Forty CI users, 26 with Cochlear device, 14 with AB device, performed the phoneme and DiN tasks at home. Their median age was 64.3 years [IQR 10.4, min 28 yrs, max 75 yrs], median experience with their CI 2.1 years [IQR 4.2 yrs, range 0.1–15.9 yrs]. Thirty-six out of forty CI users had progressive hearing loss. Twenty-seven participants wore a hearing aid contralaterally (CI-HA), eight persons had one CI, three persons bilateral CIs, and two persons 1 CI and residual hearing. The participants’ average pure tone average (PTA4, average of 500, 1,000, 2,000, 4,000 Hz), determined in free field at the clinic with their CI only, was 26.4 dB HL (SD 5.3). All participants presented with a postlingually acquired profound hearing impairment, and they communicated through spoken language in their daily life. The median period of education was 12.5 years [IQR 3.3].

These participants participated in a more extensive study dealing with the efficacy of a personalized listening training program LUISTER compared to a non-personalized one (Magits et al., under revision). In that study, participants were asked to practice segmental and suprasegmental speech tasks five times per week for 15 to 20 min on a tablet at home. The efficacy of the two training tasks was based on the SiN scores (pre- versus post-training) assessed at the clinic. Once a week, before practicing with a training program, the participants were asked to complete a DiN test twice and either a vowel or a consonant phoneme identification task (in quiet) at home. At home, the stimuli were streamed via Bluetooth and a streaming device to one CI. The participant chose which CI if they had two. Speech understanding in noise (SiN, pre-and post-training) was assessed at the clinic, via streaming. Three conditions were tested: (1) SiN presented via streaming to one CI (same as DiN and phoneme in quiet, “SiN streaming”), (2) in sound field to the CI only (“SiN CI-SF”), and (3) in sound field as in daily life (with CI and HA if applicable, “SiN daily settings”). The same CI devices were used at home and at the clinic. Logged data were automatically transferred to a repository hosted on the server of the research group via a restricted one-way communication from tablet to server.

Participants provided written informed consent, and the Ethics Committee approved the study of the University Hospitals Leuven (approval no. B322201731501). Participants were paid for the testing sessions but not compensated for the practicing sessions at home. The study protocol is registered on ClinicalTrials.gov (I.D. = NCT04063748).

Statistical analyses were performed using IBM SPSS Statistics for Windows version 27 (2020). Data were tested for normality and homogeneity of variance. The Shapiro-Wilk showed that the DiN data distribution did not significantly differ from normal, W(40) = 0.946, p = 0.057, but the SiN data did W(40) = 0.907, p = 0.003. Vowel identification scores were normally distributed: W(40) = 0.978, p = 0.628, as well as consonant identification scores: W(40) = 0.984, p = 0.845. The pure tone average (PTA) data were also normally distributed, W(40) = 0.967, p = 0.296, but not “years of CI use,” W(40) = 0.836, p < 0.001. Since the SiN data were not normally distributed we opted for median and interquartile ranges when presenting SiN with other performance measures. The non-parametric Spearman’s Rho was used to determine the strength of an association between SiN and other variables, while Pearson correlation (r) was used for the normally distributed performance measures. Linear regression analyses were performed to study the relationship between different performance measures and to determine how much the different predictors explain the response. Potential differences in performance between the start and end of the 16-week trial were analyzed with the non-parametric Friedman test of differences among repeated measures, followed by a Wilcoxon signed rank test for paired comparisons.

Figure 1 illustrates SiN (streaming) and DiN (A) and percentage vowel and consonant identification in quiet (B) for each of the 40 participants separately. Participants are ranked according to increasing (poorer) SiN scores determined at the end of the 16-week trial. These scores range from –5.2 to + 11.6 dB SNR. Generally, DiN scores are in line with the SiN ones by 3–4 dB improvement. The median SRT of the last five trials for DiN is –3.8 dB SNR [IQR, 5.0), and for SiN streamed to the device –0.3 dB SNR [IQR, 4.5). The difference between the DiN and the SiN in this study, about 3.5 dB, is also in line with the difference between the norm values of the SiN for normal hearing young persons (−7.8 dB SNR, van Wieringen and Wouters, 2008) and the norm values of the DiN (−11.7 dB SNR, Jansen et al., 2014).

