Despite the success of speech separation approaches for dry (non-reverb) speech mixtures, speech separation in naturalistic, spatial, and reverberant acoustic environments remains challenging. This limits the effectiveness of current speech separation methods for assistive hearing devices as well as neuroprosthetic devices such as cochlear implants (CIs). Here, we investigate whether a deep neural network model for speech separation can utilize the spatial information in naturalistic listening scenes as captured by a CI’s microphones to improve separation performance. We examined the impact of latent spatial cues (inherently present in two-channel speech mixtures, but need to be learned from these mixtures), as well as pre-computed spatial cues added to the speech mixtures as auxiliary input features (inter-channel level and phase differences, ILDs and IPDs). Specifically, we introduce a two-channel version of the SuDoRM-RF speech separation model, which takes as input speech mixtures recorded with two CI microphones and shows that latent spatial cues enhance separation performance without affecting model efficiency in terms of model complexity and inference latency. Pre-computed spatial cues – especially IPDs – enhanced separation performance even more, but simultaneously reduced model efficiency. Finally, simulating a CI user’s listening experience with a vocoder showed that the beneficial effect of spatial cues on DNN speech separation persists even if the separated speech streams are spectrotemporally degraded as in the output of a CI.

Background: Cochlear implant users (CI) experience degraded speech intelligibility (SI) in noise. A contralateral hearing aid (HA) benefits listening in noise, and directional microphones and directivity steering can further enhance SI.Objectives: Four directional microphones were evaluated for bimodal listeners, including a monaural and binaural broadband front-facing beamformer (BFM and BFB) and a novel monaural and binaural narrowband side-beamformer (BFMS, BFBS). Broadband directivity steering (DS) was also evaluated. Methods: BFM, BFB, and DS were tested in diffuse noise with speech administered frontally or to the HA side (S0Ndif and SHANdif). BFMS and BFBS were tested with speech presented on the CI side and noise on the HA side or vice versa (SCINHA, SHANCI). SI was assessed by determining speech reception thresholds (SRTs). Results: In the S0Ndif condition, BFB significantly improved SI by 2.9 dB SNR and BFM by 1.3 dB SNR, but not significantly. In the SHANdif condition, BFM and BFB did not benefit SI, but DS improved SI by 3.6 dB. BFBS significantly improved SI by 1.1 dB SNR, but BFMS was ineffective. Conclusion: Binaural beamformers outperform their monaural counterparts. The novel BFBS is a promising technique for bimodal listeners.

Objectives:
The substantial variability in speech perception outcomes after cochlear implantation complicates efforts to develop valid predictive models of these outcomes. Existing predictive regression models are too unreliable for clinical application, possibly because speech intelligibility (SI) after cochlear implant (CI) rehabilitation is often based on a limited number of assessments. The development of SI after CI has rarely been detailed, although knowing the shape of the learning curve can potentially improve predictive modeling. Knowing the learning curve after CI could also aid in setting expectations about SI immediately after implantation, and the duration of rehabilitation. The current objectives were to construct learning curves to estimate baseline SI at 1 week (B), maximal SI after rehabilitation (M), and rehabilitation time (time to reach 80% of the learning effect; t[M − B]80%), and to subsequently deploy these outcomes for multiple-regression modeling to predict CI outcomes.

Design:
To assess rehabilitation after cochlear implantation, we retrospectively fitted learning curves using clinically available SI assessments from 533 postlingually deaf, unilaterally implanted adults. SI was assessed with consonant-vowel-consonant words (CVC) in quiet, with phoneme score as the outcome measure. Participants were followed for up to 4 years, with SI measurements collected at fixed intervals. SI was commonly assessed 1, 2, 4, and 8 weeks after device activation. B, M, and t(M − B)80% were determined from the fitted learning curves. Predictive multiple-regression analyses were performed on these three outcome measures based on eight previously identified preoperative demographic and audiometric predictor variables: age at implantation, duration of severe-to-profound hearing loss, best-aided CVC phoneme score (in the free field), unaided ipsilateral and contralateral residual hearing and CVC phoneme scores (measured with headphones), and education type (regular or special education).

Results:
At 1 week after CI activation, raw phoneme scores had increased from 40% preoperatively (best-aided condition) to 51%, with further improvement to approximately 78% at 4 years. SI increased significantly until 1 year after activation and then plateaued. Fitted learning curves supported better estimates of these parameters, showing that average baseline SI at 1 week after CI activation was 51%, increasing to 85% after rehabilitation. The asymptotic score exceeded the raw average after 4 years because many cases had not yet plateaued. The median t(M − B)80% was 1.5 months. Predictive modeling identified duration of hearing loss, age at implantation, best-aided CVC phoneme score, and education type as the most robust predictors for postoperative SI. Despite the statistically significant correlations, however, the combined predictive value was ~19% for B, 10% for M, and 2% for t(M − B)80%.

Conclusions:
This study is among the few to generate detailed learning curves after cochlear implantation. By including clinical SI measures in the earliest rehabilitation period, we report a median rehabilitation time with CI of 1.5 months. This implied rapid learning effect emphasizes the value of monitoring SI in the first few weeks after rehabilitation. According to multiple-regression analyses, the most commonly used preoperative variables correlated significantly with postoperative outcomes, but with limited predictive value for the clinic. By fitting learning curves through data reported in the literature, we show that the increase in SI during rehabilitation is an important predictor for t(M − B)80%.

