
expectations. In the case of restoration, the interruption itself does not prompt speech 
repair; it is the interruption plus the presence of the noise that prompts repair. The presence 
of noise in the interruption serves as a plausible masker and creates the illusion of an intact 
speech signal continuing through the noise interruption. Thus, NH listeners show a speech 
understanding benefit with noise-burst interruptions, but not with silent-gap interruptions.  
42 
 
 
 
The previous research on restoration in adult CI users has revealed that they show 
atypical or negligible restoration in situations where NH listeners show restoration benefits. 
It is unknown if front-end preprocessing, which aims to remove noise from incoming 
signals, reduces opportunities for CI users to use speech repair. This project aims to 
investigate this question and measure whether changes to front-end preprocessing can 
support speech repair in CI users, and ultimately improve their ability to understand speech 
in noisy listening environments.  
 One hypothesis, then, is that noise-reduction algorithms in CI front-end 
preprocessing are reducing restoration in CI users (Experiment #1). This hypothesis is 
tempered by the fact that some bilateral CI users, meaning users with two CIs, have a 
?poorer? and ?better? performing ear. This is typically shorthand for how well the ear is 
encoding sound in the peripheral auditory system?that is, where the CI uses electrical 
pulses to activate surviving cells in the auditory nerve. Poor encoding could be caused by 
dead regions of auditory neurons in the cochlea and/or ineffective placement of the CI 
electrode array. A second hypothesis is that an ear with fewer dead regions and/or better 
electrode array placement may be better equipped to repair speech due to higher-quality 
bottom-up signals, and thus may be unaffected by the front-end preprocessing ? like 
compression ? that could be affecting the restoration illusion (Experiment #2). A third 
hypothesis is that an ear with poor encoding may not be able to repair speech at all, as 
bottom-up signals are too degraded for interaction with top-down linguistic knowledge, 
when that knowledge is made available (Experiment #3). 
To summarize, this project will help us evaluate how CI sound processing strategies 
affect speech repair. From this work, we could develop individualized front-end 
43 
 
 
 
preprocessing strategies, dependent on peripheral auditory health and the provision of 
semantic information, to increase opportunities for CI users to restore degraded, interrupted 
speech. Improving speech repair in people with CIs will support this population as they 
communicate and interact with others in realistic listening environments.  
44 
 
 
 
Chapter 2: Experiment 1: The impact of front-end preprocessing 
algorithms on restoration 
Understanding speech in noisy listening environments such as restaurants and 
classrooms is an important part of everyday communication. However, for people with 
severe to profound hearing loss who have received a CI, understanding speech in noise can 
be particularly difficult (Fetterman & Domico, 2002; Fu & Nogaki, 2005; Sladen & 
Zappler, 2015; Zhao et al., 1997). The CI is an auditory prosthesis that can restore a 
sensation of sound, and typically works well at conveying speech signals in quiet listening 
environments. However, the ability of the CI to encode speech is negatively impacted by 
the presence of noise (Fetterman & Domico, 2002; Sladen & Zappler, 2015). Improving 
CI effectiveness in noisy environments is therefore currently a critical area of need.  
One current approach in the field to combat noise interference and improve speech 
understanding in CI users is to rely on front-end preprocessing algorithms. These 
algorithms are part of the signal processing in CIs that convert acoustic information into 
electrical pulses. In so doing, these algorithms are designed to select the most important 
acoustic information to transmit to the user. Modern processing algorithms also seek to 
distinguish between sounds of interest and noise, and to reduce this noise prior to 
transmission (NIDCD fact sheet: Cochlear implants, 2016). Generally, these algorithms 
appear to improve CI users? speech understanding in certain noisy conditions (Davidson et 
al., 2010; Gifford & Revit, 2010; Gilden et al., 2015; Mauger et al., 2014; Wolfe et al., 
2015). However, this approach of using front-end preprocessing algorithms to reduce noise 
in speech signals may disallow use of a speech-repair strategy called ?restoration.? 
Determining whether this is the case is the main purpose of this study. 
45 
 
 
 
 Restoration can occur when a noise masks part of a speech signal, but the speech is 
still heard as illusorily intact and is properly interpreted (Warren, 1970). During restoration, 
listeners can use bottom-up acoustic cues ? the preceding and following speech information 
as well as the noise burst itself ? to perceive the speech as continuing intact through the 
noise burst. Listeners can also use one?s knowledge about the rules of language (semantic, 
syntactic, etc.), expectations, experience, and context to appropriately fill-in the lost speech 
information (Bashford et al., 1992; Ba?kent et al., 2010; Bregman, 1990). In summary, 
repairing noise-burst interrupted speech is possible through restoration, and involves 
access to bottom-up information and the use of top-down confirmation processes.  
NH listeners can obtain excellent speech understanding in the presence of 
intermittent noise by using this speech-repair strategy (Bashford et al., 1992; Verschuure 
& Brocaar, 1983; Warren, 1970). CI users, in contrast, often show atypical or negligible 
restoration ability in similarly noisy conditions (Bhargava et al., 2014). If we discover the 
underlying mechanisms that influence speech repair and cause its failure in CI users, we 
could improve speech understanding in noise in CI users and potentially develop 
improvements to CI speech processing. 
Previous research on restoration in CI users has focused on changing qualities of 
the interruption itself. For example, Bhargava et al. (2014) varied the SNR of the speech 
and interrupting noise, as well as the percentage of speech that was interrupted. The present 
study expands our knowledge of restoration mechanisms in CI users by measuring the 
effects of varying parameters at the level of the device. Device variables are, to some 
extent, more controllable than the presence of environmental noise (e.g., it is easy to change 
a program on a CI processor; it is hard to avoid many background noises). Therefore, 
46 
 
 
 
deeper understanding of restoration in CI users could lead to developing better speech-in-
noise processing strategies. 
We will test the hypothesis that the locus of restoration failure is at the level of CI 
front-end preprocessing. Specifically, we hypothesize that restoration in CI users will be 
worse when front-end preprocessing is turned on compared to when it is turned off; the 
reason being that these front-end preprocessing features were designed to remove or reduce 
the impact of noise during speech perception (which could cause noise-burst interrupted 
stimuli to be perceived more similarly to silent-gap interrupted stimuli). By removing or 
reducing noise, these algorithms also remove or reduce the salience of the cue that prompts 
restoration, and thus could paradoxically decrease speech intelligibility in a setting with 
intermittent noise. The rationale for this study is that we can determine if CI users' deficits 
in speech restoration are at least partly accounted for by software in the CI itself. It is our 
expectation that front-end preprocessing features in the CI will decrease restoration ability 
significantly, eliminating any ?useful? aspects of noise used in the speech repair process. 
New noise reduction strategies that allow for some of these ?useful? aspects of noise to 
become part of CI speech processing could then be conceived. 
 
2.1 Methods 
2.1.1 Participants 
 Eleven bilaterally implanted adult CI users participated in this study. Participant 
demographics are presented in Table 1. The mean age was 62.8 years (SD=14.1 years, 
range=32 to 79 years). The average age at onset of non-normal hearing in at least one ear 
was 26.9 years (SD=18.8 years, range=2 to 55 years), and the average duration of non-
47 
 
 
 
normal hearing in at least one ear prior to first implantation was 25.3 years (SD=19.8 years, 
range=0.5 to 59 years). Participants were native monolingual speakers of American 
English and self-reported normal or corrected-to-normal vision. All participants used a 
Cochlear-brand CI in both ears with N6 processors. All participants had had their most 
recent CI activated for at least 6 months prior to testing, which ensured adequate experience 
with the devices, especially bilaterally. All participants passed the Montreal Cognitive 
Assessment (MoCA) with a score of >22 (Nasreddine et al., 2005), indicating a lack of 
mild/moderate cognitive impairment.  
 
Table 1. Participant demographics for Experiment 1. 
Participant Variable Mean SD Range 
Age (years) 62.8 14.1 32 to 79 
Average age at onset of non-normal hearing (years) 26.9 18.8 2 to 55 
Average duration of non-normal hearing prior to first 
25.3 19.8 0.5 to 59 
implantation (years) 
Vocabulary (age-corrected standard score) 105.6 13.1 93 to 134 
Working memory (age-uncorrected standard score) 98.0 12.8 78 to 113 
Processing speed (age-uncorrected standard score) 100.2 18.1 63 to 122 
Attention (age-uncorrected standard score) 100.7 6.3 95 to 113 
Baseline speech scores in the ON condition (in %) 95.6 6.7 82 to 100 
Baseline speech scores in the OFF condition (in %) 91.4 8.0 76 to 99 
 
 Participants also completed a battery of cognitive tests available from the NIH 
Toolbox (Gershon et al., 2013), with scores presented in Table 1. The age-uncorrected 
standard scores were collected for these participants. For working memory (?List Sorting 
Working Memory Test Age 7+?), the mean score was 98.0 (SD=12.8, range=78 to 113). 
For processing speed (?Pattern Comparison Processing Speed Test Age 7+?), the mean 
score was 100.2 (SD=18.1, range=63 to 122). For attention and executive functioning 
48 
 
 
 
(?Flanker Inhibitory Control and Attention Test Age 12+?), the mean score was 100.7 
(SD=6.3, range=95 to 113). Finally, participants completed the PPVT-4 (Dunn & Dunn, 
2007), a vocabulary test, with a mean standard score of 105.6 (SD=13.1, range=93 to 134).  
 
2.1.2 Stimuli 
Stimuli were 320 declarative sentences (created by the author), which contained 5-
12 words with a range of speech sound and semantic content. The sentences were recorded 
by a female speaker with a Standard American English dialect. Twenty sentences were left 
?intact? (i.e., uninterrupted by silent gaps or noise) and presented in quiet as a baseline 
measure. Fifty percent of the remaining sentences (n=150) were interrupted with silent 
gaps by applying a 5-Hz periodic nominally square wave with an 80% duty cycle8 to the 
signal, with 1-ms raised cosine on/off ramps. The 80% duty cycle means that within each 
200-ms long speech segment, the first 160 ms was left intact and the following 40 ms was 
replaced with a silent gap. Meyer, Brand, and Kollmeier (2011) reported the average 
English phoneme durations to be between 103 ms and 255 ms, depending on speaking rate; 
the 80% duty cycle may remove short-duration phonemes like /b/ and /?/ while having less 
of an effect on longer-duration phonemes like most vowels. Interruptions always began 
with a full-duration ?on? phase. The remaining sentences (n=150) were interrupted with 
                                                 
8 Pilot testing and previous research (Bhargava et al., 2014) have revealed that duty cycle, or the amount of 
intact speech in a sentence, impacts restoration. For the present study, the 80% duty cycle was chosen based 
on pilot testing with four adult CI users. On average, this duty cycle produced a perceptual restoration 
effect?an improvement in performance with noise bursts compared to silent gaps?of approximately 6% 
with IEEE sentences. In previous restoration research in NH listeners, a 50% duty cycle has been used 
(Bashford et al., 1992; Newman, 2004; Powers & Wilcox, 1977; Saija et al., 2014; Verschuure & Brocaar, 
1983), though some studies have included other duty cycles (Bhargava et al., 2014; Bologna et al., 2018). In 
the present study?s pilot testing, other duty cycles (50, 60, 70, and 90%) produced no restoration effect on 
average. The overall average accuracy at the 80% duty cycle, collapsed across interruption types, was 64% 
(SD=27%). 
49 
 
 
 
speech-shaped noise bursts instead of silent gaps. The noise bursts were not modulated by 
the speech envelope that would have appeared in the missing speech segment. Though 
speech-envelope-modulated noise bursts have been shown to increase restoration (Shinn-
Cunningham & Wang, 2008) over and above non-modulated noise bursts, noise bursts 
encountered in a naturalistic listening environment would be non-modulated, and thus our 
method provided more of a realistic challenge to participants attempting to restore speech. 
Noise bursts were presented at 65 dB SPL with a ?10-dB SNR, meaning that noise bursts 
were 10 dB more intense than the average level of the target speech signal. This negative 
SNR was chosen because previous literature has shown that negative SNRs are typically 
necessary for the strongest restoration effects to occur and are more likely to prompt the 
auditory illusion of speech ?continuing? through noise (Bhargava et al., 2014). Logically, 
a noise that is less intense than speech would not ?mask? the speech, and thus the illusion 
of continuity is less likely to occur (Ba?kent, 2012). 
 
2.1.3 Equipment 
Stimuli were presented through two modes: one mode with several front-end 
preprocessing features turned on (ON condition) and one mode with these front-end 
preprocessing features turned off (OFF condition). Participants wore two research N6 CI 
sound processors, for which features can be adjusted to be on or off. Each participant?s 
own clinical maps were uploaded to the research processors prior to the experiment. 
Stimuli were presented in the free field from two loudspeakers.  
 In the ON condition, several noise reduction algorithms could affect speech 
processing. These included SCAN (a scene analyzer that adapts microphone 
50 
 
 
 
directionality), SNR-NR (a noise reduction algorithm), and WNR (a wind noise reduction 
program). SCAN automatically detects information about the listening environment and 
initiates specific programs for enhancing speech understanding based on the scene selected 
(e.g., Speech, Music, and Speech in Noise). These programs may change microphone 
settings, and the scene classifier is updated about every second (Z. Smith, Cochlear Ltd. 
research scientist, personal communication, July 12, 2018). However, SCAN's 
classifications do not always properly identify the sound environment presented (Mauger 
et al., 2014), potentially introducing variability across and within listeners in how noise is 
processed during restoration. SNR-NR and WNR both attempt to reduce background noise 
and irrelevant loud sounds; these algorithms thus may affect how the noise bursts?the key 
to restoration?are perceived alongside the speech segments. The SNR-NR program 
attempts to detect which frequency channels in the CI contain continuous noise, like a 
generator or the sound of a car?s engine. The channels below a certain SNR are then 
attenuated (Z. Smith, personal communication, July 12, 2018). This attenuation process, 
and potential sluggishness in the way the algorithm is applied, could lead to over-
attenuation of several channels during and/or after the presentation of a quick noise burst. 
This over-attenuation could be leading to reduced perceptual restoration. The WNR 
program detects decorrelation of sound signals recorded between a single sound 
processor?s microphones, the presence of which is potentially indicative of turbulence from 
wind. This program likely has little effect during the restoration paradigm, though it is 
technically a noise-reduction algorithm.  
Two more components were activated in the ON condition: ADRO (adaptive 
dynamic range optimization) and ASC (auto-sensitivity control). These front-end 
51 
 
 
 
preprocessing strategies were expected to create compression effects (Gilden et al., 2015; 
Wolfe et al., 2015). Compression could change the relationship between speech and noise 
signals during noise burst conditions, perhaps by changing the effective SNR. As 
mentioned previously, in typical restoration paradigms, a negative SNR is needed for the 
illusion to occur (Ba?kent, 2012), so changes in the effective SNR could reduce the 
perception of the illusion. ADRO works by slowly changing the channel gains based on 
input from the listening environment, which could, over time, impact the effective SNR. 
ASC is one adjustable loop of the tri-loop AGC (automatic gain control) in Cochlear-brand 
CIs and works by attempting to reduce signal intensity. Other aspects of AGC and how it 
could be influencing restoration will be described more in depth in Experiment 2. In that 
study, we investigate the extent to which AGC may be negatively affecting restoration 
ability in CI users.  
In the OFF condition, SCAN, SNR-NR, WNR, ADRO, and ASC were turned off. 
Comparing speech understanding results between the ON and OFF conditions allows us to 
examine the effects of SCAN, noise reduction, and some compression algorithms on 
restoration. These changes in listening mode, which involved changing front-end 
preprocessing settings from the CI users? typical programming settings, were expected to 
have minimal negative impact on uninterrupted speech understanding in quiet, unlike, for 
example, a change to a frequency-to-electrode allocation, which would necessitate training 
and time for adaptation.  
Finally, a standard omnidirectional microphone setting was used in OFF conditions, 
with the Fixed and Adaptive (Beam) modes turned off. This ensured that all participants in 
were using the same microphone setting in this condition, and that microphone 
52 
 
 
 
characteristics were not automatically changing during the experiment (as could occur with 
Adaptive/Beam). The standard omnidirectional microphone does not apply level 
attenuation to any incoming signals, from any direction, beyond what would occur in a NH 
listener ? that is, slight attenuation from the sides and back of the listener. In contrast, the 
Fixed and Adaptive (Beam) microphones attenuate noise either from any direction besides 
0 degrees (Fixed) or from any direction (Adaptive [Beam]). In the ON conditions, SCAN 
was able to adjust microphone settings.  
 
2.1.4 Procedure 
Order of listening mode (ON vs. OFF) was counterbalanced across participants. 
The first ten sentences in each listening mode were intact, containing no interruptions. All 
other sentences (n=300) were randomly assigned to listening mode and interruption type 
(silent gap vs. noise burst) for each participant. Therefore, any sentence had an equal 
chance of being presented in either of the two listening modes and with either of the two 
interruption types. 
 Participants were seated in a soundproof booth (Industrial Acoustics, Inc., New 
York, NY), one meter away from a pair of loudspeakers located at ?45? from the seated 
participant. Loudspeakers were calibrated using speech-weighted noise to 55 dB SPL. The 
experimenter sat in the booth with the participant and controlled the presentation of 
sentences.  
 On each trial, one sentence was presented, and participants reported aloud what 
they heard. Responses were recorded on a voice recorder and graded in terms of number 
of words correct. Grading was lax, in that incorrect verb conjugations were accepted, 
53 
 
 
 
similar to the grading used in Jaekel et al. (2018). To ensure accurate grading, a second 
grader separately graded a subset of responses (n=8 participants) by listening to the voice 
recordings. Inter-rater reliability was 88.1% based on number of sentences agreed on. 
Inconsistencies were resolved by averaging the scores of the two graders for that specific 
trial.  
 
2.2 Results 
Before adding interruptions to stimuli, participants completed baseline intact 
speech understanding measures in both the ON and OFF conditions (results presented in 
Table 1). In the ON condition, average speech understanding was 95.6% (SD=6.7%, 
range=82 to 100%). In the OFF condition, average speech understanding was 91.4% 
(SD=8.0%, range=76 to 99%). Per a paired-samples t-test, intact speech understanding in 
the OFF condition was significantly poorer than in the ON condition, t(10)=4.19, p=0.002. 
Prior to the main analysis, correlations among the participant variables (see Table 
1) were analyzed to determine which variables could be feasibly included in the mixed 
effects model. Including highly correlated variables in the mixed model analysis can cause 
the model to be less stable and lead to higher standard errors. The variables entered into 
the Pearson correlation analysis were: age, average age at onset of non-normal hearing, 
average duration of non-normal hearing prior to first implantation, average baseline intact 
speech score, age-corrected standard scores on the PPVT-4 (vocabulary test), and age-
uncorrected standard scores for the NIH toolbox tests of processing speed, working 
memory, and attention/executive functioning.  
54 
 
 
 
Age significantly negatively correlated with processing speed (r=?0.81, p=0.003), 
and attention (r=?0.78, p=0.004), and significantly positively correlated with vocabulary 
score (r=0.72, p=0.013); thus, older participants tended to have poorer processing speed 
and attention skills but stronger vocabularies. Average onset of non-normal hearing 
significantly negatively correlated with average duration of non-normal hearing (r=?0.73, 
p=0.011), meaning that participants who experienced hearing loss later in life tended to 
have shorter durations of deafness. Processing speed significantly positively correlated 
with attention (r=0.75, p=0.008) and significantly negatively correlated with vocabulary 
(r=?0.82, p=0.002). No other correlations were significant. Because so many participant 
variables were found to be correlated, a principal components analysis was conducted in 
order to create components that were uncorrelated with one another and could be added to 
the mixed effects model. Principal components analysis with varimax rotation revealed that 
age, vocabulary, attention, and processing speed were all loading as a single component 
(Eigenvalue = 3.4). Average age at onset of non-normal hearing and average duration of 
non-normal hearing loaded as a second component (Eigenvalue = 1.9). Based on this 
analysis, the first component (age, cognition, and vocabulary), the second component 
(hearing history), working memory score, and average baseline intact speech scores were 
retained for mixed effects modeling. 
In addition to these components and participant variables, the variables of 
interruption type (effect coded: ?.5 for silent gaps, +.5 for noise bursts), front-end status 
(effect coded: ?.5 OFF, +.5 ON), their interaction, and their interactions with the 
components and participant variables were included in the mixed effects model. 
Components and participant variables were standardized and centered so that a score of 0 
55 
 
 
 
represented the average of the sample, and scores of ?1 and +1 indicated one standard 
deviation from the mean. Model building followed recommendations from Hox, 
Moerbeek, and van de Schoot (2018). 
The final reduced model for Experiment 1, presented in Table 2, included the fixed 
main effects of interruption type, front-end status, working memory score, and average 
baseline intact scores. The following interactions were also included: interruption type ? 
front-end status and front-end status ? working memory. The components for age, 
vocabulary, and cognition and for hearing history were not significant and were not 
included in the reduced model. The model also included the random intercept for 
participant and the random participant slopes of interruption type and front-end status. 
These random effects indicated that participant intercepts explained 29% of the variance, 
interruption effects across participants explained 10% of the variance, front-end status 
effects across participants explained 2% of the variance, and 58% of the variance was left 
unexplained (residual).   
 