Spearman’s rho indicates a statistically significant relationship between the SRTs of SiN and DiN (rs [40] = 0.767, p < 0.001). Linear regression analyses showed that DiN significantly predicts SiN, thereby explaining 74% of the variance, F(1,38) = 621.34, p < 0.0001. The model for SiN is y = 4.52 + (1.098 *score) with a narrow 95% confidence interval to predict SiN from the DiN score [3.5–5.4].

The bottom panel (Figure 1B) illustrates phoneme identification in quiet for each of the participants. All participants performed well above chance (10% for vowels and 8.3% for consonants), but a wide range of performance is observed. Median vowel identification is 70.0% [IQR 17.8], median consonant identification is 64.4% [IQR 24.1]. Vowel and consonant perception in quiet are highly correlated [r(40) = 0.678, p < 0.001], the difference between the two measures is in the same order of magnitude for most participants.

Spearman’s rho indicated a significant negative relationship between SiN assessed at the clinic and vowel identification in quiet assessed at home (rs [40] = −0.611, p < 0.001), and a significant negative relationship between SiN and consonant identification in quiet (rs [40] = −0.587, p < 0.001). In other words, the more negative (better) sentence identification in noise, the higher the vowel and consonant recognition in quiet. Vowel and consonant recognition significantly predict SiN, with the linear regression model explaining 41% of the variance (p < 0.001). Semi partial correlations, which explain the unique contribution of each predictor variable, are 28% for vowel identification, p = 0.031, and 26% for consonant identification, p = 0.043.

As with SiN, a significant negative relationship was observed between DiN and vowel identification: r (40) = −0.537, p < 0.001, and between DiN and consonant identification: r (40) = −0.520, p = 0.001.

Vowel and consonant identification also significantly predict DiN, albeit somewhat less than SiN: the model explains 30% variance (p < 0.001). Semi partial correlations show that vowels predict 25% and consonants predict 21% of DiN.

Figure 2A illustrates the average percentages for each of the three vowel features for each 40 participants, together with the percentage correct score (blue stars). Participants are ranked according to increasing (poorer) SiN, as in Figure 1. Averaged over participants, “duration” is 50.4% (SD (19.0%), “F1” is 57.1% (SD 19%) and “F2” is 72.5% (SD 21%). F2 predicts 65% of the variance for vowel identification and was significant F(1,305) = 574.0, p < 0.0001. The linear regression analysis was performed on the individual data. Due to multicollinearity, the features “duration” and “F1” were dropped from the model (r > 0.7).

Information transmission analyses show that CI users with similar percent correct recognition scores can make different errors. Compare, for instance, vowel recognition of participants 15, 20, and 26 in Figure 2A. While percentage correct scores are similar (57%), the distributions of errors are different: participant 26 perceives the high-frequency spectral information much better (F2 cue, 76%) than participant 15 (40%) or participant 20 (57%). Likewise, participant 30 discriminates long and short vowels much better (80%) than participant 31 (53%) despite similar percentage correct scores (80%). Compare the data of participants 2 & 25, 21 & 23, and others who have similar percentage correct scores but perceive the different speech features differently.

Figure 2B illustrates the average percentages of the seven features per participant, together with the percentage correct scores ranked from low to high (blue crosses). Again, the order of participants is according to Figure 1 (SiN). Perception of voicing (AVG 61.6%, SD 27.1%), and place of articulation (AVG 47.0%, SD 19.1%) remain difficult, but the perception of plosives (AVG 79.2%, SD 23.4%), the coding of temporal envelope cues (envelop, AVG 88.2%, SD 16.6%), manner of articulation (AVG = 87.1%, SD 13.8%), fricatives 86.2% (SD 18.6%), and nasals (AVG 84.8%, SD 20%) are generally good. As with vowel identification, the feature transmission analyses of the consonant confusions provide additional information on differential performance.