Objective: Measuring listening effort using pupillometry is challenging in cochlear implant (CI) users. We assess three validated speech tests (Matrix, LIST, and DIN) to identify the optimal speech material for measuring peak-pupil-dilation (PPD) in CI users as a function of signal-to-noise ratio (SNR). Design: Speech tests were administered in quiet and two noisy conditions, namely at the speech recognition threshold (0 dB re SRT), i.e. the SNR where speech intelligibility (SI) was 50%, and at a more favourable SNR of +6 dB re SRT. PPDs and subjective ratings of effort were obtained. Study sample: Eighteen unilaterally implanted CI users. Results: LIST sentences revealed significantly different PPDs between +6 and 0 dB re SRT and DIN triplets between quiet and +6 dB re SRT. PPDs obtained with the Matrix test were independent of SNR and yielded large PPDs and high subjective ratings even in quiet. Conclusions: PPD is a sensitive measure for listening effort when processing LIST sentences near 0 dB re SRT and when processing DIN triplets at more favourable listening conditions around +6 dB re SRT. PPDs obtained with the Matrix test were insensitive to SNR, likely because it is demanding for CI users even in quiet.

Objectives: Cochlear implants (CIs) are the primary treatment for severe-to-profound hearing loss. For CI users, speech intelligibility (SI) is often excellent in quiet yet degrades dramatically in background noise. Scientific and clinical testing of the effects of noise on SI is routinely performed with speech-in-noise tests. The sensitivity of these tests to signal to noise ratio depends on the slope of their psychometric curve. This slope is not always known for CI users, and direct comparisons between typical hearing (TH) listeners and CI users are lacking. Design: We present a comparative study of a digit test (DIN), a Matrix sentence test, and an everyday sentence test (LIST) for a group of CI users and TH listeners, with use of word (digit) and sentence (triplet) scoring in the free field. We report descriptive statistics and effect size measures of the psychometric slope and the speech reception threshold (SRT) for each speech test. Results: For CI users, the slopes of the psychometric curve were significantly shallower and SRTs significantly higher than those of TH listeners. The shallowest slope was seen with the Matrix test. However, the small variances of the slope and the SRT resulted in effect size estimates that fell between those of the other two tests. The DIN test was associated with steeply sloped psychometric curves with low variance. The scoring method did not substantially affect slopes and SRTs for the DIN test and LIST sentences, but word scoring resulted in shallow slopes and substantially worse SRTs for CI users. Conclusions: The DIN test stood out in this study as an attractive speech-in-noise test for CI users, with steep slopes and low variance in slopes and SRTs among participants. Digit and keyword scoring appear to be viable options for the DIN test and LIST sentences, respectively, potentially increasing the number of available test items. For the Matrix test, sentence scoring yielded shallow slopes and deteriorated SI, especially for the CI group. We recommend word scoring for the Dutch-Flemish Matrix test.

Introduction: Cochlear implantation is the standard treatment for severe to profound hearing loss. While cochlear implant (CI) users can communicate effectively in quiet environments, speech understanding in noise remains challenging. Bimodal hearing, combining a CI in one ear and a hearing aid (HA) in the other, has shown advantages over unilateral electrical hearing, especially for speech understanding in noisy conditions. Beamforming is a technique used to improve speech understanding in noise by detecting sound direction and enhancing frontal (speech) sounds while attenuating background noise. One specific beamformer, Stereozoom, combines signals from microphones in both ears to create a focused beam toward the front resulting in a binaural beamformer (BB), in order to improve speech intelligibility in noise for bilateral and bimodal CI users. Methods: A prospective crossover study involving 17 bimodal CI users was conducted, and participants were tested with various device configurations (CI, HA, CI + HA) with and without BB. Speech recognition testing with the Dutch/Flemish matrix test was performed in a sound-attenuated booth with diffuse noise to simulate realistic listening conditions. Results: The results showed a statistically significant benefit of bimodal hearing over the CI configuration and showed a statistical significant benefit of BB for the CI and CI + HA configuration. The benefit of BB in the HA configuration was not statistically significant probably due to the higher variance. The benefit of BB in the three configurations did not differ statistically significant. Conclusion: In conclusion, bimodal hearing offers advantages for speech understanding in noise for CI users. BB provides a benefit in various device configurations, leading to improved speech intelligibility when speech comes from the front in challenging listening environments.

Objectives:
We investigated whether listening effort is dependent on task difficulty for cochlear implant (CI) users when using the Matrix speech-in-noise test. To this end, we measured peak pupil dilation (PPD) at a wide range of signal to noise ratios (SNR) by systematically changing the noise level at a constant speech level, and vice versa.

Design:
A group of mostly elderly CI users performed the Dutch/Flemish Matrix test in quiet and in multitalker babble at different SNRs. SNRs were set relative to the speech-recognition threshold (SRT), namely at SRT, and 5 and 10 dB above SRT (0 dB, +5 dB, and +10 dB re SRT). The latter 2 conditions were obtained by either varying speech level (at a fixed noise level of 60 dBA) or by varying noise level (with a fixed speech level). We compared these PPDs with those of a group of typical hearing (TH) listeners. In addition, listening effort was assessed with subjective ratings on a Likert scale.

Results:
PPD for the CI group did not significantly depend on SNR, whereas SNR significantly affected PPDs for TH listeners. Subjective effort ratings depended significantly on SNR for both groups. For CI users, PPDs were significantly larger, and effort was rated higher when speech was varied, and noise was fixed for CI users. By contrast, for TH listeners effort ratings were significantly higher and performance scores lower when noise was varied, and speech was fixed.

Conclusions:
The lack of a significant effect of varying SNR on PPD suggests that the Matrix test may not be a feasible speech test for measuring listening effort with pupillometric measures for CI users. A rating test appeared more promising in this population, corroborating earlier reports that subjective measures may reflect different dimensions of listening effort than pupil dilation. Establishing the SNR by varying speech or noise level can have subtle, but significant effects on measures of listening effort, and these effects can differ between TH listeners and CI users.