The sample size was computed based on the measure of subjective listening effort. Power analysis was performed using R software (version 4.3.1) based on the simulation of the data from five pilot participants. A linear mixed effects model was fitted to the data and the mixed power function from the mixedPower package (Kumle et al., 2021) was applied to estimate power for different sample size. The simulation was based on 1,000 runs with a varying sample size from 18 to 26. To ensure that a smaller significant effect size can also be detected, a new parameter SESOI (smallest effect size of interest) was added to the mixedPower function. The simulation results were based on the dataset as well as the defined estimates at the same time. Results suggested that a sample size of 20 subjects should be sufficient with a power over 0.8.

Twenty-five Mandarin-speaking adults (9 males and 16 females) aged between 26 and 77 years old (M = 46.5, SD = 18.4) were recruited. All participants were diagnosed with sloping severe-to-profound hearing loss. The average pure-tone audiometric thresholds between 500 and 4,000 Hz for both ears were above 65 dB HL. Figure 1 shows the group average pure-tone thresholds with standard deviation. All participants met the following inclusion criteria: (1) symmetric sloping sensorineural hearing loss (i.e., interaural difference ≤ 15 dB at all octave frequencies from 250 to 8,000 Hz) with air-bone gaps at each frequency ≤ 15 dB; (2) normal middle ear function as indicated by tympanometry and otoscopy examinations; (3) normal cognitive ability screening with MoCA-basic version of Chinses (passing score: 23/30; Chen et al., 2016). (4) At least 1 year hearing aid usage. (5) native Mandarin speakers in daily life. Among the 25 participants, 17 used Phonak brand hearing aids, 4 used Oticon, 2 used Resound, and 2 used Starkey.

During the study, each participant wore a pair of Phonak (Stäfa, Switzerland) Naida Paradise 90 UP hearing aids programmed in Phonak Target fitting software (v. 7.1) with Noahlink Wireless. To ensure proper amplification of sound and to limit potential acoustic feedback, custom hard silicone ear molds were made, and the vent size was selected based on the recommendation of the fitting software. The feedback measurement was then performed using the “Real ear and feedback measurement” module in the Phonak Target fitting software to ensure that there was no feedback problem or gain limitation. The APD-Contrast (Adaptive Phonak Digital Contrast) 2.0 fitting prescription was chosen as the target. The initial gain level was set to 100%. Fine-tuning was accomplished by playing a practice list of CMNmatrix (Hu et al., 2018) sentences at 65 dBA SPL from a speaker 1 meter in front of the participant. If the loudness was not appropriate, the broadband frequency gain for the 65 dB SPL input level was adjusted until the participant was satisfied. If participants reported an occlusion problem, the “Occlusion Compensation” feature was turned on and adjusted according to their feedback. To avoid touching by mistake, the button function for volume adjustment was disabled. For tests with speech in noise (SpiN) program, three manual programs based on SpiN were created. These three manual programs differ only in NR strength, namely off, moderate (12) and strong (20) accordingly. For tests with speech in loud noise (SpiLN) program, the same three manual programs were created similarly. These two different base programs were tested in separate appointments. Except from NR setting, the other advanced features such as noise reduction, directionality, etc. were kept the same within each base program.

The NR algorithm used in this study (Dynamic Noise Cancelation, DNC) is a spatial noise reduction feature that works in combination with a directional beamformer to improve the signal-to-noise ratio in challenging situations. It is microphone dependent and is only activated when the beamformer is fully activated and switches off as soon as the SNR reaches 18 dB and higher. NR can be adjusted from off to weak, moderate and strong levels in Phonak Target and in myPhonak APP. In Target, the NR appears on a 20-point slider: 0 is off, 1–9 is weak, 10–16 is moderate and 17–20 is strong. A stronger NR setting leads to higher attenuation of noise, which in turn provides a higher SNR benefit (Figure 2).