The perception of the seven features can vary widely for a similar percentage correct score: compare, for instance, the data of participants 14 and 21 who both have similar recognition scores (54%). However, participant 21 mainly has difficulty perceiving place of articulation and perceives the other cues very well (>80%). In contrast, participant 14 has difficulty perceiving the correct place of articulation, perception of the burst, and frication (all below 50%). Compare also data of participants 6, 7, & 8, 9 & 11 to name a few. These participants yield the same percentage correct score, yet a different distribution of errors. Analysis of the errors can guide the mapping of the device and the rehabilitation process.

Of the seven consonantal speech features, “voicing” and “frication” significantly predict 53% of the variance of consonant identification in quiet, F(2,323) = 186.1, p < 0.0001 (n = 326). Both features contribute uniquely to predicting consonant recognition in quiet (r = 0.433 for “frication” and r = 0.3 for “voicing.” Due to multicollinearity (r > 0.7) the features “plos,” or perception of release burst (r = 0.77), envelope (r = 0.72), place of articulation (r = 0.85), manner of articulation (r = 0.8), and nasal (r = 0.72) were removed from the model.

A primary reason to use the DiN test is the limited content learning and thus the possibility to use the paradigm repeatedly (Smits et al., 2013). During the 16-week training trial (Magits et al., under revision), participants performed the DiN at home each week. Figure 3 illustrates the speech reception thresholds (SRTs) as a function of time for each participant separately (ranked in the same order as in previous figures). Each data point is the average of 2 SRTs administered consecutively at the beginning of each week. For the sake of clarity each figure illustrates the data of 5 participants. Most participants yield low SRTs that vary minimally with time, especially those with very good SiN scores (participants 1–10). Test-retest reliability was determined for each participant by taking the standard deviations of the differences between the two consecutive scores, divided by square 2. This procedure outbalances a procedural learning effect (Smits and Houtgast, 2005). Averaged over all participants, the measurement error is 2.0 dB (SD 0.9). Individual measurement errors are indicated next to the participant number in Figure 3 and range from 0.8 dB (participant 3) to 4.7 dB (participant 32). Generally, the values are larger for the participants with poorer DiN (and SiN), cf participants 30–40. These higher and more variable measurement errors were not related to the age of the participants (r (40) = −0.595, p = 0.09).

Closed-set phoneme identification in quiet also offers the possibility to monitor changes with time. Recall that the vowel and consonant tasks were performed every other week during the 16-week trial. Summary scores and speech features of individual trials are illustrated for each participant under Supplementary Material. Vowel and consonant data of the 40 participants are presented in the same order as before. Many participants, especially the lowest ranked ones, show little improvement with time because they already perceive the different features very well. However, others show improvements with time, possibly because of practicing the listening training tasks or due to the weekly testing regime of DiN and phoneme identification to know whether a change in summary score is meaningful, histograms of the differences between consecutive scores were constructed for vowels and consonants separately. These are illustrated in Figure 4. The standard deviation of the distribution can be used as a guideline to make changes to the mapping of the device or to the audiological management of the client. For the current data, changes smaller than 10% may be meaningful, but changes larger than 10% certainly are.

We also compared potential performance differences between the first and the last measurement. For this, we compared 1 SiN value (streaming mode), the average of 2 DiN scores, and one vowel and one consonant identification score assessed at the beginning of the trials with the same outcomes assessed at the end. Phoneme data were transformed to RAU scores (Studebaker, 1985) for the statistical analyses. Figures 5A,B illustrate the difference in performance between the beginning and end of the period. Wilcoxon signed-rank tests for paired comparisons revealed that speech perception scores were significantly lower (better) after 16 weeks than before for all outcome measures: for SiN z = −2.88, p = 0.004, for DiN, z = −4.7, p < 0.0001, for vowel identification in quiet z = −4.5, p < 0.0001 and consonant identification in quiet z = −5.0, p < 0.0001. Median and IQR scores are presented in Table 1.