 
 
 
 
 
 
 
 
 
 
56 
 
 
 
Table 2. Final reduced mixed effects model for Experiment 1. SG = silent gaps. NB = noise bursts. 
Fixed Effects Coefficient SE t p 
Intercept 0.598 0.058 10.34 <0.001* 
Interruption type (effect coded = ?.5 SG, .5 NB) ?0.159 0.036 ?4.40 0.001* 
Front-end status (effect coded = ?.5 OFF, .5 ON) ?0.049 0.025 ?1.95 0.07 
Average intact score (standardized) 0.189 0.028 6.75 <0.001* 
Working memory (standardized) 0.083 0.030 2.80 0.03* 
Interruption type ? Front-end status 0.045 0.021 2.13 0.03* 
Front-end status ? Working memory 0.084 0.021 3.99 0.003* 
     
Random Effects Variance SD   
Participant intercept 0.036 0.191   
Interruption type slope 0.013 0.115   
Front-end status slope 0.003 0.056   
Residual 0.072 0.269   
 
 
Figures 1 and 2 display individual and average data from this experiment. Figure 1 
shows speech understanding in percent correct, and Figure 2 shows perceptual restoration 
effects in percent, calculated as performance in the silent-gap condition subtracted from 
performance in the noise-burst condition. As a reminder, the two independent variables in 
this study were interruption type and front-end status. Interruption type significantly 
affected performance, with scores for noise-burst interrupted sentences approximately 
15.9% lower than scores for silent-gap interrupted sentences (p=0.001; see Figs. 1 and 2). 
This is the opposite of the expected restoration effect, where we would expect to see better 
performance with noise-burst compared to silent-gap interruptions. Front-end status was 
not significant as a main effect (p=0.07; Fig. 1), but did significantly interact with 
interruption type (p=0.03). The interaction indicated that in the ON condition (compared 
to the OFF condition), the performance gap between noise-burst interrupted and silent-gap 
interrupted sentences shrank by 4.5%. This was driven mainly by an increase in scores with 
57 
 
 
 
noise-burst interrupted sentences in the ON condition (Fig. 1). The hypothesized effect was 
that noise-burst interruption performance would be greater than silent-gap interruption 
performance, particularly in the OFF conditions; instead, the opposite effect was found. 
Since the ON condition should, in theory, reduce the level of the interrupting noise, making 
noise-burst conditions similar to silent-gap conditions, it is perhaps unsurprising that 
performance was more similar across interruption types in the ON condition. 
 
 
Figure 1. Individual (n=11) and mean speech understanding scores for Experiment 1. Scores with silent-gap 
interrupted sentences are in gold, and scores with noise-burst interrupted sentences are in red. Scores in the 
ON conditions are in a darker hue, and scores in the OFF conditions are in a lighter hue. Mean data is 
presented on the right. SG = silent gaps. NB = noise bursts. 
 
Interrupted Speech Understanding Across 
Front-End Preprocessing Conditions
100%
80%
60%
40%
20%
0%
CAD CAF CAK CAP CAS CAT CAY CBF CCA CCR CDB Mean
SG (ON) NB (ON) SG (OFF) NB (OFF)
 
 
 
 
 
 
58 
 
 
Percent correct
 
Figure 2. Perceptual restoration effects in ON (dark) and OFF (light) conditions. Circles indicate individual 
data, the ? symbol indicates the mean, and the line within the boxplot indicates the median. The top whisker 
indicates the maximum value, and the bottom whisker indicates the minimum value; the top of each box 
indicates the third quartile, and the bottom of each box indicates the first quartile. A circle beyond the 
whiskers indicates an outlier. 
 
 
 
Among participant variables, the following had significant main effects. Average 
baseline intact speech score was a significant main effect (p<0.001), meaning that for each 
1-SD increase above the average intact score for the sample, performance on the 
experiment overall increased by 18.9%. That is, individuals who were better able to 
interpret speech in general also were better at doing so in the presence of interruptions. 
Working memory was also a significant main effect (p=0.03), meaning for each 1-SD 
59 
 
 
 
increase above the average working memory score for the sample, performance on the 
experiment overall increased by 8.3%. The interaction of front-end status and working 
memory was also significant (p=0.003). For each 1-SD increase above the average working 
memory score for the sample, performance in the ON condition increased 8.4%, meaning 
that participants with higher working memory scores performed particularly well in the 
ON compared to the OFF conditions.  
Looking at individual data (Figs. 1 and 2), it was clear there was a wide range of 
performance on the task. While the average trends were consistent across participants (e.g., 
no participant scored better with noise bursts than with silent gaps; only one participant 
scored better with OFF than with ON, with all other participants scoring better with ON 
than with OFF), the variability should be noted. For the noise-burst conditions, 
performance ranged from 9 to 91% correct. For the silent-gap conditions, performance 
ranged from 25 to 97% correct. For high-performing participants (n=6)?those whose 
overall task performance was above 50%?the difference between noise-burst and silent-
gap performance was 10.4%: noise-burst performance was 77.5%, and silent-gap 
performance was 87.9%. In contrast, for lower-performing participants (n=5)?those 
whose overall task performance was below 50%?the difference between noise-burst and 
silent-gap performance was 22.5%: noise-burst performance was 21.0%, and silent-gap 
performance was 43.4%. Thus, lower-performing participants showed a larger gap in 
performance for noise-burst interrupted versus silent-gap interrupted speech conditions.  
 
60 
 
 
 
2.3 Discussion 
 This experiment aimed to evaluate whether front-end preprocessing strategies in 
the Cochlear N6 processor negatively impacted perceptual restoration in CI users. Two 
surprising results emerged. First, we were unable to detect restoration ability in any of our 
CI users. Unlike Bhargava et al. (2014), who found that CI users could restore speech in 
certain conditions with benefits of approximately 8.8%, and unlike results from our own 
pilot testing with four CI users, we found that noise bursts acted as interferers rather than 
facilitators for speech understanding in the perceptual restoration paradigm. On average, 
performance with noise bursts was 15.9% lower than with silent gaps. Second, while front-
end preprocessing strategies did not impact perceptual restoration (as perceptual restoration 
was not observed in either the ON or OFF conditions), noise-burst interrupted speech was 
most improved by access to front-end preprocessing strategies. Improvements from the 
OFF to ON conditions for noise-burst interruptions was 8%, while improvements for silent-
gap interruptions were only 3%. If our hypothesis about restoration had been supported, 
the opposite effect would have occurred: performance with noise-burst interrupted speech 
would be negatively impacted by front-end preprocessing strategies, not enhanced. In 
actuality, the front-end preprocessing served its intended purpose, even with interrupted 
speech: it improved speech-in-noise perception. The real question, then, is why noise 
bursts, compared to silent gaps, was always harmful rather than helpful for CI users. The 
illusion of restoration, and its benefit for speech understanding in realistic listening 
environments, rests on the idea that noise serves a useful rather than harmful purpose, at 
least when it does not further degrade the speech signal itself.  
61 
 
 
 
 Of the components and participant variables, only working memory scores helped 
explain front-end preprocessing effects; no components and participant variables helped 
explain interruption effects. We developed two components based on participant variables: 
one component was composed of participant age, vocabulary score, processing speed, and 
attention; the second component was composed of duration of non-normal hearing prior to 
first implantation and onset of non-normal hearing. Both components made logical sense: 
it is unsurprising that older age is related to higher vocabulary scores and poorer cognitive 
abilities (Drag & Bieliauskas, 2010; Park et al., 2002). Furthermore, it is logical that many 
participants with early onsets of non-normal hearing would also be the participants with 
longer durations of non-normal hearing prior to implantation, as early implantation of 
children did not become common practice until fairly recently. Unfortunately, neither of 
these components helped explain effects in this experiment, and were not even significant 
predictors of general performance (i.e., they were not significant main effects). While it 
was predicted that age and interruption type would interact in a way that would benefit 
noise-burst interrupted speech performance, as was observed in work by Jaekel et al. (2018) 
and Saija et al. (2014), this was not the case in the present study. Aging is thought to be 
associated with better perceptual restoration because older adults can draw on the 
vocabularies they have developed over a lifetime to process speech, as crystallized 
intelligence like vocabulary is believed to be unaffected by aging processes, unlike some 
cognitive functions (Saija et al., 2014). However, older CI users in the present study did 
not show extra ability to utilize the perceptual restoration effect, as was observed in older 
NH listeners presented unprocessed (Saija et al., 2014) and vocoded (Jaekel et al., 2018) 
speech. Furthermore, greater vocabulary knowledge or stronger linguistic skills (apart from 
62 
 
 
 
age) have previously been shown to be associated with greater amounts of restoration, at 
least in NH listeners presented unprocessed speech (Bashford et al., 1992; Benard et al., 
2014), and to a lesser extent in NH listeners presented vocoded speech (Jaekel et al., 2018). 
Taken together, aging and vocabulary appear to affect interrupted speech understanding 
differently in CI users compared to NH listeners. While they are helpful factors among NH 
listeners, they are not predictive of interrupted speech understanding in CI users. Collison, 
Munson, and Carney (2004) reported that adult CI users appeared to show less of an effect 
of linguistic processing on speech recognition skills, concluding that the poor-quality input 
CI users experience makes it more difficult for them to link incoming speech sounds to 
representations in their lexicons. Perhaps the addition of interruptions to speech stimuli, as 
occurred in the present experiment, made this process even more difficult, disallowing the 
use of vocabulary and other crystallized intelligence during the restoration task in our older 
adult CI users. 
 Working memory interacted with front-end status, but did not interact with 
interruption type. The latter finding is perhaps unsurprising when considering previous 
work in this area, which has shown that working memory did not mediate restoration in 
younger- or older-NH listeners presented unprocessed speech or vocoded speech (Benard 
et al., 2014; Jaekel et al., 2018). Instead, in the present study, working memory interacted 
significantly with front-end status, meaning that compared to people with average working 
memory scores, CI users with higher working memory achieved particularly higher 
performance in ON, compared to OFF, conditions. Purdy et al. (2017) similarly found a 
?trend? where CI users with higher working memory scores achieved higher speech-in-
noise understanding with specifically SNR-NR turned ON compared to OFF. CI users with 
63 
 
 
 
lower working memory scores, in contrast, showed no differences in performance between 
the two SNR-NR conditions (Purdy et al., 2017). Perhaps CI users with lower working 
memory scores in the current study are generally negatively affected by the distortions 
introduced by interruptions, regardless of type of interruption. Comparatively, CI users 
with higher working memory can effectively process interrupted speech that is less 
distorted (presumably the ON condition produces a signal that is less changed by the 
presence of noise interruptions) but fail to do so in the more distorted condition (OFF 
condition), when working memory resources may be particularly taxed (Finke, B?chner, 
Ruigendijk, Meyer, & Sandmann, 2016). Research in the hearing aid literature, cited by 
Purdy et al. (2017), seems to support the idea that listeners with hearing loss and higher 
working memory scores are more resistant to distortions in speech signals (Arehart, Souza, 
Baca, & Kates, 2013; Ohlenforst, Souza, & MacDonald, 2016). 
  In summary, we cannot conclude whether front-end preprocessing strategies affect 
restoration in CI users, as we were unable to detect restoration in our sample. Interestingly, 
noise-burst interrupted speech performance was improved by access to front-end 
preprocessing strategies, but not by enough to overtake performance with silent-gap 
interrupted speech. There are several possibilities that could explain these findings: (1) 
front-end preprocessing strategies do not negatively impact noise-burst interrupted speech 
processing via noise-reduction and compression effects, (2) participants found the general 
quality of the speech signal to be improved with the front-end preprocessing strategies 
turned on, and thus were more able to effectively process the difficult noise-burst 
interrupted speech, or (3) since almost all participants? daily maps included access to at 
least some of the front-end preprocessing strategies, experience with speech processed in 
64 
 
 
 
this way resulted in better noise-burst interrupted speech performance in this condition. A 
final possibility is that other aspects of CI processing outside of controllable front-end 
preprocessing algorithms prohibited the ability of these CI users to restore speech. Thus, 
perhaps other areas of interest in the speech processing pathway ? namely, the quality of 
peripheral auditory encoding and/or the access to semantic information ? may better serve 
to promote restoration in these listeners.  
 
  
65 
 
 
 
Chapter 3: Experiment 2: Measuring effects of presentation level 
and peripheral auditory encoding on restoration 
The intensity of speech stimuli, especially in relation to noise bursts, is important 
for the restoration effect to occur in NH listeners (Bashford et al., 1992; Ba?kent, 2012). It 
is important that noise bursts be perceived as louder than speech in this paradigm, as the 
noise is then more able to serve as a plausible masker of the speech and promote an illusion 
that speech is continuing through the noise, intact and uninterrupted (Ba?kent, 2012). 
Previous work has analyzed how noise burst intensity affects restoration in CI users: 
Bhargava et al. (2014) studied how CI users restored meaningful Dutch sentences 
interrupted with either silent gaps or noise bursts at various SNRs. Speech was always 
presented at 60 dB(A), and noise was presented at either 55, 60, 65, or 70 dB(A), which is 
equivalent to +5-, 0-, ?5-, or ?10-dB SNRs. Contrary to the hypothesized importance of 
negative SNRs for restoration, CI users showed no variation in restoration effects across 
SNRs: at the 50% duty cycle, no restoration was observed at any SNR; at the 75% duty 
cycle, the restoration effect was similar at every SNR. Thus, CI users were not showing 
typical, expected effects of level differences between speech and noise during the 
restoration task.  
One aspect of CI processing that could be affecting the results observed by 
Bhargava et al. (2014) is the role of automatic gain control (AGC). Because the dynamic 
range for a CI user is smaller compared to that of a NH listener, compression is necessary 
to convey speech signals in the range from audible to comfortable loudness (Khing et al., 
2013). The SNRs presented by Bhargava et al. (2014) may not have been perceived as such 
by the CI users due to this dynamic compression from the AGC. Fast-acting AGC will 
66 
 
 
 
respond quickly (<10 ms attack time) to sudden loud noises like a door slam, while slow-
acting AGC will adjust the intensity of the incoming speech signals over time (Khing et 
al., 2013). The level at which AGC begins to compress the signal in Cochlear-brand CIs is 
around 70 dB SPL (Z. Smith, personal communication, July 12, 2018). Therefore, the most 
negative SNR measured in the study by Bhargava et al. (2014), which contained noise 
bursts at the 70 dB(A) level and is the SNR that we would expect to show a large, prominent 
restoration effect, may have been impacted by compression, reducing the perceived 
loudness differences between speech and noise bursts. Besides the general presence of 
AGC in the processing algorithm of the CI, the speed with which the AGC turns on/off 
could also impact restoration mechanisms (Ba?kent et al., 2009). On/off changes distort the 
amplitude envelope of speech, and resulted in reduced restoration in NH listeners presented 
vocoded speech (Ba?kent et al., 2009). Thus, our hypothesis was that CI users may 
experience a reduced restoration effect when noise bursts engage AGC, due to envelope 
distortions and changes to the effective SNR, compared to when noise bursts are less 
intense and do not engage AGC.    
 Experiment 2 investigated how AGC affected restoration and added the further 
element of how AGC and peripheral auditory encoding ? that is, the transmission of the 
electrical signals to various regions of the auditory nerve ? interacted to affect restoration. 
Bilateral CI users often report having a poorer-performing and a better-performing ear 
(Kan, 2018). For the present study, the poorer-performing ear was expected to have poorer 
electrode/nerve interface, resulting in poorer-quality bottom-up acoustic information due 
to less effective encoding of the incoming signals. Additional methods used to classify 
better ears and poorer ears include the extent of neural survival, order of implantation, ease 
67 
 
 
 
of telephone use, and etiology of hearing loss (Kan, 2018; Litovsky, Parkinson, Arcaroli, 
& Sammeth, 2006). Our second hypothesis was that perhaps AGC has less of an impact on 
restoration in better-performing ears, which can better encode differences between speech 
and noise, regardless of SNR. AGC is based on the signal and a maximum knee-point level 
(approximately 70 dB SPL), and is not determined with reference to the individual?s own 
hearing.  
 
3.1 Methods 
3.1.1 Participants 
 Twelve bilaterally implanted adult CI users participated in this study. Information 
about these participants is presented in Table 3. The mean age was 64.3 years (SD=14.4 
years, range=32 to 81 years). Because this study tested each ear separately, Table 3 presents 
age at onset of non-normal hearing and non-normal hearing duration prior to implantation 
for each ear individually. Participants were native monolingual speakers of American 
English and self-reported normal or corrected-to-normal vision. All participants used 
Cochlear-brand CIs in both ears with N6 processors. All participants had had their most 
recent CI activated for at least 6 months prior to testing, ensuring adequate experience with 
the devices. All participants passed the Montreal Cognitive Assessment (MoCA) with a 
score of >22 (Nasreddine et al., 2005), indicating a lack of mild/moderate cognitive 
impairment.  
 Participants also completed a battery of cognitive tests available from the NIH 
Toolbox (Gershon et al., 2013). The age-uncorrected standard scores were collected for 
these participants. For working memory (?List Sorting Working Memory Test Age 7+?), 
68 
 
 
 
the mean score was 98.6 (SD=12.4, range=78 to 113). For processing speed (?Pattern 
Comparison Processing Speed Test Age 7+?), the mean score was 98.8 (SD=17.8, 
range=63 to 122). For attention and executive functioning (?Flanker Inhibitory Control and 
Attention Test Age 12+?), the mean score was 100.8 (SD=6, range=94 to 113). Finally, 
participants completed the PPVT-4 (Dunn & Dunn, 2007), a vocabulary test, with a mean 
age-corrected standard score of 105.8 (SD=12.5, range=93 to 134).  
 
  
69 
 
 
 
Table 3. Participant variables for Experiment 2. 
 
70 
 
 
 
3.1.2 Stimuli 
Stimuli were 580 sentences (created by the author). The design of the experiment 
was two ear presentations (poorer, better ear) ? four interruption/level conditions (speech 
at 55 dB SPL with silent gaps, speech at 65 dB SPL with silent gaps, speech at 55 dB SPL 
with 65 dB SPL noise bursts, speech at 65 dB SPL with 75 dB SPL noise bursts), for a total 
of eight conditions. The condition with speech at 55 dB SPL and noise at 65 dB SPL tested 
restoration below the AGC knee-point, meaning AGC was not expected to be engaged; the 
condition with speech at 65 dB SPL and noise at 75 dB SPL tested restoration above the 
AGC knee-point, meaning AGC was expected to be engaged. The engagement of AGC 
was expected to change the effective SNR by compressing peaks of incoming sound signal 
to reduce overall levels; such compression could weaken the illusion of speech continuing 
through noise and ultimately decrease restoration. Additionally, participants completed a 
baseline (intact speech) condition and two control conditions (speech at 55 dB SPL with 
75 dB SPL noise bursts, and speech at 65 dB SPL with 65 dB SPL noise bursts). The control 
conditions were necessary for confirming that any restoration observed in the test 
conditions was not due solely to speech presentation level (i.e., perhaps speech at 65 dB 
SPL is just generally easier to restore than speech at 55 dB SPL, regardless of SNR). It is 
not possible to deactivate the AGC, unlike the other front-end preprocessing features (see 
Experiment 1); thus, we needed to manipulate the intensity of our presented stimuli to 
analyze the effects of AGC on restoration. 
 For the eight main test conditions, 480 of the 580 sentences were allotted. The 
sentences were interrupted following the paradigm described in Experiment 1, generating 
240 silent-gap interrupted sentences and 240 noise-burst interrupted sentences at an 80% 
71 
 
 
 
duty cycle. One-hundred-twenty sentences of each interruption type were presented to each 
ear, termed the ?poorer? or ?better? ears. Sixty sentences of each interruption type, at each 
ear, had speech presented at 65 dB SPL, and sixty sentences had speech presented at 55 dB 
SPL. For noise-burst interrupted sentences, the noise was presented at ?10-dB SNR, based 
on the average level of the speech.  
 For the baseline test, 20 intact sentences were presented to the poorer ear and 20 to 
the better ear. Ten of these sentences presented to each ear were presented at 55 dB SPL, 
and ten were presented at 65 dB SPL. For the control conditions, the remaining 60 
sentences were allotted to the following conditions. Fifteen sentences with speech 
presented at 55 dB SPL and noise-burst interruptions at 75 dB SPL were presented to the 
poorer ear and the better ear (n=30). Fifteen sentences with speech presented at 65 dB SPL 
and noise-burst interruptions at 65 dB SPL were presented to the poorer ear and better ear 
(n=30). Results from these control conditions were expected to help identify speech 
presentation level effects versus SNR effects. The summary of sentence types and 
presentation conditions is presented in Table 4.  
 
 
 
 
 
 
 
 
 
 
72 
 
 
 
Table 4. Design of Experiment 2. 
 
 Speech at 55 dB SPL Speech at 65 dB SPL 
? 10 intact sentences ? 10 intact sentences 
? 15 control sentences (noise-burst ? 15 control sentences (noise-burst 
interrupted; noise at 75 dB SPL) interrupted; noise at 65 dB SPL) 
? 60 experimental sentences ? 60 experimental sentences 
Poorer ear 
(silent-gap interrupted) (silent-gap interrupted) 
? 60 experimental sentences ? 60 experimental sentences 
(noise-burst interrupted; noise at (noise-burst interrupted; noise at 
65 dB SPL) 75 dB SPL) 
? 10 intact sentences ? 10 intact sentences 
? 15 control sentences (noise-burst ? 15 control sentences (noise-burst 
interrupted; noise at 75 dB SPL) interrupted; noise at 65 dB SPL) 
? 60 experimental sentences ? 60 experimental sentences 
Better ear 
(silent-gap interrupted) (silent-gap interrupted) 
? 60 experimental sentences ? 60 experimental sentences 
(noise-burst interrupted; noise at (noise-burst interrupted; noise at 
65 dB SPL) 75 dB SPL) 
 
3.1.3 Equipment 
 Participants completed half of the experiment with their self-reported ?better? ear 
only and completed the other half of the experiment with their self-reported ?poorer? ear 
only. During the task, participants wore a calibrated research N6 processor, containing the 
participants? clinical maps associated with that ear with all front-end preprocessing 
algorithms turned on, in the target ear. Participants removed the CI from the non-target ear. 
Participants who reported residual acoustic hearing had the ear(s) with residual hearing 
plugged.  
 
3.1.4 Procedure 
Participants were seated in a soundproof booth, with the same setup as that 
described in Experiment 1. The order of the four listening blocks was randomized for each 
participant. The four listening blocks were combinations of poorer vs. better ear 
73 
 
 
 
presentation and speech at 55 dB SPL vs. speech at 65 dB SPL (see Table 4). Within each 
block, the first ten sentences were intact, containing no interruptions and presented in quiet. 
The next fifteen sentences were the associated control sentences for that block, containing 
noise-burst interruptions. The next 120 sentences were the test sentences, 60 of which 
contained noise-burst interruptions and 60 of which contained silent-gap interruptions. 
Sentences were randomly assigned to condition and interruption type; any sentence had an 
equal chance of being presented in any of the four listening conditions as an intact sentence, 
control sentence, or test sentence. To ensure accurate grading, a second grader separately 
graded a subset of responses (n=7 participants) by listening to the voice recordings. Inter-
rater reliability was 88.0% based on number of sentences agreed on. Inconsistencies were 
resolved by averaging the scores of the two graders for that specific trial.  
 