All participants were tested at three separate sessions to complete all the measures within a 2-week period. The interval between each visit was 3 to 4 days. Tests were conducted in a sound booth with background noise below 30 dB A in Beijing Tongren Hospital. Twelve speakers (Audioengine 2⁺) were arranged in a circle with a radius of 1.4 m. The touchscreen monitor (Philips) used in listening effort and sound quality rating to present response interface was 50 cm long and 30 cm wide.

In the first session, after obtaining the informed consent from participants, otoscopy, pure tone audiometry (both air conduction and bone conduction) and tympanometry were implemented to check hearing condition, followed by MoCA screening. Individual ear impression was taken at last. In the second visit, speech recognition in noise test, subjective listening effort rating and sound quality assessment were conducted for three manual programs of different NR strengths (NR_off, NR_mod, NR_str) with SpiN as the base program. In the third visit, all the tests and settings were kept the same as those in the second visit, except that SpiLN was used as the base program to create manual programs. In these two sessions, the order of three outcome measures was fixed, but the order of NR strength conditions was randomized within each measure. Break was scheduled between tests or whenever required by the participants.

The above three outcomes were measured repeatedly with three NR strengths under each of the two HA programs (SpiN and SpiLN). For all three measures, data of SpiN and SpiLN was analyzed separately. The individual rated listening effort was fitted as a two-slope function of the SNR using the BX fitting method (Oetting et al., 2014) which was recommended by Krueger et al. (2017). The data points of the rating category “only noise” were excluded. The SNR values at “moderate effort” (7 ESCU-Effort Scaling Category Units) for each participant and condition were extracted. A repeated measures analysis of variance (rmANOVA) with SNRs at “moderate effort” as the dependent variable and NR strength condition as the independent variable was performed to see if SNRs required to reach “medium effort” were different between NR strength conditions. Data of speech recognition in noise test was analyzed by the repeated measures one-way ANOVA with SRT50 as the dependent variable and NR strength conditions as the independent variable. For sound quality assessment, rating scores were averaged across the test and retest for analysis. Repeated measures one-way ANOVA with rating scores as the dependent variable and NR strength conditions as the independent variable was performed for each of the four attributes. Friedman’s test was used instead if the normal distribution assumption was not met. Statistical analyses were conducted using R software (version 4.3.2; R Core Team, 2021) with “bruceR” package (Bao, 2023) for ANOVA and “ggplot2” package (Wickham, 2016) for graphs.

Figure 7 shows boxplots of SRTs of three NR strength levels for the SpiLN and for SpiN programs. The mean SRTs of −3.0, −3.7, and −3.5 dB were obtained in the three NR conditions for the SpiLN program, respectively. Great variability in results was observed, with SDs of 5.5 dB, 5.1 dB, and 4.9 dB, respectively, in the three NR conditions. The median SRTs of −2.8, −2.9, and −3.0 dB, obtained in the three NR conditions for condition of SpiN, respectively. SRT50s were compared between three NR levels (NR_off, NR_mod, NR_str) in two programs (SpiN or SpiLN) separately. SRTs in each test condition of SpiLN program were normally distributed as verified by the Shapiro–Wilk test. There is no significant effect of noise reduction (NR), F(2, 42) = 1.133, p = 0.332. For SpiN program, SRTs in NR_mod were not normally distributed. Therefore, a non-parametric test Friedman Test was applied, which also indicated no significant effect of NR, χ²(2) = 3.7, df = 2, p = 0.157.