For the sake of comparison, Table 1 also lists the median SiN thresholds for speech stimuli presented to only the CI (CI-SF) and in the daily settings condition (with CI and HA if applicable). These three different SiN outcomes do not differ statistically from each other at the beginning of the trial. However, at the end of the trial, the non-parametric Friedman test of differences among repeated measures rendered a Chi-square value of 12.1, which was significant (p < 0.005). A Wilcoxon signed-rank tests subsequently showed that the SiN “daily settings” was significantly better than the SiN in streaming mode (x = −2.574, p = 0.010) and SiN CI-SF (X = −3.968, p < 0.0001).

Understanding speech in noise is the most common complaint of persons with hearing impairment, and several indices can be used to document listening difficulties and guide hearing rehabilitation. The present study reports three different indices for speech perception in the same 40 CI users, of which two are administered at home. Where possible, we present individual results instead of group mean average to better understand individual differences in speech recognition outcomes (which, in turn, enables personalized rehabilitation).

SiN performance of contemporary CI users is excellent in the current study, even when the sentences are only streamed to the implanted side. Candidacy criteria for cochlear implantation have changed considerably over the past years, and several CI users have residual hearing (Snel-Bongers et al., 2018). Nevertheless, variability in performance is large, and to understand the source of variability, it is important to look at individual performance and do this from different perspectives. Word and sentence identification remain important measures to evaluate an intervention (e.g., Zhang and Coelho, 2018; Kelsall et al., 2021). While open-set word and sentence understanding lack full external generalizability to speech perception in daily life, they are most closely related to capturing some of the real-world listening difficulties. These measures involve phonological, lexical, grammatical skills, and semantic/contextual knowledge (Heald and Nusbaum, 2014), especially when administered using an open-set response format (Clopper et al., 2006). In an open-set task, listeners compare the stimulus to all possible candidate words in lexical memory.

In contrast, in closed set tests, the listeners need to make only a limited number of comparisons among the response alternatives. An advantage of SiN above word identification is the steeper slope of the performance intensity function of the former. The slope measures how rapidly performance changes with a change in level or signal to noise ratio (Leek, 2001).

However, SiN cannot be assessed too often (due to learning and limited test materials), while DiN can be used repeatedly and without a clinician. The high correlation between DiN and SiN is in line with the results of Smits et al. (2013) for persons with normal hearing and Kaandorp et al. (2015), Kaandorp et al., 2016, 2017Cullington and Aidi (2017), and Zhang et al. (2019) for persons with cochlear implants, thereby indicating that the two measures share some common mechanisms. The difference between the two is in the order of 4 dB SNR in the current study. DiN may even be more sensitive than SiN to capture changes in auditory performance, which can be done for each ear separately. The large dataset of the current study did reveal individual differences in performance which do not necessarily change with time. For some participants, especially those struggling most with SiN, performance varied substantially. Here, measurement error of subsequent trials can be used as a guideline for potential changes in performance.

Phoneme identification in quiet sheds additional light on variability in speech perception. While sentence and word recognition in quiet often yield ceiling scores, phoneme scores provide specific information on the perception of speech cues in the absence of context cues. Phoneme perception has a long history in research (Miller and Nicely, 1955) but is not often used as a standard metric in the clinic. At least two arguments plead in favor of incorporating phoneme identification in clinical care. First, it is essential to know how vowel and consonant identification in quiet relate to performance on tests in noise. Our study shows that vowel and consonant recognition in quiet contribute (uniquely) to SiN (and to DiN) and yields additional information on the audibility and discriminability of speech cues. At the clinic, phoneme recognition is often assessed via meaningful words (phoneme score of a word recognition test). However, nonsense syllables are preferred over meaningful words, because context can affect the recognition of phoneme scores in the latter (Donaldson and Kreft, 2006). With a nonsense syllable task, each phoneme can be presented an equal number of times. The task takes only a few minutes and can easily be done remotely. In the future, phoneme perception in noise will also be considered.