3.2 Results 
3.2.1 Main Findings 
Prior to the main analysis, correlations among the participant variables were 
analyzed. Including highly correlated variables in the mixed model analysis can cause the 
model to be less stable and lead to higher standard errors. The variables entered into the 
Pearson correlation analysis were: age, age at onset of non-normal hearing for each ear, 
average duration of non-normal hearing prior to implantation for each ear, age-corrected 
standard scores on the PPVT (vocabulary test), and age-uncorrected standard scores for the 
NIH toolbox tests of processing speed, working memory, and attention/executive 
functioning.  
74 
 
 
 
Age significantly negatively correlated with processing speed (r=?0.82, p=0.001), 
and attention (r=?0.73, p=0.008), and significantly positively correlated with duration of 
non-normal hearing in the left ear (r=0.59, p=0.045) and vocabulary score (r=0.68, 
p=0.015). Older listeners therefore tended to have poorer cognition, better vocabularies, 
and a longer period of non-normal hearing in one of their ears. Processing speed 
significantly positively correlated with attention (r=0.72, p=0.009) and significantly 
negatively correlated with vocabulary (r=?0.80, p=0.002). Duration of non-normal hearing 
in the left ear was significantly negatively correlated with age at onset of non-normal 
hearing in the left ear (r=?0.80, p=0.002), and, similarly, duration of non-normal hearing 
in the right ear was significantly negatively correlated with age at onset of non-normal 
hearing in the right ear (r=?0.77, p=0.003). Finally, duration of non-normal hearing in the 
right ear was significantly negatively correlated with working memory (r=?0.59, p=0.044), 
and age at onset of non-normal hearing in the right ear was significantly positively 
correlated with vocabulary (r=0.64, p=0.026).  
Principal components analysis with varimax rotation revealed that age, attention, 
processing speed, and vocabulary were all loading as a single component (Eigenvalue = 
3.6, Percent of variance = 36.1%). This component indicated that older age was associated 
with better vocabulary, slower processing speed, and poorer attention skills. Onset of non-
normal hearing and duration of non-normal hearing in the left ear, as well as average 
baseline intact score, loaded on a second component (Eigenvalue = 2.3, Percent of variance 
= 23.4%). This component indicated that shorter durations of non-normal hearing in the 
left ear was associated with later onsets of non-normal hearing in that ear, as well as better 
average baseline intact speech scores. Onset of non-normal hearing and duration of non-
75 
 
 
 
normal hearing in the right ear, as well as working memory score, loaded on a third 
component (Eigenvalue = 2.2, Percent of variance = 21.9%). This component indicated 
that, like the second component, shorter durations of non-normal hearing in the right ear 
was associated with later onsets of non-normal hearing in that ear, as well as higher 
working memory scores. These three components were included in the mixed effects model 
analysis.  
This experiment aimed to compare effects across ears, to determine the impact of 
peripheral auditory encoding on perceptual restoration in different compression conditions. 
Unfortunately, two of the 12 participants (CAF and CCA) were unable to complete testing 
in their poorer ear, due to very low intact speech understanding scores in that ear. Thus, 
CAF and CCA only contributed data to the better ear condition. For the rest of the sample, 
ear differences in performance were not large on average: average interrupted speech 
scores with the better ear were 51%, and average interrupted speech scores with the poorer 
ear were 47.5%, a difference of only 3.5%. When this smaller sample?s data (n=10 
participants) were entered into a mixed effects model containing the random intercepts of 
participant and ears within participants and the fixed main effects and interactions of ear 
presentation, interruption type, and level, the main effect of ear presentation was not 
significant (p=0.11), nor were any interactions with ear presentation (p values ranged from 
0.10 to 0.60). With this in mind, the ear effect was dropped from the analysis, and only the 
better ear scores were ultimately analyzed, which maintained our sample size of 12.  
Results for the mixed effects model analysis are presented in Table 5. Individual 
and average results are plotted in Figure 3, and perceptual restoration effects are plotted in 
Figure 4. The mixed effects model (Table 5) revealed the following. First, overall 
76 
 
 
 
performance in the study was 51.3% correct. Noise-burst interruptions significantly 
decreased performance compared to silent-gap interruptions (p<0.001; see Fig. 3), the 
opposite of the restoration effect. Increasing the intensity of the stimuli from 55 to 65 dB 
SPL did not significantly change overall performance (p=0.13); however, level and 
interruption type did significantly interact (p=0.03). Performance with silent gaps 
decreased from 57.6% to 55.3% as level increased from 55 to 65 dB SPL, while 
performance with noise bursts increased from 43.0% to 46.4%.  
In terms of participant variables (i.e., components), only component 2 was found 
to have a significant effect on performance: with each 1-SD increase in scores on this 
component, overall performance increased by 15.2% (p=0.002). As a reminder, higher 
scores on component 2 were associated with shorter durations of non-normal hearing in 
the left ear, later onsets of non-normal hearing in the left ear, and better average baseline 
intact speech scores. Finally, variability in participant intercepts explained 40% of the 
variance in the study, participants? interruption type, level, and interruption type ? level 
effects each explained 2 to 4% of the variance in the study, and 52% of variance was left 
unexplained (residual).  
 
 
 
 
 
 
 
 
77 
 
 
 
Table 5. Results for the mixed effects model analysis for Experiment 2. SG = silent gaps. NB = noise 
bursts. 
 
Fixed Effects Estimate SE t p 
Intercept 0.513 0.067 7.61 <0.001* 
Interruption type (effect coded = ?.5 SG, .5 NB) ?0.117 0.023 ?5.15 <0.001* 
Level (effect coded = ?.5 55 dB SPL, .5 65 dB SPL) 0.032 0.020 1.63 0.13 
Component 2 (Left Ear Duration/Onset & Average Intact Score) 0.152 0.044 3.43 0.002* 
Interruption type ? Level 0.058 0.025 2.32 0.03* 
     
Random Effects Variance SD   
Participant intercept 0.054 0.233   
Interruption type slope 0.005 0.071   
Level slope 0.003 0.056   
Interruption type ? Level slope 0.003 0.053   
Residual 0.069 0.263   
 
Figure 3. Individual and average results for Experiment 2, in the better ear only. Performance with silent 
gaps is in gold, and performance with noise bursts is in red. Dark gold and dark red represents performance 
in the condition where speech was presented at 55 dB SPL; pale gold and pale red represents performance in 
the condition where speech was presented 65 dB SPL. SG = silent gaps. NB = noise bursts.  
 
Interrupted Speech Understanding Across 
Level Conditions
100%
80%
60%
40%
20%
0%
CAD CAF CAK CAP CAS CAT CAY CBC CBF CCA CCR CDB Mean
SG (55 dB SPL) NB (55 dB SPL) SG (65 dB SPL) NB (65 dB SPL)
 
 
78 
 
 
Percent correct
 
Figure 4. Perceptual restoration effects across level conditions in Experiment 2. Circles indicate individual 
data, and the ? symbol indicates the mean for each condition. The line within each boxplot indicates the 
median for each condition.  
 
 
Overall, we could not confirm restoration effects either above or below the AGC 
knee-point in the better ear (see Figs. 3 and 4). That is, noise-burst performance was never 
higher than performance with silent-gap interrupted speech in the more intense conditions 
engaging AGC (65 dB SPL conditions) and the less intense conditions not engaging AGC 
(55 dB SPL conditions), on average. On an individual level, two participants showed a 
small perceptual restoration benefit: CBF in the 55 dB SPL condition (0.5% benefit) and 
CCR in the 65 dB SPL condition (0.3% benefit). Individual variability in performance 
across participants was also observed, with participants? overall average scores ranging 
from 9% to 89% correct, and average restoration effects ranging from ?30 to ?1%.  
79 
 
 
 
3.2.2 Control Conditions 
 We aimed to test whether presenting noise above the AGC knee-point affected 
restoration; however, by changing our stimuli from 55 dB SPL to 65 dB SPL, we were also 
changing the intensity of both the noise and the speech signal. Thus, control conditions 
were presented to participants to test whether more intense speech and/or more intense 
noise were generally easier to restore. Results from these control conditions are presented 
in Figure 5. 
  
Figure 5. Average results from test and control conditions, separated by speech presentation level. Test 
results with silent-gap interrupted speech (?SG?) are presented in gold. Test results with noise-burst 
interrupted speech (?Test NB?) are presented in dark red. Control conditions (n=15 sentences per condition) 
with noise-burst interrupted speech (?Control NB?) are presented in gray. White boxes indicate the noise 
level and SNR of noise bursts compared to the associated silent-gap condition.  
 
Comparing Test and Control Conditions in 
Experiment 2
60%
50%
40%
30%
65 dB SPL  75 dB SPL  
20% (?10 SNR) (?10 SNR) 
10% 75 dB SPL  65 dB SPL  
(?20 SNR) (0 SNR) 
0%
Speech at 55 dB SPL Speech at 65 dB SPL
SG Test NB Control NB
 
 
80 
 
 
Percent correct
 
The best noise-burst interrupted speech performance occurred when speech was 
more intense (65 dB SPL, compared to 55 dB SPL; see Fig. 5) and noise was equivalently 
intense (i.e., had a 0 dB SNR). Performance dropped slightly when the level of speech was 
maintained at 65 dB SPL but noise bursts grew more intense, decreasing to a ?10 dB SNR. 
Performance with noise-burst interrupted speech decreased overall when the level of 
speech was less intense (i.e., 55 dB SPL). At this speech level, performance was better with 
a less negative SNR (i.e., ?10 dB SNR) compared to a more negative SNR (i.e., ?20 dB 
SNR). Based on results from these control conditions, more intense speech signals in the 
context of noise-burst interruptions may be easier to understand (though in the context of 
silent-gap interruptions, performance with more intense speech signals appears to decrease; 
see Fig. 5). Furthermore, engaging AGC ? that is, presenting noise bursts at 75 dB SPL ? 
appears to always decrease performance with noise-burst interruptions, regardless of the 
level of the speech. Changing the level of the noise but keeping the level of the speech 
constant, and, conversely, changing the level of the speech but keeping the level of the 
noise constant, did not reveal equivalent performance across conditions. Therefore, in 
general, the level of speech information does not seem to be driving restoration; speech at 
55 dB SPL or 65 dB SPL does not appear to be particularly more restorable, and noise at 
65 dB SPL or 75 dB SPL does not appear to be particularly more restorable. Our main 
findings likely reflect a change in performance resulting from changing both speech and 
noise, together, from both being below the AGC knee-point to crossing the AGC knee-
point. 
   
81 
 
 
 
3.2.3 Poorer Ear Performance 
 Ten of twelve participants were able to complete the experimental task in their 
poorer ear. Poorer ear performance was characterized, unsurprisingly, by overall lower 
performance. Compared to the better ear, the restoration effect in the poorer ear at 55 dB 
SPL was ?12% (compared to the better ear?s ?15%). In this condition, silent-gap 
performance was 1% lower and noise-burst performance was 4% lower in the poorer ear 
than in the better ear. The restoration effect in the poorer ear at 65 dB SPL was ?7% 
(compared to the better ear?s ?9%). In this condition, silent-gap performance was 5% lower 
and noise-burst performance was 7% lower in the poorer ear than in the better ear. Thus, 
restoration effects were quite consistent across the two ear conditions. The most stable 
performance across ears was at 55 dB SPL with silent gaps, and the least stable 
performance was at 65 dB SPL with noise bursts. As mentioned above, when these data 
were entered into a mixed effects model, neither the main effect of ear presentation nor any 
interactions with ear presentation were significant. For this sub-sample, then, impacts of 
peripheral auditory encoding differences between ears did not interact with AGC effects or 
interruption types.  
 
3.3 Discussion 
 The restoration illusion requires the presence of a plausible masker within a speech 
signal interruption (Bashford et al., 1992; Ba?kent, 2012), usually in the form of a noise 
that is louder than the surrounding speech. Compression in CIs, which occurs via an 
algorithm called AGC and is necessary due to dynamic range constraints in CI users, may 
generally change the intensity of noise in relation to speech and distort speech envelopes 
82 
 
 
 
(Ba?kent et al., 2009; Khing et al., 2013). Both of these factors could reduce restoration 
when AGC turns on at the knee point of 70 dB SPL.  
 This study tested 12 participants to detect whether restoration ability was reduced 
when speech and noise stimuli straddled the AGC knee point vs. when speech and noise 
stimuli were presented at levels below the AGC knee point. Furthermore, ears were tested 
separately to detect whether peripheral auditory encoding interacted with compression 
effects. Previous research has reported that bilateral CI users may have a ?better? and 
?poorer? performing ear, potentially due to differences in electrode/nerve interface (Kan, 
2018; Litovsky et al., 2006). The overall hypothesis was that AGC might be less negatively 
impactful on restoration in a ?better? ear, as the ear should presumably be better able to 
encode speech information and would be less affected by distortions introduced by 
compression. 
 Only ten participants were able to complete the task in both ears; two participants 
had such poor performance in their poorer ears with intact uninterrupted sentences that it 
was not reasonable to test them with interrupted sentences. For the main analysis, only the 
better ear data were analyzed, so that data from all tested participants could be included. 
We found that, similar to Experiment 1, restoration was not observed on average (Figure 
4). That is, speech understanding with noise-burst interruptions was generally much poorer 
than with silent-gap interruptions. Two participants technically showed restoration 
benefits, one with speech at 55 dB SPL and one with speech at 65 dB SPL, but these 
restoration benefits were extremely small.  
 Differences between silent-gap and noise-burst performance shrank when stimuli 
straddled the AGC knee point, and appeared to be primarily driven by improvements in the 
83 
 
 
 
noise-burst interruption condition. Thus, when compression was engaged, with all its 
potential concomitant envelope distortions, the processing of noise-burst interrupted 
speech actually improved, contrary to our hypothesis. Our participants generally showed 
improved performance with smaller SNRs (see 0 dB SNR control condition in Fig. 5), so 
perhaps reductions in the effective SNR due to compression was actually helpful in the 65 
dB SPL level test condition. This potential influence of SNR on restoration performance 
was not observed by Bhargava et al. (2014), who presented CI users with a range of SNRs, 
from ?10 to +5. While NH listeners require negative SNRs to perceive a noise as a plausible 
masker and thus prompt the restoration illusion, CI users may need positive SNRs in order 
to successfully process noise-burst interrupted speech. Future research presenting positive 
SNRs in a restoration paradigm to CI users would help confirm this notion. Furthermore, 
having access to more intense speech information in the 65 dB SPL conditions appeared to 
be helpful, at least in the context of noise-burst interruptions (see Fig. 5). More intense 
speech, compared to less intense speech, may be more resilient against distortions 
introduced by compression.  
 We were initially interested in how a better vs. poorer ear interacted with 
compression effects to affect restoration. Accuracy and restoration effects in the two level 
conditions were similar across ears. It was difficult to characterize what made an ear 
?poorer? or ?better.? Self-report did not always reflect performance. Six of 12 participants 
showed a discrepancy between their self-reported ?better? ear and the ear with the highest 
intact speech understanding score. Participants were likely using other criteria than their 
own perception of their own intact speech understanding ability to decide which ear was 
?better.? Interestingly, of these six participants, three had very small differences in intact 
84 
 
 
 
speech scores between ears (1.9 to 3.1%), and three had quite dramatic differences in scores 
(9.4 to 20.5%), meaning the ?misidentification? was not a result of participants having two 
similarly performing ears.  
 We hypothesized that a poorer ear would struggle to encode the distortions 
introduced by compression algorithms, and thus lead to worse performance. This is because 
a poorer ear was expected to have poorer electrode/nerve interface (Kan, 2018), potentially 
due to longer durations of non-normal hearing, which could cause auditory neurons to 
deteriorate in that ear (Fetterman & Domico, 2002). However, seven out of 12 participants 
in this experiment had fairly symmetrical durations of non-normal hearing across ears (?5 
years difference). Furthermore, nine out of 12 participants had fairly symmetrical onsets of 
non-normal hearing across ears (?5 years difference). If these two factors ? duration of 
non-normal hearing and age at onset of non-normal hearing ? are expected to help identify 
a poorer vs. better ear, our sample was perhaps too symmetric to allow for such a 
categorization. Future research should attempt to characterize what makes an ear a poorer-
performing one, and what metrics CI users use when identifying their poorer vs. better ears. 
Perhaps some bilateral CI users feel they have symmetric hearing across ears, and thus 
asking them to identify a poorer ear is leading to a false forced choice.  
 To summarize, restoration benefits were again not observed among participants 
with CIs. Contrary to our hypothesis, engaging AGC appeared to improve noise-burst 
interrupted speech understanding. This finding contradicts previous research in this area, 
and supports the notion that perhaps the traditional restoration paradigm is not functional 
with CI users. More favorable SNRs may be required for restoration to work in this 
population ? the opposite of what is needed for NH listeners. 
85 
 
 
 
Chapter 4: Experiment 3: Measuring interactions of peripheral 
auditory encoding and prior linguistic knowledge on restoration 
Priming has been shown to strongly enhance the restoration effect (Samuel, 1981). 
In the present study, we tested if providing a semantic cue ? here, a single word 
meaningfully related to the content of the upcoming sentence ? can effectively ?prime? a 
listener for the upcoming sentence and increase restoration. Semantic cues can activate 
meaningful associations that allow for faster and more efficient processing of upcoming 
speech (McNamara, 2005). However, semantic cues may enhance restoration differently 
based on the quality of the bottom-up acoustic information, namely whether that acoustic 
information is highly degraded. A bilateral CI user may experience differing levels of 
degradation across their two devices due to having a ?better ear? and a ?poorer ear? (Kan, 
2018). A poorer ear may have experienced less success with a CI, possibly due to poor 
electrode/nerve interface, and this could be reflected in poorer speech understanding 
scores. We would expect such an ear to experience high levels of signal degradation. In 
contrast, less severe signal degradation may occur in an ear with comparatively better 
electrode/nerve interface (i.e., a ?better ear?). Experiencing more signal degradation, 
specifically, having access to fewer channels of spectral information, has been shown to 
reduce restoration in NH listeners (Ba?kent, 2012; Bhargava et al., 2014; Clarke et al., 
2016). By testing bilateral CI users, we can measure the extent to which ears contributing 
potentially different levels of degradation on bottom-up acoustic cues interact with top-
down linguistic knowledge, all while holding linguistic skill constant (i.e., utilize a within-
subjects design). This integration of top-down knowledge and bottom-up acoustic 
86 
 
 
 
information is key to understanding speech in noisy environments (Ba?kent, 2012; Patro & 
Mendel, 2016; Shinn-Cunningham & Wang, 2008).  
 We hypothesized that CI users would be able to utilize linguistic knowledge to 
restore speech with a better ear, but would fail to do so with a poorer ear. This is because 
greater degradation of the bottom-up acoustic information in the poorer ear would prevent 
successful integration with top-down knowledge.  
 
4.1 Methods 
4.1.1 Participants 
Eighteen CI users participated in Experiment 3. Two additional CI users were 
tested, but were dropped from the analysis due to equipment failure during experiment 
presentation. Table 6 presents information about tested participants. The mean age of 
participants was 63.7 years old (SD=13.3 years, range=32 to 81 years). Because this 
experiment utilized an ear presentation manipulation, information about each ear 
individually is presented in the table. On average, left ears experienced earlier ages at onset 
of non-normal hearing (m=19.7 years, vs. right ears at m=26.9 years) and longer durations 
of non-normal hearing prior to implantation (m=34.1 years, vs. right ears at m=28.2 years).  
The ?better ear? designation was based on performance with ten baseline intact 
sentence scores presented at 55 dB SPL; for this group, 10 participants had a right ?better 
ear? and eight participants had a left ?better ear.? These designations often conflicted with 
patient self-report of which ear was their ?better ear?: eight participants reported the 
opposite ear as their ?better ear.? For some of these participants, ear performance at 
baseline was quite similar across ears, so it is unsurprising that an opposite ear was reported 
87 
 
 
 
as the ?better ear.? Only three participants with ?mismatched? ear designations had 
comparatively large (i.e., greater than 4%) performance differences between their self-
reported better ear and best-performing ear on the baseline test (see participants CAY, 
CBC, and CBF in Table 6).  
  
88 
 
 
 
Table 6. Demographic information for participants in Experiment 3. 
 
 
  
89 
 
 
 
In general, baseline scores were similar across ears; only six of 18 participants 
showed performance differences between ears greater than 10%. On average, right ears had 
a baseline score of 91.5% (SD=7.8%), and left ears had a baseline score of 86.3% 
(SD=14.8%), an across-ear difference of 5.2%. When considering better ear vs. poorer ear 
performance, better ears had an average baseline score of 94.2% (SD=5.7%), and poorer 
ears had an average baseline score of 83.2% (SD=14.0%), an across-ear difference of 
11.7%.  
When age at onset of non-normal hearing was considered for better ears vs. poorer 
ears (rather than left vs. right ears), the average age at onset in a better ear was 27.0 years 
(SD=24.1, range=0 to 70 years) and in a poorer ear was 19.5 years (SD=17.4, range=0 to 
55 years). Thus, better ears tended to experience non-normal hearing at later ages compared 
to poorer ears. In a similar vein, the average duration of non-normal hearing prior in a better 
ear was 27.6 years (SD=21.8, range=1 to 58 years) and in a poorer ear was 34.6 years 
(SD=23.5, range=0 to 68 years). Thus, better ears tended to have shorter durations of non-
normal hearing compared to poorer ears. 
Processing speed, working memory, attention, and vocabulary scores were 
generally near standard scores of 100, with the greatest standard deviations observed for 
processing speed and vocabulary (Table 6). Thus, this group had generally average 
executive functioning and average vocabulary knowledge. Finally, all participants passed 
the Montreal Cognitive Assessment (MoCA) with scores of 24 or greater (m=27.1), 
indicating a lack of mild or moderate cognitive impairment (Nasreddine et al., 2005). 
 