Figures 8, 9 show estimated two-slope functions of listening effort for each individual and the group mean of the results for three NR strengths within the SpiN program an SpiLN program, respectively. The mean listening effort function was calculated by averaging the SNRs for each rating category over all participants. In both HA programs, there is a large amount of variation in the participant’s individual listening effort function. For the group mean function, the perceived listening effort decreased with increasing SNR. Across the entire range of listening effort scaling, the NR_off required higher SNR values to achieve similar effort ratings compared to the NR activate conditions (NR_mod and NR_str). Such advantages of NR in SpiN program appeared to be more pronounced roughly between “little effort” (5 ESCU) to “medium effort” (7 ESCU), but more discrete for “no effort” (1 ESCU) and “extreme effort” (13 ESCU). As for SpiLN program, NR activated conditions (NR_mod and NR_str) provided more advantages roughly from “medium effort” (7 ESCU) to “considerable effort” (9 ESCU). In terms of quantitative analysis, in Figure 10, the SNR values at 7 ESCU were extracted because it was the midpoint of the smooth area of the two-slope listening effort function. For SpiN program, the mean SNR at 7 ESCU was 0 dB for the NR_off, −1.3 dB for the NR_mod, and −1.1 dB for the NR_str with standard deviations of 3.3, 3.1, and 3.3 dB, respectively. There is a significant effect of noise reduction (NR) settings. Specifically, the difference from NR_off was significant for NR_mod [t(22) = −3.032, p = 0.018] but not for NR_str [t(22) = −2.23, p = 0.11]. No significant difference was found between NR_mod and NR_str [t(22) = 0.36, p = 1]. In the SpiLN program, the mean SNR at 7 ESCU was −1.1 dB for NR_off, −3.3 dB for NR_mod, and −3 dB for NR_str, with standard deviations of 4.0 dB, 3.8 dB, and 4.2 dB, respectively. Significant differences from NR_off were observed with both NR_mod [t(24) = −6.17, p < 0.001] and NR_str [t(24) = −6.08, p < 0.001]. No significant difference was found between NR_mod and NR_str [t(24) = 1.03, p = 0.944].

Figure 11 plots the average sound quality ratings from the 24 participants in the three NR strengths. One participant’s data was missing. For the rating of noise suppression in the SpiLN program, the analysis yielded significant main effect of NR [F(1.461, 33.594) = 18.43, p < 0.001]. The participants rated noise suppression significantly higher with NR_str than with NR_off [t(23) = 4.78, p < 0.001] and with NR_mod [t(23) = 2.95, p < 0.05]; NR_mod was rated significantly higher than NR_off [t(23) = 4.39, p < 0.001]. For SpiN program, the main effect of NR was significant [F(1.344, 30.908) = 18.57, p < 0.001]. The participants rated noise suppression significantly higher with NR_str [t(23) = 4.45, p < 0.001] and NR_mod [t(23) = 5.19, p < 0.001] than with NR_off.

In terms of listening comfort and satisfaction ratings, the main effect of NR was significant for both the SpiLN program [Listening comfort: F(2, 46) = 11.35, p < 0.001; Satisfaction: F(1.308, 30.094) = 12.87, p < 0.001] and the SpiN program [Listening comfort: F(2, 46) = 16.56, p < 0.001; Satisfaction: χ²(2) = 17.58, p < 0.001]. For SpiLN program, participants rated significantly higher listening comfort and satisfaction with NR_str [Listening comfort: t(23) = 3.35, p < 0.01; Satisfaction: t(23) = 3.45, p < 0.01] and with NR_mod [Listening comfort: t(23) = 4.38, p < 0.001; Satisfaction: t(23) = 4.80, p < 0.001] than with NR_off. For SpiN program, both sound attributes were rated significantly higher with NR_str [Listening comfort: t(23) = 4.81, p < 0.001; Satisfaction: Z = 37, p < 0.01] and NR_mod [Listening comfort: t(23) = 4.41, p < 0.001; Satisfaction: Z = 50, p < 0.01] than with NR_off.

For clarity ratings, the analysis revealed a significant effect of NR for both the SpiLN program [F(2, 46) = 9.09, p < 0.001] and the SpiN program [F(2, 46) = 15.36, p < 0.001]. Participants rated significantly higher speech clarity with NR_str than with NR_off for both the SpiLN program [t(23) = 4.26, p < 0.001] and the SpiN program [t(23) = 4.68, p < 0.001]. The speech clarity was rated significantly higher with NR_mod than with NR_off only in the SpiN program [t(23) = 3.84, p < 0.01].