Second, phoneme identification can also be used as a diagnostic tool. While percentage correct is a summary score of phoneme perception, the information transmission analyses reveal which spectral and temporal cues are most challenging for the recipient. This information can help optimize the mapping and provide targeted rehabilitation. For instance, a low score for duration discrimination in vowels could guide the clinician to provide tasks to improve discriminability between short and long vowels. A low score on “frication” or “voicing” could guide the clinician to optimize the mapping of high- and low-frequency cues, respectively. The value of this metric was recognized several years ago (van Wieringen and Wouters, 2000) but seemed cumbersome to implement in clinical care. With novel, cost-effective technologies, the benefit of assessing phoneme perception at regular intervals can be reconsidered. Note that sufficient data scores are required to draw conclusions from the information transmission analyses. The maximum-likelihood estimate for information transfer is biased to overestimate its true value when the number of stimulus presentations is small (Sagi and Svirsky, 2008).

Learning either the content or the procedure of a test could improve performance when presented repeatedly. During the 16-weeks, the participants also practiced training modules (Magits et al., under revision). Comparison of SiN pre- versus post-training showed a significant improvement in speech understanding in noise for both the personalized LUISTER and the non-personalized listening training programs. Since the same sentences were never presented twice to the same participant, the observed differences are more likely to result from practicing than repeated testing. However, it is difficult to determine whether the observed improvements for DiN and phoneme identification result from the content of the listening training (perceptual learning) or procedural learning (repeated listening to a task). All perceptual experiments involve some procedural learning, such as getting acquainted with a voice, the characteristics of the speech material, etc., (Nygaard and Pisoni, 1998; Yund and Woods, 2010). A procedural learning effect is larger for a closed-set than for open-set response format, but Smits et al. (2013) report that procedural learning with DiN is accomplished after 1 trial with normal-hearing persons. Nevertheless, de Graaff et al. (2019) report that DiN data of CI users reveal improvements in speech recognition over time, without a clear relation to fitting appointments with an audiologist. These improvements could result from procedural learning or improved perception of speech perception in general.

Rehabilitation following cochlear implantation is demanding and requires several visits to the clinic to fine-tune the device. With the growing number of clients, improved technology, and public health concerns surrounding the COVID-19 pandemic, remote testing has sparked a lot of interest to complement care at the clinic. The shared responsibility between professional and client may also empower clients to take action if needed. Home-based testing has the potential to change and improve the hearing care pathway. It would not only lead to a reduction in the required number of visits and thus reduction in cost− and time savings for both clinics and patients, it would even improve the quality and richness of data obtained during audiological rehabilitation. The importance of speech in noise testing cannot be overestimated, but note that it entails more than the perception of the auditory signal in noise which can be captured with a DiN task. When the acoustical signal is difficult to perceive, as in noisy conditions, speech understanding places more demands on linguistic knowledge and executive functioning (Mattys et al., 2012; Moberly et al., 2019; Zhan et al., 2020). Remote monitoring of speech-in-noise performance possibly makes room to assess neurocognitive abilities that differentially explain speech in noise performance, which may lead to a personalized holistic management of hearing impairment.

For remote testing to be successful, the obtained data should be clinically valid and accurate, and clients should feel confident handling the device. In our study, all participants felt comfortable doing tests remotely because the professional had provided sufficient information prior to the trial and was online available to address any concerns or technical problems. Data collection with wireless streaming was reliable as repeated testing yielded similar results in the same CI user.

During the COVID-19 pandemic, face-to-face care was brought to a halt, and interest in tele-audiology surged out of necessity. While a recent survey reports that audiologists are generally positive about teleaudiology, infrastructure and training should not be underestimated, and hybrid care remains necessary (Saunders and Roughley, 2021). Also, more research is needed to examine reimbursement and cost-effectiveness of remote services (Bush et al., 2016). Such factors may represent a barrier to the practical delivery of telemedicine services, and these topics represent areas for further research. Technical advances in connectivity now allow for wireless streaming capabilities for current CI systems. Wireless streaming provides good quality audio and is less susceptible to noise or signal processing introduced by the connection cable. Only one calibration is needed for a given digital communication set. In the current study, only the implanted side was assessed at home, but stereo streaming is possible. It remains important to evaluate the whole hearing pathway in the sound field too.