90 
 
 
 
4.1.2 Stimuli 
Stimuli were 240 IEEE sentences, which are declarative sentences containing 5-12 
words (Rothauser et al., 1969). The sentences were recorded by an adult male speaker of 
Standard American English dialect. The two interruption types (silent gaps, noise bursts) 
were be applied to sentences in the same manner as described in Experiment 1, such that 
120 sentences were interrupted with silent gaps and 120 sentences were interrupted with 
noise bursts.  
 A semantic cue (a single word meaningfully related to the content of the sentence 
about to be presented) was presented visually on a computer monitor prior to each sentence, 
for 120 sentences (60 of which were silent-gap interrupted, and 60 of which were noise-
burst interrupted). Semantic cues were generated in the following way. Three assistants 
without knowledge of the present experiment were provided a list of the 720 total IEEE 
sentences, and were asked to generate one to two words for each sentence that were 
meaningfully related to the sentence content. The answers were compiled and the most 
commonly reported related word, or the word judged most appropriate by the experimenter, 
was chosen as that sentence's ?semantic cue? word. For example, the word ?fish? was 
chosen for the IEEE sentence ?A rod is used to catch pink salmon.? The assistants were 
instructed that words in the target sentence could not serve as cues, nor could any 
conjugation of a verb in the target sentence.  
 One cue word was associated with each of the 720 sentences through this method. 
To determine which 240 of these sentences had the ?best? cue words, to be used in the real 
experiment, a short pilot study was conducted. Four young NH adults participated in the 
pilot study, which was conducted using PsychoPy2 (Peirce et al., 2019). Each participant 
91 
 
 
 
was presented all 720 sentences auditorily over headphones, one at a time in random order. 
After the presentation of each sentence, a written word appeared on the computer monitor 
and participants decided via a lexical decision task whether the word was a real word or 
non-word, using a keyboard press to indicate their choice. Each participant was presented 
180 IEEE sentences (one-fourth of the set) followed by the related cue word, 180 sentences 
followed by an unrelated word (a cue word associated with a different sentence in the set), 
and 360 sentences followed by a non-word (Rastle, Harrington, & Coltheart, 2002). The 
target 180 sentences (the ones with the related cue word of interest) were rotated among 
the four participants; thus, for Participant 1, IEEE sentences 1 through 180 were followed 
by the related cue word; for Participant 2, IEEE sentences 181 through 360 were followed 
by the related cue word; and so on. Accuracy and reaction times in the lexical decision task 
were reviewed following testing.  
Accuracy was high, with participants achieving an average of 96.3% correct in their 
lexical decisions. Reaction times were, on average, 0.539 seconds for related words, 0.550 
seconds for unrelated words, and 0.620 seconds for non-words. All reaction time data for 
related words from all four participants was compiled into a single document, and the 240 
sentences with the fastest reaction times (and correct classifications) to the cue words were 
selected for the main experiment. In this subset, the difference between reaction times for 
unrelated vs. related words was 0.03 seconds, meaning the related words were identified 
0.03 seconds faster on average than unrelated words. A better method to consider for future 
work in this area would be instead to select the subset of sentences based on the unrelated 
word reaction time ? related word reaction time for the same sentence/word pair, which 
would directly capture the priming effect. With this method, we could have achieved an 
92 
 
 
 
unrelated vs. related words difference of 0.16 seconds, had we selected the 240 sentences 
with the highest values on this measure. As it is, our subset contains words with which the 
participants likely had the most experience (i.e., words with high frequency in their 
lexicons). Still, our main manipulation in the main experiment was the presentation of a 
real word vs. no word at all, so while our real words may not have always been among the 
strongest primed sentence/word pair from pilot testing, the real word should be high 
frequency enough to be easily accessible by CI users.  
Any one participant in the main experiment was only randomly presented 120 
sentences from this set of 240 sentences with related cue words. For another 120 sentences 
(not duplicates of the first 120), no semantic cue was provided to the participant. Instead, 
during these trials, a series of X symbols (equal in length to the target sentence?s associated 
semantic cue word) was presented prior to the target sentence. No sentences appeared twice 
to the same participant as both a cued and uncued sentence. Sentences were randomly 
assigned to cued vs. uncued conditions for each participant.  
 
4.1.3 Equipment 
 Participants were seated in a soundproof booth, 1 meter in front of a computer 
monitor located at eye-level at 0?. Sentences were presented over a pair of loudspeakers 
located at ?45?. A computer keyboard was used for participants to record responses. 
MATLAB 2018b (Mathworks; Natick, MA) was used to administer the experiment. 
 To measure linguistic knowledge, participants completed the Peabody Picture 
Vocabulary Test, Fourth Edition (PPVT-4), which measures receptive vocabulary size 
(Dunn & Dunn, 2007). Scores from this test were considered a proxy measure of the 
93 
 
 
 
participant's receptive language ability. Furthermore, participants completed a battery of 
cognitive tests available via the NIH Toolbox Cognition Battery iPad application (Gershon 
et al., 2013; Tulsky et al., 2014). For attention and executive functioning, Flanker 
Inhibitory Control and Attention Test Age 12+ was used; for working memory, List Sorting 
Working Memory Test Age 7+ was used; for processing speed, Pattern Comparison 
Processing Speed Test Age 7+ was used. The test battery was administered on an iPad 2 
(Apple, Inc.; Cupertino, CA) in a quiet location. The test battery was completed in 15 
minutes or less. 
 
4.1.4 Procedure 
 The design of the experiment was ear presentation (two levels: ?better? ear, 
?poorer? ear), semantic cue (two levels: present, absent), and interruption (two levels: silent 
gap, speech-shaped noise burst), for a total of eight conditions with 30 sentences per 
condition.  
 Before the presentation of each sentence, participants focused on a cross hair 
presented in the middle of the computer screen. In the semantic cue ?present? condition, a 
word semantically related to the sentence (e.g., ?FISH?) replaced the cross hair for two 
seconds and disappeared. The sentence was then immediately auditorily presented (e.g., 
?A rod is used to catch pink salmon?). In the semantic cue ?absent? condition, a series of 
X characters (equal in length to the semantic cue word associated with that sentence, e.g., 
?XXXX?) replaced the cross hair instead. Speech with noise interruptions was presented 
at ?10-dB SNR, with speech presented at 55 dB SPL. Participants were instructed to listen 
to each sentence and repeat aloud what they heard into a voice recorder. When finished, 
94 
 
 
 
participants were instructed to press the space bar to begin the next trial. The experiment 
was self-paced. Two experimenters graded responses separately, one live and one off the 
voice recording. Experimenters recorded the number of words correct for each sentence. 
Inter-rater reliability for the full dataset (n=18) was 82.3% based on number of sentences 
agreed on. Inconsistencies were resolved by averaging the scores of the two graders for 
that specific trial. 
 
4.2 Results 
4.2.1 Multilevel Model Analysis 
Prior to running the model, the extent to which any participants showed restoration 
was inspected by analyzing the mean data for each participant in each of the ear 
presentation ? semantic cue conditions (see Figure 6). Seven participants (out of 18) always 
performed better with silent-gap interruptions compared to noise-burst interruptions, and 
thus never showed restoration. Among the remaining participants, three participants 
showed restoration in the better ear/semantic cue present condition; three participants 
showed restoration in the better ear/semantic cue absent condition; five participants showed 
restoration in the poorer ear/semantic cue present condition; and six participants showed 
restoration in the poorer ear/semantic cue absent condition. Four participants (out of 18) 
showed restoration in multiple conditions; for example, CBC always showed restoration 
with the poorer ear, and CCA always showed restoration when a semantic cue was present.  
 
 
 
 
95 
 
 
 
Figure 6. Illustration of which participants showed restoration effects across the various listening conditions. 
 
Participants who show restoration in each condition 
 
Better Ear,  Better Ear,  Poorer Ear,  Poorer Ear,  
Cue No Cue Cue No Cue 
CAP    
CCA  CCA  
CDQ CDQ  CDQ 
 CAS   
 CAK CAK CAK 
  CBH  
  CES  
  CBC CBC 
   CAB 
   CAD 
   CDB 
No Restoration: CAQ, CAT, CAY, CBF, CBR, CCO, CDS 
  
Next, we investigated whether participant variables showed multi-collinearity, 
which can result in a non-converging multilevel model. Any participant variables 
exhibiting multi-collinearity could be submitted to a principal component analysis in order 
to generate uncorrelated composite variables. The variables entered into the Pearson 
correlation analysis were age, age at onset of non-normal hearing for better and poorer ears, 
duration of non-normal hearing prior to implantation for better and poorer ears, processing 
speed, working memory, attention, and vocabulary scores. Age was significantly 
negatively correlated with processing speed (r=?0.85, p<0.001) and attention (r=?0.64, 
p=0.005), and was significantly positively correlated with vocabulary (r=0.70, p=0.001) 
and duration of non-normal hearing in the poorer ear (r=0.63, p=0.005). Thus, older 
participants typically had poorer cognition, while also having larger vocabularies and 
longer durations of hearing loss in their poorer ear specifically.  
96 
 
 
 
In terms of cognitive measures, processing speed was significantly positively 
correlated with attention (r=0.66, p=0.003), and significantly negatively correlated with 
vocabulary (r=?0.68, p=0.002) and duration of non-normal hearing in the poorer ear 
(r=?0.59, p=0.01). Attention was significantly negatively correlated with duration of non-
normal hearing in the poorer ear (r=?0.66, p=0.003). Working memory did not correlate 
with any other measure.  
In terms of hearing history measures, age at onset of non-normal hearing in the 
better ear was significantly positively correlated with age at onset of non-normal hearing 
in the poorer ear (r=0.57, p=0.014), and significantly negatively correlated with duration 
of non-normal hearing in the better ear (r=?0.85, p<0.001). This indicated that participants 
with later ages of onset of non-normal hearing in one ear were also more likely to have a 
later age of onset in the other ear; furthermore, a later age at onset was associated with a 
shorter duration of hearing loss in that ear. This pattern was also observed for age at onset 
of non-normal hearing in the poorer ear, which was significantly negatively correlated with 
both the duration of non-normal hearing in the better ear (r=?0.69, p=0.002) and in the 
poorer ear (r=?0.81, p=<0.001). Duration of non-normal hearing in the better ear was 
positively correlated with duration of non-normal hearing in the poorer ear (r=0.61, 
p=0.007).  
Principal components analysis was used for compiling these highly correlated 
participant variables onto components. Using varimax rotation and Kaiser normalization, 
two components were extracted with Eigenvalues greater than 1 (see Fig. 7). We found that 
age, processing speed, attention, and vocabulary loaded as a single component 
(Eigenvalue=3.4, percent of variance=43.0%). That is, older age was associated with 
97 
 
 
 
slower processing speeds, poorer attention, and larger vocabularies. We found that 
durations of non-normal hearing in the better and poorer ears, and ages at onset for the 
better and poorer ears loaded as a second component (Eigenvalue=3.1, percent of 
variance=38.7%). That is, later onsets of non-normal hearing in the better and poorer ears 
were associated with shorter durations of non-normal hearing in the better and poorer ears. 
These two components were included as the participant variables in the multilevel model. 
A third participant variable, working memory, was also included in the multilevel model. 
Working memory did not significantly correlate with any other participant variable, and 
did not primarily load on either of the two principal components outlined above. 
 
Figure 7. Component plot in rotated space of the eight variables entered into the principal components 
analysis. All variables were entered in standardized, z-score form. Blue symbols represent variables loading 
primarily on Component 1. Yellow symbols represent variables loading primarily on Component 2. 
 
98 
 
 
 
In addition to the participant variables, the following variables were included as 
fixed factors in a linear multilevel model: semantic cue (absent vs. present; effect-coded as 
?.5 and +.5, respectively), interruption type (silent gap vs. noise burst; effect-coded as ?.5 
and +.5, respectively), ear presentation (poorer ear vs. better ear; ?.5 and +.5, respectively), 
and all interactions. The dependent variable was percent words correct per sentence. 
Model-building followed recommendations outlined in Hox et al. (2018). The final reduced 
model, presented in Table 7, included the random intercepts of participant (explaining 31% 
of the variance), ears within participants (explaining 15% of the variance), and sentence 
(explaining 8% of the variance), as well as the sentence random slope of interruption type 
(explaining 8% of the variance). Per the residual value, 38% of the variance in results was 
left unexplained. Figure 8 shows average overall performance of participants across the 
various conditions, and Figure 9 shows perceptual restoration effects.  
 
Table 7. The best-fitting mixed effects model for Experiment 3. SG = silent gaps. NB = noise bursts. 
 
Fixed Effects Estimate SE t p 
Intercept 0.45 0.05 8.74 <0.001* 
Semantic cue (effect coded = ?.5 no cue, +.5 cue) 0.05 0.01 6.81 <0.001* 
Interruption type (effect coded = ?.5 SG, +.5 NB) ?0.06 0.01 ?5.68 <0.001* 
Ear presentation (effect coded = ?.5 poorer, +.5 better) 0.09 0.05 1.77 0.095 
Component 1 (Age, Cognition, Vocabulary) ?0.12 0.05 ?2.25 0.036* 
Interruption type ? Ear presentation ?0.05 0.02 ?3.40 <0.001* 
Interruption type ? Component 1 0.03 0.01 3.73 <0.001* 
     
Random Effects Variance SD   
Participant intercept 0.04 0.19   
Ear within participant intercept 0.02 0.14   
Sentence intercept 0.01 0.11   
Interruption type slope (by sentence) 0.01 0.10   
Residual 0.05 0.23   
  
99 
 
 
 
Figure 8. Average performance with interrupted speech across ear and cue conditions. Error bars indicate 
standard error. SG = silent gaps. NB = noise bursts. 
 
Average Interrupted Speech Understanding 
Across Ear and Cue Conditions
60%
50%
40%
30%
20%
10%
0%
Better Ear, Better Ear, Poorer Ear, Poorer Ear,
Cue No Cue Cue No Cue
SG NB
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
100 
 
 
Percent correct
 
Figure 9. Perceptual restoration effects in Experiment 3, across ear and cue conditions. 
 
 
 
 
In terms of overall average accuracy, performance with silent gaps was always 
better than with noise bursts (see Fig. 8; p<0.001), and performance with semantic cues 
present was always better than when semantic cues were absent (p<0.001). The effect of 
ear presentation was not significant (p=0.095), but the interaction of interruption type with 
ear presentation was significant (p<0.001). Figure 10 illustrates this interaction. Using the 
emmeans package available in R (R Core Team, Vienna, Austria), we found that the 
interaction was significant because with silent gaps, the better ear performed significantly 
better than the poorer ear (p=0.024), while with noise bursts, both ears performed similarly 
(p=0.22). This means that the better ear had a particular advantage over the poorer ear in 
101 
 
 
 
the silent-gap condition, specifically. The better ear achieved 53% accuracy with silent 
gaps, and the poorer ear achieved 44% ? a difference of 9%. In contrast, the better ear 
achieved 45% with noise bursts, and the poorer ear achieved 41% ? a difference of 3%. In 
both ear conditions, performance with silent gaps was always significantly higher than 
performance with noise bursts (for better ears, p<0.001; for poorer ears, p=0.014), 
demonstrating the lack of a restoration effect in either ear.  
 
Figure 10. Average performance with interrupted speech across ears (collapsed across cue type). The 
interaction of ear and interruption type was significant, per the mixed effects model. Error bars indicate 
standard error. SG = silent gaps. NB = noise bursts. 
 
Interaction of Ear and 
Interruption Type
60%
55%
50%
45%
40%
35%
30%
Better Ear Poorer Ear
SG NB
 
 
The main effects of working memory and Component 2 were not significant and 
did not interact with any other variables, and so were removed from the model. The main 
effect of component 1, associated with age, processing speed, attention, and vocabulary, 
was significant. As scores on Component 1 increased (meaning participants were older, 
102 
 
 
Percent correct
 
had slower processing speeds, poorer attention, and larger vocabularies), overall 
performance significantly decreased (p=0.036). However, as Figure 11 illustrates, 
Component 1 interacted significantly with interruption type (p<0.001). This indicated that 
performance with noise bursts improved relative to performance with silent gaps as 
Component 1 scores increased. Thus, more positive restoration effects were associated with 
older CI users with poorer cognitive skills and larger vocabularies. While this relationship 
was significant, very few participants actually experienced a restoration benefit (Fig. 11). 
 
Figure 11. Differences in performance with noise bursts vs. silent gaps (i.e., restoration effects) are plotted 
against Component 1 scores for each participant. As scores on Component 1 increase, restoration effects 
become more positive, meaning performance with noise bursts begins to improve relative to performance 
with silent gaps. 
 
Restoration Effects Correlated 
with Component 1
10%
5%
0%
-5%
-10%
-15%
-20%
-2 -1 0 1 2
Component 1 Score
 
  
103 
 
 
Percpetual Restoration Effect
 
4.2.2 Cluster Analysis 
While the multilevel model analysis helped identify several important aspects of 
the data, it mainly focuses on average results across the participants. As shown previously 
in Figure 6, several participants in the study did show restoration effects, albeit across 
several different listening conditions without a clear pattern of results. However, looking 
only at the multilevel model, one would conclude that no participants on average achieved 
restoration benefit. Thus, one more analysis was used that considered participant variables 
and how they affected perceptual restoration ability. 
 Options for this participant variable analysis included considering comparing 
participants who experienced non-normal hearing prior to full language development (i.e., 
?prelingual? participants) to those participants who lost hearing after language 
development (i.e., ?postlingual? participants). Possibly non-normal language development 
could lead to less stable lexical representations, putting participants with prelingual hearing 
loss at a disadvantage for perceptual restoration. Another option was comparing 
participants with symmetric versus asymmetric hearing histories ? that is, participants who 
reported non-normal hearing loss in both ears at the same age, versus participants who 
reported non-normal hearing loss at different ages for different ears. Possibly continued 
access to normal hearing in at least one ear would reduce the chance that perceptual 
restoration ability deteriorates. A third option was comparing participants with asymmetric 
baseline scores across their better and poorer ears to those with symmetric baseline scores 
across ears. For example, CAY, CBC, CBF, CBR, CCO, and CES all had greater than 10% 
differences in performance across their better and poorer ears. These participants were 
104 
 
 
 
functionally affected by asymmetric hearing, and perhaps this would impact restoration 
ability in one ear versus the other.  
 All of these possibilities could be evaluated with a cluster analysis. Cluster analysis 
is a method of sorting participants into groups based on their scores on variables of interest. 
Participants in each group are thus more similar to one another on those variables (Pastor, 
2010). In the present study, the variables by which participants were classified were the 
standardized variables of age at testing, processing speed score, attention score, working 
memory score, vocabulary score, onsets of non-normal hearing in the better and poorer 
ears, and durations of non-normal hearing in the better and poorer ears (n=9 variables). 
Standardization of variables prior to cluster analysis is important to prevent different scales 
from biasing the clustering process. Unfortunately, intact speech scores could not be 
included in the cluster analysis, as two participants did not have intact speech score data 
and would have been removed entirely from the analysis (see Table 6).  
 A hierarchical cluster analysis was conducted using a between-groups linkage 
method and squared Euclidean distance. Three clusters emerged from the analysis based 
on inspection of the dendrogram (Fig. 12). Cluster 1 contained seven participants, Cluster 
2 contained seven participants, and Cluster 3 contained four participants. 
 
 
 
 
 
 
105 
 
 
 
Figure 12. Dendrogram for cluster analysis. The y-axis lists participants. The x-axis denotes rescaled 
distances. The three clusters are highlighted in navy, blue, and pale blue.  
 
 
 
 Table 8 presents how each cluster differed from one another on each of the variables 
entered into the cluster analysis (on the original scales), and also lists the mean scores on 
each variable for the sample as a whole. Cluster 1 was characterized largely by their hearing 
histories; this group experienced early, symmetrical onset of non-normal hearing in both 
ears, and experienced similarly long durations of non-normal hearing. This group was 
called the ?symmetrical group with early onsets/long durations of non-normal hearing.? 
Cluster 2 was characterized by a mixture of age, cognitive abilities, and hearing histories; 
they were typically younger than the rest of the sample and had better cognitive scores. 
This group tended to have symmetric hearing histories across ears, experiencing hearing 
106 
 
 
 
loss in their 30s on average, with ears experiencing about a decade of loss prior to 
implantation. This group was called the ?symmetrical group with late onsets/short 
durations of non-normal hearing.? Cluster 3 was also characterized by a mixture of age, 
cognitive abilities, and hearing histories; participants in this group were typically the oldest 
participants in the sample, and had large vocabularies. Their hearing histories were 
asymmetric, with a late onset and short duration of loss in the better ear. This group was 
called the ?older asymmetrical group.?  
 
Table 8. Average scores on each of the variables for each cluster, as well as for the sample as a whole. 
Standard deviations are listed within parentheses. 
 
Cluster 1 Cluster 2 Cluster 3 Entire Sample 
(n=7) (n=7) (n=4) (n=18) 
 
Age (years) 67 (9) 52 (10) 78 (3) 64 (13) 
Processing Speed 92 (8) 113 (10) 80 (15) 98 (17) 
Working Memory 94 (9) 104 (7) 98 (15) 99 (10) 
Attention 96 (4) 106 (6) 97 (3) 100 (7) 
Vocabulary 104 (13) 101 (7) 118 (15) 106 (13) 
Better Ear - Age at Onset (years) 4 (4) 30 (9) 63 (8) 27 (24) 
Poorer Ear - Age at Onset (years) 4 (4) 31 (10) 27 (22) 20 (17) 
Better Ear - Duration of Loss (years) 53 (7) 14 (7) 7 (5) 28 (22) 
Poorer Ear - Duration of Loss (years) 54 (7) 11 (8) 41 (27) 35 (24) 
 
 
 Cluster designation was entered into a mixed effects model also containing the 
variables of interruption type, semantic cue, and ear presentation. The final reduced model 
is presented in Table 9. Cluster 1 was used as the referent category in this analysis. As a 
reminder, Cluster 1 is the ?symmetrical group with early onsets/long durations of non-
normal hearing.? For Cluster 1, average performance overall was 36% words correct; 
107 
 
 
 
adding noise bursts decreased performance (p<0.001), and adding semantic cues increased 
performance (p<0.001). For Cluster 1 and all other clusters, adding noise bursts to speech 
presented to the better ear decreased performance further (p=0.001). This is the same 
interaction observed in the previous multilevel model analysis (see Table 7 and Fig. 10).  
 Cluster 2, the ?symmetrical group with late onsets/short durations of non-normal 
hearing,? performed significantly better overall compared to Cluster 1 (p=0.038), 
achieving approximately 61% words correct. This group experienced an even bigger 
decrease in performance with noise bursts compared to Cluster 1 (p=0.01), while the 
helpful effect of semantic cue on performance was the same across the two groups. 
 Cluster 3, the ?older asymmetrical group? performed overall similarly to Cluster 1, 
achieving approximately 36% words correct. In contrast to Cluster 1 and Cluster 2, 
participants in Cluster 3 showed an increase in performance with noise bursts, i.e., a 
restoration effect (p<0.001). This interaction of cluster and interruption type is presented 
in Figure 13. Cluster 3 also experienced a significant interaction of semantic cue and ear 
presentation, with the provision of a semantic cue particularly increasing scores in the 
better ear specifically (p=0.02).  
  In summary, the main takeaway from this analysis is that Cluster 3 was the only 
cluster to benefit from a restoration effect. This cluster is characterized by older age, 
asymmetrical hearing, and strong vocabularies. This restoration effect benefit for Cluster 
3 was not impacted by the provision of a semantic cue or ear presentation. 
 