No significant difference in speech intelligibility was observed between HA settings with and without NR. While the NR algorithm utilized in this study did not lead to improved speech recognition, it is noteworthy that it did not negatively affect performance, which has also been observed with certain other NR algorithms (Ricketts and Hornsby, 2005). This is consistent with our hypothesis that speech recognition in noise may not be improved by NR processing, but neither is degraded. Previous research has suggested that speech degradation caused by high levels of noise or distorted speech cues could lead to delays in lexical processing (Ben-David et al., 2011; McMurray et al., 2017). On the one hand, the application of NR processing could attenuate a certain amount of background noise, but on the other hand it would introduce some distortion of speech cues (Kates, 2008). If the level of speech cue distortion is high enough to counteract the effect of noise reduction, listeners may not benefit from NR due to the increased cognitive workload in speech processing (Winn et al., 2015; Ward et al., 2017).

For individuals with mild to moderate hearing loss, several studies have reported improvements in speech intelligibility with NR algorithms (Peeters et al., 2009; Ohlenforst et al., 2018; Wong et al., 2018). Ohlenforst et al. (2018) found that sentence recognition performance was significantly better when NR was enabled than when it was not. The study by Wong et al. (2018) revealed that when NR was activated, listeners demonstrated improved speech perception abilities and reported reduced noise annoyance, resulting in improved speech clarity. These findings have been attributed in part to predictable relationships between degree of impairment and auditory-cognitive abilities in individuals with mild to moderate hearing loss (Humes, 2007; Souza et al., 2007). These associations diminished for listeners with more severe hearing loss, whose abilities became more variable and unpredictable.

In this study, NR settings on the hearing aids, being new settings for the participants, might have been perceived as unfamiliar input. This unfamiliarity could lead to an increased reliance on working memory, as participants needed to utilize more cognitive resources to resolve the unfamiliar stimuli (Souza et al., 2015; Ronnberg et al., 2019, 2022). The speech tests employed in this study included cafeteria noise and the International Female Fluctuating Masker (IFFM), incorporating both energetic and informational maskers. These maskers pose distinct challenges: energetic maskers distract, while informational maskers impact more significantly as they engage semantic long-term memory (SLTM; Sorqvist and Ronnberg, 2012). In situations involving informational maskers, working memory capacity (WMC) is a crucial predictor of performance. The findings suggest that high dependency on working memory was not mitigated by prolonged use of hearing aids; a decade of hearing aid use does not diminish this reliance, especially with four-talker (4 T) maskers (Ng and Ronnberg, 2020). Additionally, the study’s speech recognition materials were akin to Hagerman matrix sentences. Ronnberg et al. (2016) found that working memory had a stronger correlation with Hagerman matrix sentences than with HINT sentences, which are driven by everyday contexts. This underscores the significant role of working memory in complex listening environments. Hearing aid users with severe-to-profound hearing loss often exhibited constraints in working memory capacity, which might impair their capability to process and decode speech signals following NR processing. Although NR technology effectively suppressed background noise, this suppression might not adequately compensate for the challenges encountered by these individuals in processing speech information and memory tasks. Consequently, focusing on traditional speech recognition tests might not fully capture the overall improvements in communication achieved through NR technology. Additionally, these tests might lack sensitivity to the enhancements in NR’s top-down processing. Therefore, it was essential to supplement these assessments with subjective evaluations and measures of listening effort to comprehensively evaluate the influence of NR.

This study investigated the effect of NR schemes on listening effort as indicated by a subjective method in a group of people with sever-to-profound hearing impairment. Our results suggested that the NR processing reduced the listening effort required in noisy conditions for individuals with severe-to-profound hearing loss. Participants preferred either NR_mod or NR_str over NR_off, were consistent with previous work demonstrating a reduction in listening effort but no improvement in speech intelligibility when enabling NR (Desjardins and Doherty, 2014; Ohlenforst et al., 2018). In addition, several other studies had shown that participants experienced reduced listening effort when using NR in noisy environments (Sarampalis et al., 2009; Brons et al., 2014; Wong et al., 2018). Despite variations in specific details, overall, these findings confirmed that listening effort evaluations could sensitively capture listeners’ perceived effort.