 
 
 
108 
 
 
 
Table 9. Mixed effects model analysis including cluster designation.  
 
Fixed Effects Estimate SE t p 
Intercept 0.36 0.08 4.65 <0.001* 
Cluster 2 (referent: Cluster 1) 0.25 0.11 2.24 0.038* 
Cluster 3 (referent: Cluster 1) -0.02 0.13 -0.12 0.91 
Semantic cue (Effect coded: ?.5 cue absent, +.5 cue present) 0.04 0.01 3.70 <0.001* 
Interruption type (Effect coded: ?.5 silent gap, +.5 noise burst) -0.05 0.01 -4.14 <0.001* 
Ear presentation (Effect coded: ?.5 poorer, +.5 better) 0.12 0.07 1.64 0.12 
Cluster 2 ? Semantic cue 0.01 0.02 0.47 0.64 
Cluster 3 ? Semantic cue 0.02 0.02 1.04 0.30 
Cluster 2 ? Interruption type -0.04 0.02 -2.58 0.01* 
Cluster 3 ? Interruption type 0.07 0.02 3.48 <0.001* 
Cluster 2 ? Ear presentation -0.05 0.11 -0.49 0.63 
Cluster 3 ? Ear presentation -0.11 0.13 -0.81 0.43 
Semantic cue ? Ear presentation -0.04 0.02 -1.62 0.10 
Interruption type ? Ear presentation -0.05 0.02 -3.36 0.001* 
Cluster 2 ? Semantic cue ? Ear presentation 0.02 0.03 0.45 0.66 
Cluster 3 ? Semantic cue ? Ear presentation 0.09 0.04 2.25 0.02* 
     
Random Effects Variance SD   
Participant (Intercept) 0.03 0.18   
Ear within participant (Intercept) 0.02 0.14   
Sentence (Intercept) 0.01 0.11   
Sentence (Interruption type slope) 0.01 0.10   
Residual 0.05 0.23   
 
 
 
 
 
 
 
 
109 
 
 
 
Figure 13. Performance across interruption types and clusters. Cluster 1 (n=7) is the ?symmetrical group 
with early onsets/long durations of non-normal hearing.? Cluster 2 (n=7) is the ?symmetrical group with late 
onsets/short durations of non-normal hearing.? Cluster 3 (n=4) is the ?older asymmetrical group.? SG = silent 
gap. NB = noise burst. 
 
Interaction of Interruption 
Type with Cluster
70%
60%
50%
40%
30%
20%
Cluster 1 Cluster 2 Cluster 3
SG NB
 
 
4.3 Discussion 
 This experiment analyzed the extent to which different levels of degradation in 
bottom-up acoustic information affected integration with top-down linguistic knowledge 
during perceptual restoration in CI users. We hypothesized that ?poorer ears,? which likely 
experience greater signal degradation, would fail to restore speech, while ?better ears,? 
with less signal degradation, would successfully interact with top-down linguistic 
knowledge to prompt restoration. We hoped to find a three-way interaction of interruption 
type, ear presentation, and provision of a semantic cue: performance with noise bursts was 
expected to be highest (and higher than performance with silent gaps) in the better ear when 
110 
 
 
Percent correct
 
a semantic cue helped prime the upcoming sentence. Previous research has shown the 
benefit of priming as a way to prompt restoration in normal-hearing listeners (Samuel, 
1981); it was unknown if a similar benefit occurred among CI users.  
 Unlike in Experiments 1 and 2, some participants in the present study showed 
restoration benefits (see Fig. 9). While 39% of participants did not show restoration in any 
ear presentation/semantic cue condition, 39% of participants showed restoration in one of 
the conditions, 11% of participants showed restoration in two of the conditions, and another 
11% showed restoration in three of the conditions. Despite this promising individual data, 
no average restoration benefits were apparent (see Figs. 8 and 9). This lack of an average 
restoration benefit among CI users was confirmed by the multilevel model analysis (Table 
7) ? the hypothesized three-way interaction was not significant, and no improvements with 
noise-burst interrupted speech over silent-gap interrupted speech were observed. 
Performance with noise-burst interrupted speech was always significantly poorer than 
performance with silent-gap interrupted speech, on average.  
 What we found, instead, was an overall helpful effect of semantic cue for repairing 
interrupted speech in general ? whether those interruptions were silent gaps or noise bursts. 
Thus, top-down linguistic knowledge appears to be utilized by CI users whenever 
interrupted speech is encountered. We also found that noise bursts were particularly 
harmful when presented to the better ear (Fig. 10), whether semantic cues were present or 
absent. We predicted that restoration would be more likely in the better ear, as bottom-up 
acoustic cues were expected to be higher in quality. High-quality bottom-up cues have been 
purported to be important for successful speech repair (Ba?kent, 2012; Bhargava et al., 
2014; Jaekel et al., 2018) and stimulating context usage (Patro & Mendel, 2016). Instead, 
111 
 
 
 
the better ear reacted to noise bursts as interferers rather than prompters of restoration, 
while silent-gap interrupted speech improved significantly. Perhaps the high-quality 
encoding of speech information was what was most useful to CI users in the better ear, 
since when this interrupted but high-quality speech information was presented alone 
(silent-gap condition), speech understanding increased. Finally, our model showed that the 
impact of noise bursts on speech understanding were less pronounced if participants were 
older, had stronger vocabularies, and poorer cognition (see Fig. 11). While this decrease in 
noise-burst interference was not pronounced enough to elicit restoration (i.e., noise-burst 
performance was still lower than performance with silent gaps), this finding does align 
with previous work in the area showing that older listeners, and people with stronger 
vocabularies, are more likely to repair noise-burst interrupted speech (Benard et al., 2014; 
Jaekel et al., 2018; Nagaraj & Magimairaj, 2017; Saija et al., 2014).  
 While the multilevel model did not reveal any significant restoration effects for CI 
users on average, some CI users appeared to be experiencing restoration (Fig. 9). To detect 
whether certain groups of CI users, based on their personal characteristics like age, 
vocabulary, cognition, and hearing histories, could achieve restoration, we conducted a 
cluster analysis. Three clusters of CI users emerged (Table 8, Fig. 12). Participants in 
Cluster 1 typically had symmetrical hearing histories across their ears, with early onsets of 
non-normal hearing and long durations of non-normal hearing. Such hearing histories are 
usually associated with poorer speech perception outcomes, with longer durations of 
deafness being correlated with poorer speech understanding in quiet (Blamey et al., 2013; 
Green et al., 2007) and in noise (Fetterman & Domico, 2002), likely due to anatomical and 
physiological changes to the auditory system. Prelingual hearing loss, meaning 
112 
 
 
 
experiencing non-normal hearing before the age of three, is also associated with poorer 
speech recognition, potentially due its impacts on spoken language acquisition and 
reorganization of the auditory cortex (Boisvert, McMahon, Dowell, & Lyxell, 2015; 
Petersen, Gjedde, Wallentin, & Vuust, 2013; Peterson, Pisoni, & Miyamoto, 2010). Cluster 
1 showed no restoration, confirmed with a second multilevel model (see Fig. 13 and Table 
9).  
 Participants in Cluster 2 also typically had symmetrical hearing histories across 
their ears, but instead had late onsets of non-normal hearing and short durations of non-
normal hearing. Despite likely being less affected by anatomical/physiological changes to 
the auditory system caused by long durations of hearing loss, and thus likely having more 
similar auditory system organization to NH listeners, this group was the most negatively 
affected by noise-burst interruptions. They showed the strongest noise interference effect 
(see Fig. 13).  
 Participants in Cluster 3 had asymmetrical hearing histories and were typically 
older. This cluster was the only group to show a restoration benefit (Fig. 13). Furthermore, 
participants in this cluster showed semantic cue and ear presentation effects: interrupted 
speech performance (in general) was higher in the better ear when semantic cues were 
available (Table 9). While we did not observe the predicted interaction of interruption type, 
semantic cue, and ear presentation in this cluster, it was encouraging to find that priming 
and less-degraded bottom-up acoustic information may boost interrupted speech 
understanding in general. Having a group of CI users that had asymmetrical hearing 
histories likely increased the chances that we could observe ear presentation effects, via 
113 
 
 
 
the designation of a poorer and better ear that more likely appreciably differed in 
degradation levels (Kan, 2018).  
 To summarize, CI users on average showed an interference effect rather than a 
restoration benefit when speech was interrupted with noise bursts. This finding goes against 
previous literature (Bhargava et al., 2014) but aligns with findings in Experiments 1 and 2. 
A closer analysis of the results from the current study revealed that some CI users were 
able to restore speech, but there were no clear patterns for which ear presentation/semantic 
cue conditions were most likely to prompt restoration. 
 When CI users were categorized into sub-groups based on their hearing histories 
and other variables, we found that older CI users with asymmetric hearing histories were 
most able to restore speech. Previous work has shown that older age is associated with 
greater restoration benefits, as older listeners tend to use context more while processing 
speech (Pichora-Fuller, 2008; Sheldon et al., 2008), increasing top-down contributions to 
noise-burst interrupted sentence understanding (Jaekel et al., 2018; Saija et al., 2014). This 
group was also most able to utilize semantic information and a better ear to process 
interrupted in speech in general (i.e., both silent-gap and noise-burst interrupted speech). 
Our original hypothesis posited that the use of semantic information and a better ear would 
particularly boost performance with noise bursts, prompting restoration; this was not the 
case. Restoration in this group was not tied to the presence of a semantic prime or ear used 
to complete the task. There appears to be a disassociation between gleaning restoration 
benefit vs. processing interrupted speech in this group of CI users. The latter is directly 
affected by access to high-quality bottom-up acoustic signals and the opportunity to apply 
top-down linguistic knowledge. The former, restoration, is unaffected by variables that 
114 
 
 
 
impact the on-line processing of speech, and instead is more affected by qualities of the 
participant him- or herself, like hearing history and age. This group had, on average, the 
latest better ear onset of non-normal hearing (63 years) and the shortest better ear duration 
of non-normal hearing (7 years). Thus, this group had the longest experience with normal 
hearing in at least one ear, and therefore likely had the most experience with utilizing 
restoration as a tactic to repair speech. Future work should recruit more CI users with these 
qualities (only four CI users in our study fit this profile) and confirm that restoration ability 
in adult CI users may depend on experience with restoration with normal hearing. It would 
also be of interest to conduct cluster analyses on larger samples of CI users to determine 
whether the listener profiles found in our analysis exist in the CI user population at large. 
Listener profiles could be used to predict speech understanding outcomes and/or inform 
expectations for CI candidates.  
 While some individual CI users may be able to restore speech, restoration still 
appears to be elusive for CI users as a group. The individual CI users who could restore 
speech in Experiment 3, and who participated in the other experiments in this project, failed 
to demonstrate restoration in Experiments 1 and 2. Thus, even CI users who can restore in 
some cases fail to do so all the time. Overall, it appears that manipulations in noise-
reduction algorithm availability (Experiment 1), compression engagement (Experiment 2), 
ear presentation (Experiments 2 and 3), and semantic cue availability generally fail to elicit 
restoration in CI users. When restoration was observed, it was not consistent within 
listeners or across speech materials. CI users are more successful at understanding silent-
gap interrupted rather than noise-burst interrupted sentences, and show no strong evidence 
of performing speech repair.   
115 
 
 
 
Chapter 5: Experiment 4: Replicating previous work with CI users 
and perceptual restoration 
 Following several failures to find consistent restoration in CI users in Experiments 
1-3, we next attempted to replicate the original, and only, restoration study performed in 
CI users to date, which was conducted by Bhargava et al. (2014). In addition to testing the 
parameters used in that original study, we tested additional duty cycles and interruption 
rates to get a better sense of how these factors influence restoration in CI users. 
 The authors of the original study were interested in measuring whether CI users? 
difficulties with speech understanding in noisy environments were due to reduced 
restoration ability, with this reduction being caused by how CI processing affects bottom-
up acoustic cues. The authors tested restoration in not only CI users, but also in NH listeners 
and NH listeners presented vocoded speech. The main findings were that under only certain 
conditions could CI users utilize restoration, and that the loss of restoration ability in other 
conditions was likely due to CI processing transmitting degraded bottom-up cues. The 
latter was concluded because when NH listeners were presented vocoded speech, which 
simulates aspects of CI processing, they struggled to show any restoration. Another finding 
from the study was that there was considerable variation in restoration benefit across the 
CI users tested. In a more difficult condition (i.e., when less speech information was 
provided between interruptions), restoration benefits ranged from ?10 to +20 RAUs in CI 
users, with negative values indicating that noise-burst interrupted speech created 
interference rather than a benefit. In this case, six of 13 CI users (46%) showed an 
interference effect rather than a benefit. In the easier condition (i.e., when more speech 
information was provided between interruptions), restoration benefits ranged from ?5 to 
116 
 
 
 
approximately +20 RAUs in CI users, with only two of the 13 CI users (15%) showing 
interference effects. The authors concluded that the determination of who showed a 
positive restoration benefit was driven by intact, non-interrupted sentence understanding 
scores in quiet, in that higher sentence understanding scores were associated with positive 
restoration benefits, at least in the more difficult listening condition. Reasons for this 
included that higher-performing CI users may be able to make better use of speech 
information and/or may be more sensitive to acoustic speech cues in general.  
 We had several goals for the present replication study. First, we aimed to replicate 
the methods used in the Bhargava study as closely as possible in order to measure 
restoration in CI users. Second, we aimed to vary specific parameters used in the Bhargava 
study to see whether they influenced restoration and could explain the differences in results 
between the original study and our Experiments 1-3. Third, we aimed to measure whether 
intact speech understanding scores were positively correlated with restoration benefits.  
   
5.1 Methods 
5.1.1 Participants 
 We recruited the same number of participants used in the original Bhargava study, 
n=13. Table 10, modified from Bhargava?s study, describes their participants. Per the 
Bhargava paper, ?N/A? for the column ?Age at onset of hearing loss (years)? indicates that 
?readings were not available in the patient record.? 
 
 
 
117 
 
 
 
Table 10. Participant information table adapted from Bhargava et al. (2014) describing the participants in 
their perceptual restoration study. AB = Advanced Bionics. 
 
Intact CVC 
Age at Age at onset 
Speech Phoneme 
Participant Gender Testing of hearing CI Brand 
Scores (in Score % 
(years) loss (years) 
RAUs) 
CI 1 F 28 N/A Cochlear 102.3 75 
CI 2 M 38 3 Cochlear 99.8 85 
CI 3 F 22 N/A Cochlear 88.2 94 
CI 4 M 23 N/A Cochlear 93.0 82 
CI 5 F 65 30 Cochlear 117.8 94 
CI 6 M 52 33 Cochlear 96.4 95 
CI 7 M 65 61 Cochlear 93.5 80 
CI 8 F 62 45 Cochlear 98.0 80 
CI 9 M 64 N/A Cochlear 95.6 85 
CI 10 F 57 N/A Cochlear 104.4 85 
CI 11 F 65 N/A Cochlear 81.1 91 
CI 12 M 35 1 Cochlear 117.8 90 
CI 13 M 55 0 AB 78.4 67 
Average  48.5 24.7  97.4 84.8 
St. Dev.  16.9 24  11.8 9.2 
Range  22 - 65 0 - 61  78.4 ? 117.8 67 ? 95 
 
 The next table (Table 11) details the participant variables for our replication study. 
On average, compared to the original study, the current study had older participants 
(average age of 60.4 years compared to 48.5 years), with lower baseline, intact sentence 
understanding scores (average of 83.2 RAUs compared to 97.4 RAUs) and lower CNC 
phoneme scores (average of 80.9% compared to 84.8%). It is unclear whether our sample 
had earlier ages of onsets of hearing loss (average onset of 12.3 years compared to 24.7 
years), because onsets were reported for only seven of 13 participants in the original study.  
It is also unclear how the Bhargava study measured age at onset of hearing loss; the values 
118 
 
 
 
in Table 11 indicate the earliest age at which non-normal hearing was noticed in at least 
one of the two ears.  
 Note that the participants in the Bhargava study spoke Dutch as their native 
language and completed the study in Dutch. In contrast, our participants used English as 
their native language and completed the study in English.  
 
Table 11. Participant information from our replication study. AB = Advanced Bionics. 
 
Age at 
onset of Intact CNC 
Age at CNC 
non- CI Speech Word 
Participant Gender testing Phoneme 
normal Brand Scores (in Score 
(years) Score (%) 
hearing RAUs) (%) 
(years) 
CAD M 79 55 Cochlear 111.4 86 95 
CAF F 72 3 Cochlear 80.6 56 74 
CAQ F 61 34 Cochlear 84.7 60 76 
CBH F 66 0 Cochlear 64.8 68 86 
CBR F 66 0 Cochlear 83.2 54 75 
CCA M 79 13 Cochlear 71.9 64 81 
CCO F 74 2 Cochlear 107.2 82 92 
CCR F 70 2 Cochlear 84.5 64 82 
CCS M 43 1 Cochlear 77.2 54 77 
CDQ F 51 3 Cochlear 97.5 64 82 
CEI F 31 1.33 Cochlear 76.8 76 86 
CEN F 45 12 AB 41.0 24 51 
CES M 48 30 Cochlear 101.1 86 95 
Average  60.4 12.0  83.2 64.5 80.9 
St. Dev.  15.3 17.2  18.8 16.7 11.5 
Range  31 - 79 0 - 55  41.0 - 111.4 24 ? 86 51 - 95 
 
119 
 
 
 
5.1.2 Stimuli 
 The original study used sentences from the Vrij University (VU) corpus (Versfeld 
et al., 2000), which were meaningful Dutch sentences. An example of one of the sentences 
is (translated into English) ?Outside it is dark and cold.? The original study drew from 38 
lists produced by a male talker, with each list containing 13 sentences, for an average of 80 
words per list. One list was used for familiarization purposes, two lists were used for 
measuring the baseline intact sentence understanding scores, and 20 lists were used for 
measuring each of the ten conditions (described below), with 2 lists randomized to each 
condition.  
 Our study used sentences from the IEEE corpus (Rothauser et al., 1969). An 
example of one of the sentences is ?The birch canoe slid on the smooth planks.? The 
sentences were recorded by a male talker, and each list contained 10 sentences. Somewhat 
similar to the original study, we used one list for familiarization purposes, two lists for 
measuring baseline intact sentence understanding scores, and 24 lists for measuring each 
of the twelve conditions (described below), with 2 lists randomized to each condition. 
 The authors of the original study processed their sentences in the following way. 
Two interruption types were applied to the sentences: silent-gap interruptions and noise-
burst interruptions. Both interruption types were periodic square waves, with an 
interruption rate of 1.5 Hz and raised 5-ms cosine ramps. The duty cycle was 50% or 75%. 
For the 50% duty cycle condition, 333 ms of speech was available in every 666 ms-long 
speech segment. For the 75% duty cycle condition, 500 ms of speech was available in every 
666 ms-long speech segment. The SNR of noise-burst interruptions was varied as well: 
restoration was measured at ?10, ?5, 0, and 5 dB SNRs. Thus, the ten tested conditions 
120 
 
 
 
were: (1) silent-gap interruptions with a 50% duty cycle, (2) silent-gap interruptions with 
a 75% duty cycle, (3-6) noise-burst interruptions with a 50% duty cycle at ?10, ?5, 0, and 
5 dB SNRs, and (7-10) noise-burst interruptions with a 75% duty cycle at ?10, ?5, 0, and 
5 dB SNRs. Speech was always presented at 60 dB SPL. 
 For the replication study, we tested two interruption types, silent gaps and noise 
bursts, in order to measure the restoration effect. We additionally tested two interruption 
rates, 1.5 Hz (used in the original study), and 5 Hz (used in Experiments 1-3), and three 
duty cycles, 65%, 75% (used in the original study), and 85% (similar to the 80% duty cycle 
used in Experiments 1-3). All combinations of these parameters were tested, resulting in 
12 total conditions (two interruption types ? two interruption rates ? three duty cycles). 
Each condition was tested twice, each time with a new sentence list of 10 sentences. Thus, 
the experiment was composed of 240 sentence trials. Order of conditions was randomized, 
and sentence list allocation to conditions was randomized. For our experiment, the SNR 
was fixed at ?5 dB; in the original study, no effect of SNR was found in CI users, so we 
did not choose to vary this factor. We presented speech at the same intensity level as the 
Bhargava study.  
 
5.1.3 Procedure 
 In the original study, participants sat in a soundbooth in front of a computer monitor 
and loudspeaker located at 0 degrees. Participants first completed the baseline speech 
understanding conditions, then familiarization, then the experiment. Familiarization 
consisted of participants being presented sentences that were processed with a different 
interruption rate, duty cycle, and SNR than those used in the main experiment. Participants 
121 
 
 
 
listened to each sentence during familiarization and repeated aloud what was heard. 
Feedback during familiarization was an auditory presentation of the intact sentence and a 
visual display of the text of the sentence. Following completion of familiarization, the main 
experiment began. 
 In our replication study, participants sat in a soundbooth in front of a pair of 
loudspeakers located at ?45 degrees. The computer screen faced the experimenter, who 
controlled the experiment and was in the soundbooth with the participant for the duration 
of the experiment. Participants were tested on baseline sentences, then familiarized with 
the stimuli, then completed the main experiment. Familiarization consisted of presenting 
participants with stimuli that were processed slightly differently from stimuli in the main 
experiment: stimuli were presented with a 1.5-Hz interruption rate (used in the main 
experiment) and 80% duty cycle (not used in the main experiment) at ?10-dB SNR (not 
used in the main experiment). The procedure and feedback for familiarization were 
identical to those used in the original study. 
In the original study, participants were scored on the number of words they 
correctly reported per list. The percent-correct scores were converted into RAUs. No 
feedback was provided to participants, and incorrect/absent scores were not penalized. In 
the replication study, sentences were graded for number of words correctly reported, using 
lax grading that accepted as correct some changes in tense (?shop? for ?shopped? but not 
?went? for ?go?) and changes in plurality (?cats? for ?cat? or ?cup? for ?cups?), in line 
with previous research practices (Jaekel et al., 2018). Similar to the original study, we 
transformed percent-correct scores into RAUs, no feedback was provided, and incorrect 
answers were not penalized. 
122 
 
 
 
5.2 Results 
 Prior to describing our own results from the replication study, we describe results 
from the original study by Bhargava et al. (2014). In the most difficult listening condition 
(50% duty cycle), CI users showed no restoration effects at any SNR; that is, performance 
with noise-burst interrupted speech was never significantly higher (or lower) than 
performance with silent-gap interrupted speech. In the less difficult listening condition 
(75% duty cycle), CI users showed significant restoration effects at every SNR; that is, 
performance with noise-burst interrupted speech was significantly higher than performance 
with silent-gap interrupted speech, indicating a restoration benefit. When examining how 
participant variables might explain these results, the original study plotted restoration 
benefits against baseline speech scores. Higher baseline scores were associated with more 
positive restoration benefits in the 50% duty cycle condition (reported as r = 0.71, p < 
0.05), while there was no relationship in the 75% duty cycle condition (reported as r = 0.13, 
p = 0.67). In summary, we should expect to observe a restoration effect in our replicating 
condition (interruption rate = 1.5 Hz, duty cycle = 75%, and SNR = ?5 dB). In this 
condition, Bhargava and colleagues (2014) found an approximate +5.6 RAU benefit for 
restoration, or approximately a benefit of 8.8%. Furthermore, we can expect that 
performance in this condition may not correlate with baseline speech understanding scores, 
but that performance in decidedly more difficult conditions (e.g., the 65% duty cycle 
conditions) may correlate positively with baseline speech understanding scores.  
 Our findings are described below. First, we compare our replicating condition with 
that in the original study, and then discuss results from the other test conditions. Compared 
to the +5.6 RAU benefit for restoration in the 1.5-Hz interruption rate, 75% duty cycle, ?5-
123 
 
 
 
dB SNR condition found by Bhargava and colleagues, we found a ?0.9 RAU interference 
effect (SE = 3.6, range = ?18.9 to +19.7). Seven participants showed restoration on average 
(+3.6, +3.6, +3.8, +4.3, +10.2, +17.4, +19.7 RAUs). Six participants showed interference 
(?18.9, ?17.9, ?16.7, ?13.6, ?4.2, ?3.1 RAUs). Thus, we were unable to find similar 
average restoration effects to the Bhargava study with the same parameters and largely 
identical procedures. In terms of accuracy in the replicating condition, the original study 
showed approximately 50 RAUs correct for the silent-gap interrupted condition, and 
approximately 55 RAUs correct for the noise-burst interrupted condition. In contrast, our 
average RAUs correct in the replicating condition for silent-gap interrupted speech was 
45.2, and our average RAUs correct for noise-burst interrupted speech was 44.2, both lower 
than that observed in the original study. Performance in these two conditions in our study 
was not statistically significantly different, per a paired-samples t test: t(12) = 0.26, 
p=0.802. 
Perhaps a certain level of performance with silent-gap interruptions is necessary to 
achieve restoration benefits; because our sample?s silent-gap interrupted speech 
understanding was lower than that seen in the original study?s, we did not see restoration 
benefits. However, a closer analysis of the individual data did not confirm this notion?in 
the replicating condition, the range of performance with silent-gap interrupted speech was 
similar for participants who achieved a restoration benefit (18 to 72 RAUs) and for those 
who experienced interference (22 to 78 RAUs).  
 Average restoration benefits and accuracy in RAUs for all of our study conditions 
are reported in Table 12. Restoration benefits were observed in two of the six conditions 
tested: the 1.5-Hz interruption rate with a 65% duty cycle and the 5-Hz interruption rate 
124 
 
 
 
with an 85% duty cycle. The highest accuracy was in the 5-Hz interruption rate with an 
85% duty cycle, and the lowest accuracy was in the 5-Hz interruption rate with a 65% duty 
cycle. The eight ?prelingual? participants ? that is, participants who experienced non-
normal hearing in at least one ear at < 3 years of age, before language development ? 
showed the most restoration benefit, with an average of +3.2 RAUs restoration effect across 
conditions. The five ?postlingual? participants ? those who experienced non-normal 
hearing after completing language development ? showed the most interference, with an 
average of ?6.5 RAUs restoration effect across conditions.   
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
125 
 
 
 
Table 12. Average restoration benefits and accuracy in RAUs for our replication study. The replicating 
condition is underlined. SG = silent gaps. NB = noise bursts. 
 
Average 
Condition Restoration SE Range 
Benefit (RAUs) 
1.5-Hz interruption rate, 65% duty cycle 
2.4 2.8 ?11.6 to 17.7 
(433.3 ms of speech per 666.6-ms segment) 
1.5-Hz interruption rate, 75% duty cycle 
?0.9 3.6 ?18.9 to 19.7 
(499.9 ms of speech per 666.6-ms segment) 
1.5-Hz interruption rate, 85% duty cycle 
?1.0 3.0 ?24.9 to 14.9 
(566.6 ms of speech per 666.6-ms segment) 
5-Hz interruption rate, 65% duty cycle 
?4.3 3.2 ?24.9 to 13.7 
(130 ms of speech per 200-ms segment) 
5-Hz interruption rate, 75% duty cycle 
2.6 3.1 ?13.6 to 20.1 
(150 ms of speech per 200-ms segment) 
5-Hz interruption rate, 85% duty cycle 
?2.1 3.4 ?21.8 to 12.1 
(170 ms of speech per 200-ms segment) 
 
   
Accuracy 
Condition SE Range 
(RAUs) 
   
 
1.5-Hz interruption rate, 65% duty cycle, SG 26.0 4.6 4.2 to 56.9 
1.5-Hz interruption rate, 65% duty cycle, NB 28.4 5.6 ?2.8 to 68.8 
1.5-Hz interruption rate, 75% duty cycle, SG 45.1 5.6 18.0 to 78.1 
1.5-Hz interruption rate, 75% duty cycle, NB 44.2 7.6 4.8 to 91.5 
1.5-Hz interruption rate, 85% duty cycle, SG 63.4 6.4 39.3 to 100.3 
1.5-Hz interruption rate, 85% duty cycle, NB 62.4 7.5 14.4 to 103.5 
5-Hz interruption rate, 65% duty cycle, SG 29.3 7.2 ?8.1 to 66.3 
5-Hz interruption rate, 65% duty cycle, NB 24.9 7.9 ?13.0 to 71.2 
5-Hz interruption rate, 75% duty cycle, SG 42.9 8.4 ?2.4 to 96.8 
5-Hz interruption rate, 75% duty cycle, NB 45.5 8.3 5.6 to 83.2 
5-Hz interruption rate, 85% duty cycle, SG 64.4 7.8 16.9 to 105.4 
5-Hz interruption rate, 85% duty cycle, NB 62.3 7.7 19.4 to 108.4 
  
 We analyzed accuracy using a repeated measures ANOVA. Unfortunately, we 
could not repeat the original study?s statistical analysis with our own data (Dunnett?s test 
for multiple comparisons). Bhargava et al. (2014) used the technique to compare several 
126 
 
 
 
noise-burst conditions to a single silent-gap condition, whereas each of our noise-burst 
conditions had a partner silent-gap condition against which it could be compared. 
 The variables of interruption type (two levels, noise-burst and silent-gap 
interruptions), interruption rate (two levels, 1.5 Hz and 5 Hz), and duty cycle (three levels, 
65%, 75%, and 85%) were entered as the independent variables to a RM ANOVA. The 
dependent variable was accuracy in RAUs. No variables violated Mauchly?s Test of 
Sphericity, so sphericity was assumed and no corrections were used. 
 The only significant main effect was duty cycle, F(2,24)=156.8, p<0.001, ?2?=0.93. 
The main effect of interruption type was not significant, F(1,12)=0.08, p=0.78, ?2?=0.01, 
nor was the main effect of interruption rate, F(1,12)<0.001, p=0.99, ?2?<0.001. No 
interactions were significant: interruption type ? interruption rate, F(1,12)=0.44, p=0.52, 
?2?=0.04; interruption type ? duty cycle, F(2,24)=0.36, p=0.70, ?
2
?=0.03; interruption rate 
? duty cycle, F(2,24)=0.04, p=0.96, ?2?=0.003; three way interaction, F(2,24)=1.76, 
p=0.19, ?2?=0.13.  
 For the main effect of duty cycle, post-hoc paired-samples t-tests Bonferroni-
corrected for multiple comparisons revealed that performance in every duty cycle was 
significantly different from every other (all p?s <0.001). The best performance was in the 
85% duty cycle, followed by the 75% duty cycle, followed by the 65% duty cycle.  
 To conclude, interruption type did not affect accuracy: similar scores were obtained 
with silent-gap interrupted sentences and noise-burst interrupted sentences. Thus, we could 
not confirm the presence of a restoration effect. Because no interactions with interruption 
type were significant, no combination of parameters (duty cycles and interruption rates) 
explained interrupted speech understanding. Interruption rate did not affect accuracy either; 
127 
 
 
 
similar scores were obtained regardless of how frequent interruptions appeared (i.e., every 
666.6 ms vs. every 200 ms). In contrast, longer duty cycles, that is, greater access to speech 
information across segments, did lead to significantly improved performance. The original 
study found a similar effect, in that performance was greater in the 75% duty cycle 
condition compared to the 50% duty cycle condition. Thus, restoration, defined as a 
significant difference between silent-gap and noise-burst conditions, was not observed via 
this analysis.  
 
5.2.1 Baseline speech understanding and its correlation with restoration benefits 
 This analysis matches one used in the original study, which correlated restoration 
benefits in each duty cycle condition (in RAUs) with baseline sentence understanding 
scores (in RAUs). We correlated baseline sentence understanding scores (in RAUs) with 
restoration benefits in each duty cycle ? interruption rate condition (in RAUs), for a total 
of six correlation analyses. The results are presented in Table 13. No significant 
correlations were found in any of the conditions tested. 
 
Table 13. Pearson correlations between restoration benefits at each interruption rate ? duty cycle 
combination with baseline speech understanding scores. No correlations were significant. 
 
1.5 Hz, 1.5 Hz, 1.5 Hz, 5 Hz, 5 Hz, 5 Hz, 
65% duty 75% duty 85% duty 65% duty 75% duty 85% duty 
  cycle cycle cycle cycle cycle cycle 
r (n=13) 0.31 0.36 0.49 0.12 0.15 ?0.10 
p 0.30 0.23 0.09 0.70 0.63 0.75 
 
128 
 
 
 
5.2.2 Are the IEEE sentences ?restorable?? 
 One possibility is that our IEEE sentences, unlike the VU sentences utilized in the 
original study, are not restorable at the parameters tested, even for NH listeners. This could 
explain why we did not see consistent restoration across conditions and especially why we 
did not replicate restoration in the replicating condition. To explore this, we tested one NH 
adult listener (age = 53 years) with the same stimuli processed by a 32-channel noise 
vocoder. Pilot testing revealed that unprocessed versions of the stimuli presented to 
separate NH listeners resulted in ceiling performance that reflected no restoration benefits 
(this issue was also observed in the original Bhargava study, for their NH unprocessed 
condition). A vocoder can mimic aspects of CI processing, degrading the stimuli enough 
to bring performance down from the ceiling. The decision to utilize 32 channels was based 
on findings from Jaekel et al. (2018); while 32 channels is higher spectral resolution than 
one would expect in real CI users (Berg et al., 2019; Croghan et al., 2017; Friesen et al., 
2001), it allows NH listeners to complete restoration tasks with vocoded stimuli without 
experiencing floor effects.  
 Our NH adult listener showed high performance on average across conditions (87.3 
RAUs). The worst performance, 60.4 RAUs, was observed with silent gaps in the 1.5-Hz 
interruption rate, 65% duty cycle condition. The best performance, 106.9 RAUs, was 
observed with noise bursts in the 5-Hz interruption rate, 85% duty cycle condition. 
Perceptual restoration was observed at every interruption rate ? duty cycle combination, 
with an average restoration benefit of +6.8 RAUs (ranging from +1.4 to +17.2 RAUs). The 
greatest amount of restoration, +17.2 RAUs, was observed in the 5-Hz interruption rate, 
85% duty cycle condition (the condition most similar to the parameters used in 
129 
 
 
 
Experiments 1-3). At the replicating condition (1.5-Hz interruption rate, 75% duty cycle), 
a restoration benefit of +6.2 RAUs was observed.  
 Thus, it appears that the IEEE sentences used in our replicating study were 
restorable across conditions. Note that this NH participant was not tested as a direct 
comparison to our CI user participants, but as a way to determine whether our sentences 
were restorable. The 32-channel noise vocoder used provides higher spectral resolution 
than a CI user could be expected to receive, and does not fully capture the listening 
experiences of CI users.  
 
5.3 Discussion 
 Experiment 4 had three goals: (1) replicate the original CI restoration study by 
Bhargava et al. (2014); (2) vary parameters that could impact restoration and assess their 
effects; and (3) measure whether intact speech understanding scores correlated with 
restoration benefits, which was a finding in the original paper. One additional goal was 
also pursued: measuring whether the sentence materials we used were ?restorable? by 
presenting them to a NH listener. 
 The study by Bhargava et al. (2014) found a +5.6 RAU (approximately 8.8%) 
restoration benefit in the replicating condition. In contrast, we found a ?0.9 RAU 
(approximately ?0.5%) restoration effect in this condition. While the participants in the 
original study found the noise-burst interruptions helpful for understanding speech in this 
condition, our participants, on average, did not show this benefit, with seven participants 
showing positive restoration and six participants showing negative restoration effects. 
Analyzed statistically, no significant difference in performance in this replicating condition 
130 
 
 
 
for the silent-gap versus the noise-burst interrupted sentences was observed. Thus, we were 
unable to confirm the findings reported by Bhargava et al. (2014) ? CI users showed no 
restoration benefits, on average, in this condition. This finding aligns with findings from 
Experiments 1-3, where participants similarly showed no advantage for noise-burst 
interruptions and instead tended to show an advantage for silent-gap interruptions, the 
opposite of the restoration effect.  
 The second goal of our study was to measure how varying the parameters of duty 
cycle and interruption rate affected restoration in CI users. For duty cycle, we tested 65, 
75, and 85% cycles. For interruption rate, we tested 1.5- and 5-Hz rates. Our RM ANOVA 
analysis, which we admit was underpowered, revealed no effects of interruption type or 
interruption rate. We did find a significant main effect of duty cycle, which indicated that 
increasing the amounts of speech available to the listener within each speech segment 
resulted in better performance overall. However, duty cycle did not interact with 
interruption type, meaning this usefulness of greater amounts of speech information did 
not boost restoration. No interactions were found to be significant with this analysis.  
 The third goal was to measure whether intact baseline sentence understanding 
scores were correlated with restoration benefits; this was found for a 50% duty cycle 
condition in the original paper (untested here), but not for the 75% duty cycle condition in 
the original paper (tested here). Like the original study, we found no significant correlation 
between performance in the replicating condition and baseline speech understanding. In 
addition, we did not find significant correlations between performance in any of the tested 
conditions and baseline speech understanding, even those closer to the 50% duty cycle 
tested in the original paper (i.e., the 65% duty cycle condition with 1.5-Hz interruptions). 
131 
 
 
 
On the one hand, perhaps the resources used in difficult listening conditions like a 50% 
duty cycle are not marshalled for the easier 65% duty cycle condition. On the other hand, 
the closest our correlations? p-values came to our alpha level was for the 1.5-Hz 
interruption rate, 85% duty cycle condition, which would not be predicted to show any 
effect of baseline speech understanding scores per the original paper.  
 A final goal for this study grew out of concerns that perhaps the IEEE sentences 
simply were not ?restorable? given the parameters tested. That is, perhaps CI users were 
not, on average, showing restoration effects because even a NH listener with clear bottom-
up acoustic information and comparable top-down knowledge would fail to show 
restoration effects with these materials. IEEE sentences were shown to be restorable in 
Jaekel et al. (2018), but at different parameters than those used in the present experiment. 
To measure IEEE sentence restorability with our current parameters, we initially tested two 
NH adult listeners on unprocessed versions of the stimuli used in this experiment. 
Unfortunately, the two NH listeners performed at ceiling in almost every condition, 
meaning no restoration effect was observed (that is, there was no ?room to grow? from the 
silent-gap interrupted conditions, since participants were performing so well already in 
such conditions). Therefore, we tested a different NH adult listener with 32-channel noise-
vocoded versions of the sentences. This adult listener was comparable in age to the CI users 
in both the original and replication studies (age = 53.9 years). This listener did not perform 
at ceiling across conditions, and thus was able to show a restoration effect in every 
condition tested. We therefore concluded that the sentence material was restorable with our 
parameters even with some acoustic degradation to the bottom-up acoustic cues; however, 
132 
 
 
 
clearly, the vocoder fails to capture the full CI listening experience, as many of our CI users 
failed to show restoration.  
 Our replication of the original CI restoration paper was not perfect; first, our sample 
was slightly older and had poorer baseline speech understanding scores compared to the 
sample in the original paper. Second, we tested some conditions that were not tested in the 
original paper, which may have affected how participants performed on the replicating 
condition, as the task was slightly different. Third, our participants were native English 
speakers completing the study in English, while the original study was performed with 
Dutch speakers in Dutch. This necessitated the use of a different sentence corpus than that 
used in the original study, which could have differed in other ways from the original corpus 
(e.g., our sentences could have been harder to comprehend in some manner). Fourth, there 
were several other small differences between our study and the original study that may 
have had an effect on performance, for example using two loudspeakers rather than one, 
having the experimenter control the experiment rather than it being self-paced by the 
participant, etc. Overall, however, we believe we have come close to repeating the target 
condition used in the original study, and thus our results can be compared to the original 
study?s. 
 To summarize, the lack of restoration effects observed in Experiments 1-3 is less 
concerning given the results of Experiment 4, where we failed to replicate a previously 
reported finding by Bhargava et al. (2014) that CI users should be able to restore speech. 
We did find considerable variation in our participants in terms of restoration, just as was 
found in the original study, but this variation did not lead to an overall average effect of 
positive restoration benefits. Furthermore, the parameters used to measure restoration did 
133 
 
 
 
not clearly elicit patterns in results: for example, one might expect higher duty cycles to 
result in more restoration (Bhargava et al., 2014), but this was not always the case; 
sometimes less frequent interruptions resulted in restoration, while other times more 
frequent interruptions resulted in restoration. In general, our results suggest that individuals 
with CIs do not show consistent perceptual restoration ? both within and across 
participants. This has important implications for real world listening in noise, which will 
be addressed in the conclusions chapter.  
134 
 
 
 
Chapter 6: Conclusion 
CI users are less successful at understanding speech in noisy listening environments 
compared to NH listeners (Fetterman & Domico, 2002). While noise reduction algorithms 
in CI front-end preprocessing generally appear to improve speech understanding in noise 
(Gifford & Revit, 2010; Mauger et al., 2014; Wolfe et al., 2015),  one question is whether 
these algorithms, along with other CI-related variables, might be negatively affecting the 
use of a speech repair mechanism called perceptual restoration (Warren, 1970). Removing 
noise from a speech signal interruption reduces, rather than improves, understanding; the 
presence of noise in a speech signal interruption can cause the listener to perceive the 
interrupted word as intact and the noise as a separate, irrelevant auditory object. This 
illusion of intactness results in better speech understanding, as the interrupted speech has 
been ?repaired,? or perceptually restored, by the brain. CI users sometimes fail to show 
perceptual restoration in conditions where NH listeners can take advantage of the 
mechanism (Bhargava et al., 2014). This dissertation aimed to discover the extent to which 
device and listener factors reduced or improved perceptual restoration in CI users. 
Strengthening perceptual restoration in CI users could improve their ability to understand 
speech in noisy, realistic listening environments.     
In order to achieve perceptual restoration, three main elements are needed: (1) the 
presence of a plausible masker (Bashford et al., 1992; Ba?kent, 2012); (2) access to high-
quality bottom-up acoustic information, for example, so that maskers may be adequately 
distinguished from the target signal (Bashford et al., 1992; Ba?kent, 2012; Bhargava et al., 
2014; Jaekel et al., 2018); and (3) opportunities to use top-down linguistic knowledge 
(Ba?kent et al., 2016; Bregman, 1990; Samuel, 1981; Shinn-Cunningham & Wang, 2008). 
135 
 
 
 
The presence of a plausible masker is potentially affected by noise reduction algorithms in 
CI front-end preprocessing (Experiment 1) and compression, which changes the effective 
SNR (Experiment 2). The quality of bottom-up acoustic information may vary in bilateral 
CI users who have a ?poorer? and ?better? ear ? meaning their ears experience different 
levels of signal degradation, potentially due to factors like duration of hearing loss and 
electrode/nerve interface (Experiment 3). Providing semantic cues may increase 
opportunities for CI users to apply top-down knowledge to incoming signals, but this could 
be less possible in CI users who experienced hearing loss prior to spoken language 
development (Experiment 3). Thus, this dissertation aimed to evaluate CI users? restoration 
abilities across all three main elements. 
 