Kestens et al. (2021) undertook a systematic review to investigate how intersubject differences in cognition could influence the aided benefit for speech understanding and listening effort with bilateral digital hearing aids. The research results demonstrated that the effects of noise reduction technology varied among hearing aid users with different cognitive abilities. For instance, the study conducted by Desjardins and Doherty (2014) did not find a significant relationship between working memory and listening effort when using noise reduction technology. In contrast, Wendt et al. (2017) found that users with better working memory capacities experienced less objectively measured listening effort compared to HA users with less memory capacity when utilizing the noise reduction feature. Furthermore, users with slower processing speeds appeared to benefit more from noise reduction technology in terms of listening effort, indicating that they relied more on hearing aids when dealing with complex listening situations (Desjardins and Doherty, 2014). Although noise reduction technology is expected to theoretically help reduce listening effort, the existing research findings remains inconsistent. This variability in outcomes could be due to the different measurement methods used in various studies, the differences in cognitive abilities among participants, and variations in hearing aid technologies.

Researches suggested that NR might not necessarily enhance speech clarity, but rather it could reduce cognitive effort, a benefit that becomes particularly apparent in listening effort (Sarampalis et al., 2009; Desjardins and Doherty, 2014; Wong et al., 2018). In the patient-centered model, testing for listening effort is crucial. The goal of rehabilitation is to reduce the degree of barriers, activity limitations, and participation restrictions for hearing aid users (World Health Organization, 2001). Subjective listening effort seemed to be related to perceived hearing impairment (Alhanbali et al., 2018), and subjective listening effort could have a negative impact on the quality of life (Pichora-Fuller et al., 2016).

Overall, the activation of NR provided clear advantages compared to its deactivation across several dimensions, including enhanced noise tolerance, improved listening comfort, increased satisfaction, and better perceived speech signal quality. These trends showed a more favorable and comfortable auditory experiences when noise was selectively attenuated and fine-tuned by NR algorithms.

The findings of our study were consistent with observations that a majority of individuals with hearing impairments tended to employ NR features in their hearing aids (Souza et al., 2018). Clearly, with noise reduction activated, speech and noise that had undergone noise reduction processing became more comfortable for listeners due to the reduction in noise levels. Overall, these subjective assessments of NR evaluation demonstrated positive effects in terms of noise acceptance, user preferences, and enhanced listening comfort. Numerous studies have previously indicated that noise reduction had a positive impact on noise perception and preferences. Studies utilizing the Abbreviated Profile of Hearing Aid Benefit (APHAB; Cox and Alexander, 1995) have consistently shown that noise reduction has a notable impact on reducing aversion to noise (Boymans and Dreschler, 2000; Palmer et al., 2006). Brons et al. (2013) demonstrated that activating noise reduction reduced annoyance caused by noise, although it found no impact on speech naturalness through paired comparison testing. Palmer et al. (2006) discovered that when NR was activated, the annoyance ratings of hearing-impaired participants closely resembled those of individuals with normal hearing. These findings support that noise reduction has the potential to enhance listening comfort and reduce noise-related annoyance. In this study, participants reported improved speech clarity when NR was activated. This apparent discrepancy between objective speech intelligibility and subjective speech clarity ratings may stem from the different conditions under which each was tested. Specifically, speech intelligibility was assessed using an adaptive procedure that varied the signal-to-noise ratio (SNR) within a negative range. It is documented that NR algorithms are more effective at positive SNRs than at negative ones, which could influence these outcomes (Fredelake et al., 2012; Smeds et al., 2015). Furthermore, the benefits in terms of listening ease and comfort found in this study did not coincide with any reduction in speech intelligibility.