6.1 Device factors affecting perceptual restoration in CI users 
 We hypothesized that both noise reduction algorithms in CI front-end 
preprocessing and compression algorithms would reduce restoration in CI users, as they 
were expected to affect the plausibility of interrupting noise serving as a masker. Although 
such algorithms may boost speech understanding in noise in general, increasing 
opportunities to utilize restoration in noisy environments could result in improved 
outcomes.  
 Experiment 1 evaluated whether disabling CI front-end preprocessing features 
improved performance in the restoration paradigm. Despite choosing parameters that have 
elicited restoration in CI users in our own pilot testing and in previous work in the area 
(Bhargava et al., 2014), we found no restoration either with or without CI front-end 
preprocessing. Thus, we could not make any conclusions as to whether restoration was 
136 
 
 
 
affected by noise-reduction algorithms. While we cannot speak to restoration benefits, we 
found that the availability of CI front-end preprocessing did particularly improve noise-
burst interrupted speech understanding. Therefore, the noise-reduction algorithms served 
their purpose: making it easier for CI users to perceive noise-interrupted speech signals, 
potentially via reducing the signal distortions introduced by loud, sudden noise bursts.   
 Experiment 2 evaluated whether compression reduced perceptual restoration in CI 
users. The level of incoming signals is decreased by a compression algorithm above a 
certain intensity level; we wanted to determine if the levels of our interrupting noises were 
being reduced by compression, thus reducing their plausibility as maskers in the restoration 
task. We presented stimuli at levels that either would or would not engage compression, 
with the expectation that when compression was engaged, restoration would be reduced. 
As in Experiment 1, we found no restoration at either sound level, and thus were unable to 
determine whether compression affected restoration ability. Speech understanding with 
noise-burst interrupted speech was best in the most intense condition, when compression 
was expected to be engaged. Noise-interrupted speech was possibly easier to process in the 
more intense condition because the effective SNR was changed to be more favorable; 
however, again, we found no evidence that CI users were able to utilize noise bursts to 
perform speech repair. 
 To summarize, we were unable to conclude whether device factors negatively 
affected restoration, as we were unable to detect restoration in CI users in general. CI users 
do not appear to perceive noise interruptions as prompters of restoration, but as interferers 
to speech processing. Previous work has found that CI users may qualitatively perceive 
noise differently from NH listeners (Oxenham & Kreft, 2014), so perhaps their approach 
137 
 
 
 
to processing speech in the restoration paradigm is also qualitatively different from that of 
NH listeners. For example, current spread is known to reduce CI users? utilization of 
temporal envelope modulations during speech processing (Oxenham & Kreft, 2014); it is 
possible that the temporal changes introduced by silent-gap interruptions were less 
detectable to CI users, and thus had less of an impact on speech understanding. In 
comparison, noise-burst interruptions could have introduced forward and backward 
masking that affected speech signals surrounding the burst and/or introduced auditory 
distraction, thus reducing speech understanding. Overall, noise-reduction algorithms and 
compression appeared to work as intended ? improving speech-in-noise understanding in 
general ? but whether they allow opportunities to use speech repair, which could further 
improve communication outcomes in noisy environments, remains an open question.  
 
6.2 Listener factors affecting perceptual restoration in CI users 
 Experiment 3 evaluated how the factors of bottom-up acoustic signal quality and 
opportunities to use top-down knowledge affected restoration in CI users. To test the 
former, we presented interrupted sentences to either the ?better? (assumed higher signal 
quality) or ?poorer? (assumed lower signal quality) ears. To test the latter, we provided a 
semantic cue prior to half of the sentences, which was meant to prime listeners to the 
content of the upcoming sentence. While some CI users showed restoration benefits in 
certain conditions of Experiment 3, no pattern emerged and our statistical model failed to 
detect average restoration benefits. Instead, we found that top-down knowledge was able 
to interact with any quality of bottom-up acoustic signal to produce an increase in speech 
understanding, and that having presumably higher-quality bottom-up acoustic information 
138 
 
 
 
did not boost noise-burst interrupted speech understanding, contrary to previous restoration 
work with degraded signals (Ba?kent, 2012; Clarke et al., 2016; Jaekel et al., 2018).  
 When CI users were grouped based on hearing histories and other participant 
variables, we found that older CI users with asymmetrical ears could achieve restoration 
benefits, while CI users with symmetrical ears did not. The asymmetry in hearing histories 
across ears itself may not be relevant; rather, the ability to repair speech may be due to the 
fact that these listeners had the longest exposure to normal hearing in one ear out of all 
participants in the study. While this group showed particular benefits with processing 
interrupted speech in general when they could utilize top-down knowledge and received 
high-quality bottom-up acoustic signals, their restoration ability was not dependent on 
these factors. Therefore, in CI users, it may be the case that extensive experience repairing 
speech signals with a NH ear may be necessary to achieve restoration with the device. Note 
that this finding was based on data from only a few participants, meaning the majority of 
the CI users we tested did not show any restoration. This again points to a possible 
qualitative difference in how CI users, in general, approach the processing of noise-burst 
interrupted signals compared to NH listeners.  
 
6.2.1 Restoration in prelingual vs. postlingual adult cochlear implant users 
Some participants in our studies experienced non-normal hearing prior to language 
development, meaning they had  ?prelingual? hearing loss and/or deafness. This may have 
impacted how these CI users were able to use restoration, a skill requiring the use of top-
down linguistic knowledge. With the implementation of universal newborn hearing 
screenings, hearing loss is now often detected much earlier in life than in the past, and with 
139 
 
 
 
the invention of CIs and FDA allowances for early implantation, children with severe 
hearing losses are receiving auditory input at much younger ages. However, we tested 
mainly older adults who were less able to take advantage of early detection and early 
implantation; our prelingual adults had long durations of non-normal hearing and had to, 
as children, develop language without clear auditory input (Pisoni, 2000). This 
development would include learning how to restore interrupted speech signals, as 
perceptual restoration is a skill that is believed to be developed over time rather than innate. 
Specifically, the restoration skill appears to be absent in NH toddlers (Newman, 2006) and 
developed (or in the midst of being developed) by 4-6 years in NH children (Ackroff, 1981; 
Koroleva, Kashina, Sakhnovskaya, & Shurgaya, 1991; Newman, 2004; Walley, 1988; 
Winstone, Davis, & de Bruyn, 2012). 
One hypothesis is that prelingual CI users would not show any restoration: having 
less robust lexical representations containing less acoustic-phonetic detail (due to never 
having access to clear speech) could make it more difficult to match incoming interrupted 
speech to stored representations during speech repair (R?nnberg, Rudner, Foo, & Lunner, 
2008). This could lead to prelingual CI users relying more on bottom-up acoustic 
information instead of utilizing top-down linguistic knowledge to restore speech. In fact, 
our prelingual CI users did restore speech in certain cases ? for example, in Experiment 4, 
several prelingual CI users showed restoration benefits. Therefore, it appears that 
prelingual CI users can develop restoration skills even without rich auditory input. Like 
their postlingual peers, however, they fail to consistently use this skill to repair noise-
interrupted speech (Experiments 1-3). While it is difficult to make strong conclusions based 
on such variable results, perceptual restoration development may be due to maturation 
140 
 
 
 
rather than exposure to and experience with auditory signals. As discussed below, future 
work could focus on the effects of auditory signal quality and quantity on perceptual 
restoration development in children with CIs.  
 
6.3 Replicating previous work in restoration 
 Prior to the present study, Bhargava et al. (2014) was the only published study to 
have directly tested restoration in CI users. We wanted to determine if we could replicate 
the restoration effects observed in some of the conditions in that original study, since we 
struggled to find consistent restoration effects in any of our experiments. Despite closely 
matching the methods used by Bhargava et al. (2014), we could not reproduce the positive 
restoration benefits they observed using their specific interruption parameters. Instead, we 
found restoration benefits in a condition with less frequent interruptions, but more speech 
information missing, and in a condition with more frequent interruptions and less speech 
information missing. Partly due to the large variability in performance in our sample, and 
partly due to the small size of these restoration benefits, no combination of parameters 
yielded statistically significant restoration benefits. Thus, once again, no consistent 
restoration was observed among the CI users tested ? matching the null findings found in 
our other experiments. Not only could we not elicit restoration in CI users by varying 
device factors and listener factors (Experiments 1-3), we could not elicit restoration by 
varying factors pertaining to the signal itself (Experiment 4).  
 
141 
 
 
 
6.4 Limitations 
 The present study was affected by some limitations. Despite pilot testing and prior 
work in the area (Bhargava et al., 2014), we were unable to set interruption parameters that 
allowed us to find a baseline restoration ability in our CI users. Because we could not elicit 
this baseline, we were unable to detect the impacts of noise-reduction algorithms and 
compression on restoration. Future work should consider administering a pre-test for each 
participant to determine the best interruption parameters to elicit restoration for that 
individual, if perceptual restoration in CI users is possible.  
 As front-end preprocessing algorithms in CIs are proprietary, there is little available 
information as to how the algorithms specifically work. While we expected to see effects 
of SNR-NR and other algorithms in the restoration paradigm, it was decidedly unclear how 
exactly the algorithms were changing the incoming speech signals. Knowing how the 
interrupted speech signals are affected by the noise-reduction algorithms and compression 
could allow us to create stimuli that can be more effectively manipulated to measure 
potential restoration effects. For example, if we knew the exact attack time for the SNR-
NR algorithm, we could interrupt our stimuli with noise bursts that have durations that 
consistently elicit attack time effects.   
 We had hoped that our sample of CI users would have clear ?poorer ear? and ?better 
ear? distinctions. In fact, our sample had quite symmetrical hearing histories, and often had 
little difference in speech understanding performance between their ears. To better 
determine the impact of bottom-up acoustic quality, a sample with more diverse, 
asymmetrical hearing histories should be recruited. Such a sample could include CI users 
with single-sided deafness or partial array insertions. CI users with single-sided deafness 
142 
 
 
 
have one ear with acoustic hearing and one ear with electric hearing; thus, the effects of 
bottom-up acoustic information quality on restoration, if any exist, would most likely be 
detected in this group of listeners. 
 
6.5 Future Directions 
 Studying the effects of listening effort and fatigue on perceptual restoration in CI 
users could elucidate the extent to which these factors might have affected performance, 
and contributed to the lack of restoration benefits, in the current project. Anecdotally, some 
CI users in the present study commented that the noise bursts were distracting and/or 
irritating. Others commented that the provision of semantic cues in Experiment 3 was 
sometimes frustrating (e.g., when they were unable to process the sentence, and felt like 
they should be able to because they had access to a ?hint?) and/or confusing (e.g., when 
they felt they had to process too many things at once ? holding a visually presented word 
in their memory while trying to listen to an interrupted sentence). The lack of restoration 
benefit could be because processing noisy sentences drains cognitive resources and 
requires greater effort than processing sentences with gaps (Finke, Sandmann, Kopp, 
Lenarz, & B?chner, 2015). The lack of benefit could also be due to a conscious choice on 
the part of the participants to refuse to evaluate difficult, noisy sentences in the first place, 
potentially due to fatigue. There exist several procedures for measuring listening effort in 
CI users, including pupillometry (Perreau, Wu, Tatge, Irwin, & Corts, 2017; Winn & 
Moore, 2018), dual-task paradigms (Hughes & Galvin, 2013), self-report, and ERP 
measures (Finke et al., 2016). Understanding the interaction of restoration and listening 
143 
 
 
 
effort might be helpful; it could be that in highly effortful listening situations, people 
demonstrate reduced restoration ability.   
 Perceptual restoration in CI users has only been studied in adults thus far; 
measuring perceptual restoration in children with CIs could reveal the extent to which the 
developmental time course of the mechanism is affected by exposure to degraded auditory 
input. By comparing the development of restoration in children with CIs to NH children, 
this line of work could determine whether children with CIs are delayed in learning to 
utilize top-down knowledge during spoken language processing in difficult listening 
environments. Poor signal quality could slow the speed with which children learn to 
integrate top-down and bottom-up information; alternatively, consistently degraded 
bottom-up acoustic information could encourage children to begin to rely more heavily on 
top-down knowledge at earlier ages. Furthermore, the FDA has approved that pediatric CI 
users over the age of six can have activated front-end preprocessing, meaning that the 
effects of noise-reduction algorithms on the development of restoration skills could also be 
evaluated.  
 
6.6 Final Conclusion 
 CI users fail to consistently repair noisy speech signals. Device factors like front-
end preprocessing and compression, which were believed to impact the plausibility of noise 
as a masker (and thus the plausibility of the restoration illusion), neither inhibited nor 
promoted restoration. Similarly, listener factors like bottom-up signal quality and top-down 
linguistic knowledge use, whose integration is believed to be key to repairing speech, did 
not produce a restoration effect. While the perceptual restoration framework appears to 
144 
 
 
 
function for NH listeners, the processing of noise-burst interrupted speech may be 
qualitatively different in CI users, and thus may require different parameters in order to 
function correctly. Noise generally serves more as an interferer rather than a promoter of 
speech repair in this population. Perceptual restoration, then, may not be a useful tool for 
CI users attempting to understand speech in noisy environments, and the inability to utilize 
restoration may be a contributor to this population?s general difficulties understanding 
speech in noise.  
  
145 
 
 
 
Bibliography 
 
AB?s Exclusive Sound Processing Technologies. (2018). Retrieved from 
https://advancedbionics.com/us/en/home/products/sound-processing.html 
Ackroff, J. M. (1981). The interrelationship of verbal transformations, phonemic 
restorations, and age Doctoral Dissertation. University of Wisconsin-Milwaukee. 
Milwaukee, Wisconsin.  
Anderson, S., Parbery-Clark, A., White-Schwoch, T., & Kraus, N. (2012). Aging affects 
neural precision of speech encoding. J Neurosci, 32(41), 14156-14164. 
doi:10.1523/JNEUROSCI.2176-12.2012 
Ardoint, M., Green, T., & Rosen, S. (2014). The intelligibility of interrupted speech 
depends upon its uninterrupted intelligibility. J Acoust Soc Am, 136(4), El275-
El280. doi:10.1121/1.4895096 
Arehart, K. H., Souza, P., Baca, R., & Kates, J. M. (2013). Working memory, age, and 
hearing loss: susceptibility to hearing aid distortion. Ear Hear, 34(3), 251-260. 
doi:10.1097/AUD.0b013e318271aa5e 
Balkany, T., Hodges, A., Menapace, C., Hazard, L., Driscoll, C., Gantz, B., . . . Payne, S. 
(2007). Nucleus Freedom North American clinical trial. Otolaryngol Head Neck 
Surg, 136(5), 757-762. doi:10.1016/j.otohns.2007.01.006 
Bashford, J. A., Riener, K. R., & Warren, R. M. (1992). Increasing the Intelligibility of 
Speech through Multiple Phonemic Restorations. Perception & Psychophysics, 
51(3), 211-217. doi:Doi 10.3758/Bf03212247 
Ba?kent, D. (2010). Phonemic restoration in sensorineural hearing loss does not depend 
on baseline speech perception scores. J Acoust Soc Am, 128(4), EL169-EL174.  
Ba?kent, D. (2012). Effect of speech degradation on top-down repair: phonemic 
restoration with simulations of cochlear implants and combined electric-acoustic 
stimulation. J Assoc Res Otolaryngol, 13(5), 683-692. doi:10.1007/s10162-012-
0334-3 
146 
 
 
 
Ba?kent, D., Clarke, J., Pals, C., Benard, M. R., Bhargava, P., Saija, J. D., . . . Gaudrain, 
E. (2016). Cognitive compensation of speech perception with hearing impairment, 
cochlear implants, and aging: How and to what degree can it be achieved? Trends 
in hearing, 20, 1-16.  
Ba?kent, D., Eiler, C., & Edwards, B. (2009). Effects of envelope discontinuities on 
perceptual restoration of amplitude-compressed speech. J Acoust Soc Am, 125(6), 
3995-4005. doi:10.1121/1.3125329 
Ba?kent, D., Eiler, C. L., & Edwards, B. (2010). Phonemic restoration by hearing-
impaired listeners with mild to moderate sensorineural hearing loss. Hear Res, 
260(1-2), 54-62. doi:10.1016/j.heares.2009.11.007 
Benard, M. R., & Ba?kent, D. (2014). Perceptual learning of temporally interrupted 
spectrally degraded speech. J Acoust Soc Am, 136(3), 1344. 
doi:10.1121/1.4892756 
Benard, M. R., Mensink, J. S., & Ba?kent, D. (2014). Individual differences in top-down 
restoration of interrupted speech: links to linguistic and cognitive abilities. J 
Acoust Soc Am, 135(2), EL88-94. doi:10.1121/1.4862879 
Bench, J., Kowal, A., & Bamford, J. (1979). The BKB (Bamford-Kowal-Bench) sentence 
lists for partially-hearing children. Brit. J. Audiol., 13, 108-112.  
Berg, K. A., Noble, J. H., Dawant, B. M., Dwyer, R. T., Labadie, R. F., & Gifford, R. H. 
(2019). Speech recognition as a function of the number of channels in 
perimodiolar electrode recipients. J Acoust Soc Am, 145(3), 1556-1564.  
Bhargava, P., Gaudrain, E., & Ba?kent, D. (2014). Top-down restoration of speech in 
cochlear-implant users. Hear Res, 309, 113-123. 
doi:10.1016/j.heares.2013.12.003 
Bidelman, G. M., & Patro, C. (2016). Auditory perceptual restoration and illusory 
continuity correlates in the human brainstem. Brain Res, 1646, 84-90. 
doi:10.1016/j.brainres.2016.05.050 
Biever, A., Gilden, J., Zwolan, T. A., Mears, M., & Beiter, A. (2018). Upgrade to 
Nucleus 6 in Previous Generation Cochlear Sound Processor Recipients. J Am 
Acad Audiol, 29, 802-813.  
147 
 
 
 
Billings, B. H., & Gray, D. E. (Eds.). (1972). American Institute of Physics Handbook 
(3rd ed.). New York: McGraw-Hill. 
Blamey, P. (2005). Adaptive dynamic range optimization (ADRO): a digital 
amplification strategy for hearing aids and cochlear implants. Trends Amplif, 9(2), 
77-98.  
Blamey, P., Artieres, F., Ba?kent, D., Bergeron, F., Beynon, A., Burke, E., . . . Lazard, D. 
S. (2013). Factors affecting auditory performance of postlinguistically deaf adults 
using cochlear implants: an update with 2251 patients. Audiol Neurootol, 18(1), 
36-47. doi:10.1159/000343189 
Boisvert, I., McMahon, C. M., Dowell, R. C., & Lyxell, B. (2015). Long-term 
asymmetric hearing affects cochlear implantation outcomes differently in adults 
with pre- and postlingual hearing loss. PLoS One, 10(6), e0129167. 
doi:10.1371/journal.pone.0129167 
Bologna, W. J., Vaden, K. I., Jr., Ahlstrom, J. B., & Dubno, J. R. (2018). Age effects on 
perceptual organization of speech: Contributions of glimpsing, phonemic 
restoration, and speech segregation. J Acoust Soc Am, 144(1), 267. 
doi:10.1121/1.5044397 
Bost, T. J. M., Versfeld, N. J., & Goverts, S. T. (2019). Effect of Audibility and 
Suprathreshold Deficits on Speech Recognition for Listeners With Unilateral 
Hearing Loss. Ear and Hearing, 40(4).  
Bregman, A. S. (1990). Auditory Scene Analysis. Cambridge, MA: The MIT Press. 
Busch, T., Vanpoucke, F., & van Wieringen, A. (2017). Auditory Environment Across 
the Life Span of Cochlear Implant Users: Insights From Data Logging. J Speech 
Lang Hear Res, 60(5), 1362-1377. doi:10.1044/2016_JSLHR-H-16-0162 
Caldwell, A., & Nittrouer, S. (2013). Speech perception in noise by children with 
cochlear implants. J Speech Lang Hear Res, 56(1), 13-30. doi:10.1044/1092-
4388(2012/11-0338) 
Caldwell, M. T., Jiam, N. T., & Limb, C. J. (2017). Assessment and improvement of 
sound quality in cochlear implant users. Laryngoscope Investigative 
Otolaryngology, 2(3), 119-124.  
148 
 
 
 
Chaikof, R. (2016). Brand Comparison Chart - Choosing a Cochlear Implant Brand. 
Retrieved from http://cochlearimplantonline.com/site/?page_id=2359 
Chatterjee, M., Peredo, F., Nelson, D., & Ba?kent, D. (2010). Recognition of interrupted 
sentences under conditions of spectral degradation. J Acoust Soc Am, 127(2), 
EL37-41. doi:10.1121/1.3284544 
Clarke, J., Ba?kent, D., & Gaudrain, E. (2016). Pitch and spectral resolution: A 
systematic comparison of bottom-up cues for top-down repair of degraded speech. 
J Acoust Soc Am, 139(1), 395-405. doi:10.1121/1.4939962 
Cochlear Implants. (2017). Retrieved from https://www.medel.com/us/cochlear-implants 
The Cochlear? Nucleus? Implant System. (2018). Retrieved from 
https://www.cochlear.com/us/home/treatment-options-for-hearing-loss/cochlear-
implants 
Collison, E. A., Munson, B., & Carney, A. E. (2004). Relations Among Linguistic and 
Cognitive Skills and Spoken Word Recognition in Adults With Cochlear 
Implants. J Speech Lang Hear Res, 47(3), 496-508. doi:10.1044/1092-
4388(2004/039) 
Croghan, N. B. H., Duran, S. I., & Smith, Z. M. (2017). Re-examining the relationship 
between number of cochlear implant channels and maximal speech intelligibility. 
J Acoust Soc Am, 142(6), EL537. doi:10.1121/1.5016044 
Davidson, L. S., Geers, A. E., & Brenner, C. (2010). Cochlear implant characteristics and 
speech perception skills of adolescents with long-term device use. Otol Neurotol, 
31(8), 1310-1314. doi:10.1097/MAO.0b013e3181eb320c 
Dawson, P. W., Mauger, S. J., & Hersbach, A. A. (2011). Clinical evaluation of signal-to-
noise ratio-based noise reduction in Nucleus(R) cochlear implant recipients. Ear 
Hear, 32(3), 382-390. doi:10.1097/AUD.0b013e318201c200 
Dorman, M. F., Loizou, P. C., Fitzke, J., & Tu, Z. (1998). The recognition of sentences in 
noise by normal-hearing listeners using simulations of cochlear-implant signal 
processors with 6?20 channels. J Acoust Soc Am, 104(6), 3583-3585. 
doi:10.1121/1.423940 
149 
 
 
 
Drag, L. L., & Bieliauskas, L. A. (2010). Contemporary review 2009: Cognitive aging. J 
Geriatr Psychiatry Neurol, 23(2), 75-93. doi:10.1177/0891988709358590 
Dunn, L. M., & Dunn, D. M. (2007). The Peabody Picture Vocabulary Test, Fourth 
Edition. Bloomington, MN: NCS Pearson, Inc. 
Fact Sheet: Cochlear Implants. (2010, October 2010). Retrieved from 
https://report.nih.gov/nihfactsheets/Pdfs/CochlearImplants(NIDCD).pdf 
Faulkner, K. F., & Pisoni, D. B. (2013). Some observations about cochlear implants: 
challenges and future directions. Neuroscience Discovery, 1(1). 
doi:10.7243/2052-6946-1-9 
Fetterman, B. L., & Domico, E. H. (2002). Speech Recognition in Background Noise of 
Cochlear Implant Patients. Otolaryngology?Head and Neck Surgery, 126(3), 257-
263. doi:10.1067/mhn.2002.123044 
Finke, M., B?chner, A., Ruigendijk, E., Meyer, M., & Sandmann, P. (2016). On the 
relationship between auditory cognition and speech intelligibility in cochlear 
implant users: An ERP study. Neuropsychologia, 87(1), 169-181.  
Finke, M., Sandmann, P., Kopp, B., Lenarz, T., & B?chner, A. (2015). Auditory 
distraction transmitted by a cochlear implant alters allocation of attentional 
resources. Front Neurosci, 9(68).  
Friesen, L. M., Shannon, R. V., Ba?kent, D., & Wang, X. (2001). Speech recognition in 
noise as a function of the number of spectral channels: comparison of acoustic 
hearing and cochlear implants. J Acoust Soc Am, 110(2), 1150-1163.  
Fu, Q. J., & Nogaki, G. (2005). Noise susceptibility of cochlear implant users: the role of 
spectral resolution and smearing. J Assoc Res Otolaryngol, 6(1), 19-27. 
doi:10.1007/s10162-004-5024-3 
Fu, Q. J., Shannon, R. V., & Wang, X. (1998). Effects of noise and spectral resolution on 
vowel and consonant recognition: acoustic and electric hearing. J Acoust Soc Am, 
104(6), 3586-3596.  
150 
 
 
 
Gershon, R. C., Wagster, M. V., Hendrie, H. C., Fox, N. A., Cook, K. F., & Nowinski, C. 
J. (2013). NIH toolbox for assessment of neurological and behavioral function. 
Neurology, 80(11 Suppl 3), S2-6. doi:10.1212/WNL.0b013e3182872e5f 
Gifford, R. H., & Revit, L. J. (2010). Speech perception for adult cochlear implant 
recipients in a realistic background noise: effectiveness of preprocessing 
strategies and external options for improving speech recognition in noise. J Am 
Acad Audiol, 21(7), 441-451; quiz 487-448. doi:10.3766/jaaa.21.7.3 
Gilden, J., Lewis, K., Grant, G., & Crosson, J. (2015). Improved hearing in noise using 
new signal processing algorithms with the Cochlear? Nucleus? 6 sound 
processor. Journal of Otology, 10(2), 51-56. 
doi:https://doi.org/10.1016/j.joto.2015.09.001 
Gnansia, D., Pressnitzer, D., Pean, V., Meyer, B., & Lorenzi, C. (2010). Intelligibility of 
interrupted and interleaved speech for normal-hearing listeners and cochlear 
implantees. Hear Res, 265(1-2), 46-53. doi:10.1016/j.heares.2010.02.012 
Gomersall, P. A., Baguley, D. M., & Carlyon, R. P. (2019). A Cross-Sectional 
Questionnaire Study of Tinnitus Awareness and Impact in a Population of Adult 
Cochlear Implant Users. Ear Hear, 40(1), 135-142. 
doi:10.1097/AUD.0000000000000601 
Gordon-Salant, S., Fitzgibbons, P. J., & Yeni-Komshian, G. H. (2011). Auditory 
Temporal Processing and Aging: Implications for Speech Understanding of Older 
People. Audiol Res, 1(1), e4. doi:10.4081/audiores.2011.e4 
Green, K. M., Bhatt, Y., Mawman, D. J., O'Driscoll, M. P., Saeed, S. R., Ramsden, R. T., 
& Green, M. W. (2007). Predictors of audiological outcome following cochlear 
implantation in adults. Cochlear Implants Int, 8(1), 1-11. 
doi:10.1179/cim.2007.8.1.1 
Greenberg, S., Carvey, H., Hitchcock, L., & Chang, S. (2003). Temporal properties of 
spontaneous speech?a syllable-centric perspective. Journal of Phonetics, 31(3), 
465-485. doi:https://doi.org/10.1016/j.wocn.2003.09.005 
Holden, L. K., Finley, C. C., Firszt, J. B., Holden, T. A., Brenner, C., Potts, L. G., . . . 
Skinner, M. W. (2013). Factors affecting open-set word recognition in adults with 
cochlear implants. Ear Hear, 34(3), 342-360.  
151 
 
 
 
Hox, J. J., Moerbeek, M., & van de Schoot, R. (2018). Multilevel Analysis: Techniques 
and Applications (Third ed.). New York, NY: Routledge. 
Hughes, K. C., & Galvin, K. L. (2013). Measuring listening effort expended by 
adolescents and young adults with unilateral or bilateral cochlear implants or 
normal hearing. Cochlear Implants Int, 14(3), 121-129. 
doi:10.1179/1754762812Y.0000000009 
Ishida, M., & Arai, T. (2016). Missing phonemes are perceptually restored but differently 
by native and non-native listeners. Springerplus, 5(1), 713. doi:10.1186/s40064-
016-2479-8 
Jaekel, B. N., Newman, R. S., & Goupell, M. J. (2018). Age effects on perceptual 
restoration of degraded interrupted sentences. J Acoust Soc Am, 143(1), 84. 
doi:10.1121/1.5016968 
Jin, S. H., Nie, Y., & Nelson, P. (2013). Masking release and modulation interference in 
cochlear implant and simulation listeners. Am J Audiol, 22(1), 135-146. 
doi:10.1044/1059-0889(2013/12-0049) 
Kaandorp, M. W., Smits, C., Merkus, P., Goverts, S. T., & Festen, J. M. (2015). 
Assessing speech recognition abilities with digits in noise in cochlear implant and 
hearing aid users. Int J Audiol, 54(1), 48-57. doi:10.3109/14992027.2014.945623 
Kan, A. (2018). Improving Speech Recognition in Bilateral Cochlear Implant Users by 
Listening With the Better Ear. Trends in hearing, 22, 2331216518772963-
2331216518772963. doi:10.1177/2331216518772963 
Khing, P. P., Swanson, B. A., & Ambikairajah, E. (2013). The effect of automatic gain 
control structure and release time on cochlear implant speech intelligibility. PLoS 
One, 8(11), e82263. doi:10.1371/journal.pone.0082263 
Koroleva, I. V., Kashina, A., Sakhnovskaya, O. S., & Shurgaya, G. G. (1991). Perceptual 
restoration of a missing phoneme: New data on speech perception in children. 
Sensornye Sistemy, 5(3), 30-40.  
Lassaletta, L., Castro, A., Bastarrica, M., de Sarria, M. J., & Gavilan, J. (2006). Quality 
of life in postlingually deaf patients following cochlear implantation. Eur Arch 
Otorhinolaryngol, 263(3), 267-270. doi:10.1007/s00405-005-0987-1 
152 
 
 
 
Leonard, M. K., Baud, M. O., Sjerps, M. J., & Chang, E. F. (2016). Perceptual restoration 
of masked speech in human cortex. Nat Commun, 7, 13619. 
doi:10.1038/ncomms13619 
Lin, F. R., Chien, W. W., Li, L., Clarrett, D. M., Niparko, J. K., & Francis, H. W. (2012). 
Cochlear implantation in older adults. Medicine (Baltimore), 91(5), 229-241. 
doi:10.1097/MD.0b013e31826b145a 
Litovsky, R., Parkinson, A., Arcaroli, J., & Sammeth, C. (2006). Simultaneous bilateral 
cochlear implantation in adults: a multicenter clinical study. Ear Hear, 27(6), 
714-731. doi:10.1097/01.aud.0000246816.50820.42 
Loizou, P. C. (2006). Speech processing in vocoder-centric cochlear implants. Adv 
Otorhinolaryngol, 64, 109-143. doi:10.1159/000094648 
Loizou, P. C., Hu, Y., Litovsky, R., Yu, G., Peters, R., Lake, J., & Roland, P. (2009). 
Speech recognition by bilateral cochlear implant users in a cocktail-party setting. 
J Acoust Soc Am, 125(1), 372-383. doi:10.1121/1.3036175 
Mauger, S. J., Warren, C. D., Knight, M. R., Goorevich, M., & Nel, E. (2014). Clinical 
evaluation of the Nucleus 6 cochlear implant system: performance improvements 
with SmartSound iQ. Int J Audiol, 53(8), 564-576. 
doi:10.3109/14992027.2014.895431 
McNamara, T. P. (2005). Semantic priming: Perspectives from memory and word 
recognition. New York: Psychology Press. 
Meyer, B. T., Brand, T., & Kollmeier, B. (2011). Effect of speech-intrinsic variations on 
human and automatic recognition of spoken phonemes. J Acoust Soc Am, 129(1), 
388-403.  
Miller, G. A., & Licklider, J. C. R. (1950). The Intelligibility of Interrupted Speech. J 
Acoust Soc Am, 22(2), 167-173. doi:Doi 10.1121/1.1906584 
Moore, B. C. J. (2012). An introduction to the psychology of hearing (6th ed.). United 
Kingdom: Emerald. 
153 
 
 
 
Nagaraj, N. K., & Magimairaj, B. M. (2017). Role of working memory and lexical 
knowledge in perceptual restoration of interrupted speech. J Acoust Soc Am, 
142(6), 3756. doi:10.1121/1.5018429 
Nasreddine, Z. S., Phillips, N. A., Bedirian, V., Charbonneau, S., Whitehead, V., Collin, 
I., . . . Chertkow, H. (2005). The Montreal Cognitive Assessment, MoCA: a brief 
screening tool for mild cognitive impairment. J Am Geriatr Soc, 53(4), 695-699. 
doi:10.1111/j.1532-5415.2005.53221.x 
Nelson, P. B., & Jin, S. H. (2004). Factors affecting speech understanding in gated 
interference: cochlear implant users and normal-hearing listeners. J Acoust Soc 
Am, 115(5 Pt 1), 2286-2294. doi:10.1121/1.1703538 
Nelson, P. B., Jin, S. H., Carney, A. E., & Nelson, D. A. (2003). Understanding speech in 
modulated interference: cochlear implant users and normal-hearing listeners. J 
Acoust Soc Am, 113(2), 961-968. doi:10.1121/1.1531983 
Newman, R. S. (2004). Perceptual restoration in children versus adults. Applied 
Psycholinguistics, 25(4), 481-493. doi:10.1017/s0142716404001237 
Newman, R. S. (2006). Perceptual restoration in toddlers. Perception & Psychophysics, 
68(4), 625-642.  
NIDCD fact sheet: Cochlear implants. (2016). National Institute of Deafness and Other 
Communication Disorders Retrieved from 
https://www.nidcd.nih.gov/sites/default/files/Documents/health/hearing/Factsheet
CochlearImplants.pdf. 
Nie, K., Barco, A., & Zeng, F. G. (2006). Spectral and temporal cues in cochlear implant 
speech perception. Ear Hear, 27(2), 208-217. 
doi:10.1097/01.aud.0000202312.31837.25 
O'Neill, E. R., Kreft, H. A., & Oxenham, A. J. (2018). Speech perception with spectrally 
non-overlapping maskers as measure of spectral resolution in cochlear implant 
users. JARO: Journal of the Association for Research in Otolaryngology, 20, 151-
167.  
154 
 
 
 
Ohlenforst, B., Souza, P., & MacDonald, E. N. (2016). Exploring the relationship 
between working memory, compressor speed, and background noise 
characteristics. Ear Hear, 37(2), 137-143.  
Ouda, L., Profant, O., & Syka, J. (2015). Age-related changes in the central auditory 
system. Cell and Tissue Research, 361(1), 337-358. doi:10.1007/s00441-014-
2107-2 
Oxenham, A. J., & Kreft, H. A. (2014). Speech perception in tones and noise via cochlear 
implants reveals influence of spectral resolution on temporal processing. Trends 
Hear, 18. doi:10.1177/2331216514553783 
Park, D. C., Lautenschlager, G., Hedden, T., Davidson, N. S., Smith, A. D., & Smith, P. 
K. (2002). Models of visuospatial and verbal memory across the adult life span. 
Psychol Aging, 17(2), 299-320. doi:10.1037/0882-7974.17.2.299 
Pastor, D. A. (2010). Cluster Analysis. In G. R. Hancock & R. O. Mueller (Eds.), The 
Reviewer's Guide to Quantitative Methods in the Social Sciences (pp. 41-54). 
New York, NY: Routledge. 
Patro, C., & Mendel, L. L. (2016). Role of contextual cues on the perception of spectrally 
reduced interrupted speech. J Acoust Soc Am, 140(2), 1336. 
doi:10.1121/1.4961450 
Patro, C., & Mendel, L. L. (2020). Semantic influences on the perception of degraded 
speech by individuals with cochlear implants. J Acoust Soc Am, 147(3), 1778. 
doi:10.1121/10.0000934 
Peirce, J., Gray, J. R., Simpson, S., MacAskill, M., H?chenberger, R., Sogo, H., . . . 
Lindel?v, J. K. (2019). PsychoPy2: Experiments in behavior made easy. Behavior 
Research Methods, 51(1), 195-203. doi:10.3758/s13428-018-01193-y 
Perreau, A. E., Wu, Y. H., Tatge, B., Irwin, D., & Corts, D. (2017). Listening Effort 
Measured in Adults with Normal Hearing and Cochlear Implants. J Am Acad 
Audiol, 28(8), 685-697. doi:10.3766/jaaa.16014 
Perry, T. T., & Kwon, B. J. (2015). Amplitude fluctuations in a masker influence lexical 
segmentation in cochlear implant users. J Acoust Soc Am, 137(4), 2070-2079. 
doi:10.1121/1.4916698 
155 
 
 
 
Petersen, B., Gjedde, A., Wallentin, M., & Vuust, P. (2013). Cortical plasticity after 
cochlear implantation. Neural Plast, 2013, 318521. doi:10.1155/2013/318521 
Peterson, N. R., Pisoni, D. B., & Miyamoto, R. T. (2010). Cochlear implants and spoken 
language processing abilities: review and assessment of the literature. Restor 
Neurol Neurosci, 28(2), 237-250. doi:10.3233/RNN-2010-0535 
Pichora-Fuller, M. K. (2008). Use of supportive context by younger and older adult 
listeners: balancing bottom-up and top-down information processing. Int J Audiol, 
47 Suppl 2, S72-82. doi:10.1080/14992020802307404 
Pisoni, D. B. (2000). Cognitive factors and cochlear implants: some thoughts on 
perception, learning, and memory in speech perception. Ear Hear, 21(1), 70-78.  
Powers, G. L., & Wilcox, J. C. (1977). Intelligibility of Temporally Interrupted Speech 
with and without Intervening Noise. J Acoust Soc Am, 61(1), 195-199. doi:Doi 
10.1121/1.381255 
Purdy, S. C., Welch, D., Giles, E., Morgan, C. L. A., Tenhagen, R., & Kuruvilla-Mathew, 
A. (2017). Impact of cognition and noise reduction on speech perception in adults 
with unilateral cochlear implants. Cochlear Implants Int, 18(3), 162-170. 
doi:10.1080/14670100.2017.1299393 
Rakszawski, B., Wright, R., Cadieux, J. H., Davidson, L. S., & Brenner, C. (2016). The 
effects of pre-processing strategies for pediatric cochlear implant recipients. J Am 
Acad Audiol, 27(2), 85-102.  
Rastle, K., Harrington, J., & Coltheart, M. (2002). 358,534 nonwords: the ARC Nonword 
Database. Q J Exp Psychol A, 55(4), 1339-1362. 
doi:10.1080/02724980244000099 
R?nnberg, J., Rudner, M., Foo, C., & Lunner, T. (2008). Cognition counts: a working 
memory system for ease of language understanding (ELU). Int J Audiol, 47 Suppl 
2, S99-105. doi:10.1080/14992020802301167 
Rothauser, E., Chapman, W., Guttman, N., Nordby, K., Silbiger, H., Urbanek, G., & 
Weinstock, M. (1969). IEEE recommended practice for speech quality 
measurements. IEEE Trans. Audio Electroacoust., 17, 225-246.  
156 
 
 
 
Saija, J. D., Aky?rek, E. G., Andringa, T. C., & Ba?kent, D. (2014). Perceptual 
restoration of degraded speech is preserved with advancing age. J Assoc Res 
Otolaryngol, 15(1), 139-148. doi:10.1007/s10162-013-0422-z 
Samuel, A. G. (1981). Phonemic restoration: insights from a new methodology. J Exp 
Psychol Gen, 110(4), 474-494.  
Samuel, A. G. (1987). Lexical Uniqueness Effects on Phonemic Restoration. Journal of 
Memory and Language, 26, 36-56.  
Sheldon, S., Pichora-Fuller, M. K., & Schneider, B. A. (2008). Priming and sentence 
context support listening to noise-vocoded speech by younger and older adults. J 
Acoust Soc Am, 123(1), 489-499. doi:10.1121/1.2783762 
Shinn-Cunningham, B. G., & Wang, D. (2008). Influences of auditory object formation 
on phonemic restoration. J Acoust Soc Am, 123(1), 295-301. 
doi:10.1121/1.2804701 
Sladen, D. P., & Zappler, A. (2015). Older and Younger Adult Cochlear Implant Users: 
Speech Recognition in Quiet and Noise, Quality of Life, and Music Perception. 
Am J Audiol, 24(1), 31-39. doi:10.1044/2014_AJA-13-0066 
Smith, K. G., & Fogerty, D. (2017). Speech recognition error patterns for steady-state 
noise and interrupted speech. J Acoust Soc Am, 142(3), El306. 
doi:10.1121/1.5003916 
Spencer, N. J., Tillery, K. H., & Brown, C. A. (2018). The effects of dynamic-range 
automatic gain control on sentence intelligiblity with a speech masker in 
simulated cochlear implant listening. Ear Hear, 40(3), 710-724.  
Studebaker, G. A. (1985). A rationalized arcsine transform. Journal of Speech and 
Hearing Research, 28, 455-462.  
Tulsky, D. S., Carlozzi, N., Chiaravalloti, N. D., Beaumont, J. L., Kisala, P. A., Mungas, 
D., . . . Gershon, R. (2014). NIH Toolbox Cognition Battery (NIHTB-CB): list 
sorting test to measure working memory. J Int Neuropsychol Soc, 20(6), 599-610. 
doi:10.1017/S135561771400040X 
157 
 
 
 
Verschuure, J., & Brocaar, M. P. (1983). Intelligibility of Interrupted Meaningful and 
Nonsense Speech with and without Intervening Noise. Perception & 
Psychophysics, 33(3), 232-240. doi:Doi 10.3758/Bf03202859 
Walley, A. C. (1988). Spoken Word Recognition by Young-Children and Adults. 
Cognitive Development, 3(2), 137-165. doi:Doi 10.1016/0885-2014(88)90016-0 
Warren, R. M. (1970). Perceptual restoration of missing speech sounds. Science, 
167(3917), 392-393.  
Warren, R. M., & Obusek, C. J. (1971). Speech perception and phonemic restorations. 
Atten Percept Psychophys, 9(3B), 358-362.  
Warren, R. M., & Sherman, G. L. (1974). Phonemic Restorations Based on Subsequent 
Context. Perception & Psychophysics, 16(1), 150-156. doi:Doi 
10.3758/Bf03203268 
Winn, M. B., & Moore, A. M. (2018). Pupillometry reveals that context benefit in speech 
perception can be disrupted by later-occuring sounds, especially in listeners with 
cochlear implants. Trends Hear, 22.  
Winstone, N., Davis, A., & de Bruyn, B. (2012). Developmental Improvements in 
Perceptual Restoration: Can Young Children Reconstruct Missing Sounds in 
Noisy Environments? Infant and Child Development, 21(3), 287-297. 
doi:10.1002/icd.749 
Wolfe, J., Neumann, S., Marsh, M., Schafer, E., Lianos, L., Gilden, J., . . . Jones, M. 
(2015). Benefits of Adaptive Signal Processing in a Commercially Available 
Cochlear Implant Sound Processor. Otol Neurotol, 36(7), 1181-1190. 
doi:10.1097/MAO.0000000000000781 
Xu, L., Thompson, C. S., & Pfingst, B. E. (2005). Relative contributions of spectral and 
temporal cues for phoneme recognition. J Acoust Soc Am, 117(5), 3255-3267. 
doi:10.1121/1.1886405 
Yathiraj, A., & Rao, A. (2013). Preprocessing strategies and speech perception in 
cochlear implant users. Journal of Hearing Science, 3(2), 50-59.  
158 
 
 
 
Zeng, F. G., Grant, G., Niparko, J., Galvin, J., Shannon, R., Opie, J., & Segel, P. (2002). 
Speech dynamic range and its effect on cochlear implant performance. J Acoust 
Soc Am, 111(1 Pt 1), 377-386. doi:10.1121/1.1423926 
Zhao, F., Stephens, S. D. G., Sim, S. W., & Meredith, R. (1997). The use of qualitative 
questionnaires in patients having and being considered for cochlear implants. 
Clinical Otolaryngology, 22(3), 254-259. doi:DOI 10.1046/j.1365-
2273.1997.00036.x 
 
159