Preferred Strength of Noise Reduction for Normally Hearing and Hearing-Impaired Listeners

Participants

Approval by the Medical Ethical Committee of AMC was obtained on 24 April 2008 (MEC 08/082). Participants were recruited from the patients of the Audiological Centre of the AMC and had to sign an informed consent prior to participation.

There were three groups of 10 participants in the study. The audiometric data are given in Appendix A. Two additional participants quit the experiment early stating that they did not hear any difference between the stimuli. Although their incomplete data sets were not used for analysis, their audiometric data are included in Appendix A for completeness.

The included number of participants was based on a power calculation with data from NH listeners from a previous study (Houben et al., 2011a). In that study, mean preference for NR strength was 7.5 dB with standard deviation of ±1.8 dB. If we assume the same measurement variance for our groups of participants, at least eight participants are required in each participant group to detect a difference in preference of 2.5 dB (between-participant design α = 0.05, β = 80, two-sided).

The audiometric inclusion criteria for the three hearing loss categories were as follows:

- NH: all hearing thresholds (octave frequencies ranging from 0.25 to 8 kHz) equal to or better than 20 dB hearing loss (HL) for both ears.

The speech reception threshold (SRT) for consonant–vowel–consonant (CVC) words in quiet was also measured. The SRT is the sound pressure level at which the participant correctly repeated 50% of the phonemes. For the first five NH participants, no SRT measurement is available due to a measurement mistake. The mean age (with standard deviation) of the three groups was 47 ± 12 years, 61 ± 15 years, and 67 ± 7.5 years, for NH, HI-mild, and HI-moderate, respectively.

Stimuli

The unprocessed stimuli were recordings of four sentences spoken by a female, taken from the VU-98 sentence materials (Versfeld et al., 2000). During the development of this Dutch speech-in-noise test, the sentences were optimized to have near-equal intelligibility in stationary noise (Versfeld et al., 2000). The sentences are, however, not homogenized for perceived listening comfort. Therefore, from these materials, four sentences (#49, 52, 58, 63) were selected to minimize possible differences in subjective preferences between different sentences. Sentence selection was done based on perceived listening comfort. The sentences were embedded in babble noise of which the long-term spectrum was matched to that of speech of the sentence materials. The sentences were rated for perceived listening comfort by three expert listeners. Perceived listening comfort was used rather than sound quality to avoid the (explicit or implicit) reference that plays a role when judging sound quality. Perceived listening comfort was rated on a 5-point scale ranging from not comfortable to very comfortable. The four sentences were selected to have equal scores and average listening comfort.

NR processing was done using a low-latency single-microphone NR algorithm (SNRA), implemented in Matlab. This algorithm has been used before to measure individual preferences for the strength of NR (Houben et al., 2011a, 2011b, 2013). The algorithm uses modulation-based spectral subtraction and has low complexity and low latency (Appleby & Groth, 2011; Groth & Nelson, 2005; Kates, 2017; Rosenstrauch, 2011). A detailed description of the algorithm can be found in Houben et al. (2011a, 2011b, 2013). Here, the algorithm is briefly described.

The incoming speech signal is first split into a signal and an analysis path. For the signal path, all filtering and processing is done in the time domain, and for the analysis path, this is done in the frequency domain (Groth & Nelson, 2005). The two paths are each analyzed with a 17-band frequency-warped filter bank (Kates & Arehart, 2005). The advantage of a frequency-warped filter bank over a conventional filter bank is that it has a non-uniform frequency representation very close to that of the auditory system.

The SNR is estimated for each frequency band using the estimation of the noise and speech signal (Rosenstrauch, 2011). During intervals in which no speech is detected, the input signal is used to calculate a noise estimate. The estimation of the speech signal is based on spectral and temporal characteristics of speech, using samples of the input signal of about 1 ms. The gain is calculated in the analysis path using the estimated SNR and Wiener optimal filtering theory (Smith, 2002), with a threshold for gain depth (=G_min). This threshold was used to limit the strength of the NR. The calculated gain was subsequently applied to the signal in the signal path. The variable under investigation, the threshold for gain depth (G_min), limits the maximum gain reduction that is applied by the SNRA to G_min. Higher values of G_min led to more gain reduction. The variable G_min can be applied independently from the NR algorithm since it does not alter the gain function but only applies a threshold of maximum gain reduction. In the literature on spectral subtraction, this limit is also known as “spectral floor” (Berouti et al., 1979; Loizou, 2007).

Figure 1 shows the estimated realized gain of our algorithm as function of the estimated input SNR. The processed sentences were analyzed based on the four sentences that were chosen for this experiment. These sentences were placed one after another and processed by the NR algorithm in loops such that at least 60 s of the processed output file could be removed to stabilize the algorithm. The gain that was applied by the algorithm was calculated by comparing the processed (output) signal to the input signal for separate time-frequency bins (window length was 512 samples, with a sample frequency of 16 kHz). Note that due to estimation errors, the maximum gain in this plot can be larger than G_min.

OPEN IN VIEWER

Figure 1.

Calculated gain function for some values of G_min used. The gain was calculated by comparing the processed and unprocessed sound signals for one sound file which consists of all four sentences concatenated. The striped vertical line at +5 dB indicates the long-term average SNR of the stimuli.

By removing the estimated noise, the loudness of the speech-in-noise signal is inevitably affected. To correct for this, NR algorithms commonly apply a correction that restores the overall gain. Our algorithm restores the gain by matching the Root mean square (RMS) of the output signal to that of the input. The correction is thus done on the overall speech-in-noise signal.

Experimental Design

At the start of the visit, hearing status was checked for each participant by means of pure-tone audiometry. Following audiometry, they participated in a paired comparison listening test which is explained in the next two paragraphs.

A between-participants design was used, the three hearing loss groups each containing 10 participants. The main experimental parameter was G_min, which is explained in the previous section. Ten values of G_min were used (0, 4, 6, 7, 8, 9, 10, 12, 16, 18 dB), each of which was compared to all the others, leading to 10 × 9 unique comparisons (including AB, BA to prevent bias in the presenting order). Each comparison was done twice (test and retest, without a pause in between) leading to 180 paired comparisons per participant. This value was a trade-off between measurement accuracy (more runs are better) and total time required to do the experiment (less is better). On average, a session took about 1.5 h.

In each trial, participants had to listen to two successive processed speech-in-noise stimuli and choose which they preferred. The participants were asked the following question “Imagine that you will have to listen to these signals all day. Which sound would you prefer for prolonged listening?” The question was intentionally stated in a broad context. The reason is that we were primarily interested in general preference and not in perceived specific signal qualities such as “speech quality” or “amount of background noise.” Before commencing with the paired comparisons listening test, a training session was performed with 20 paired comparisons with similar stimuli. The data obtained during the training session were discarded.

The experimental set-up consisted of a Focusrite Saffire Pro 10 audio interface and a Presonus HP40 headphone buffer with Sennheiser HDA200 headphones. All stimuli were presented bilaterally with an input SNR of +5 dB. This value was chosen because it is representative of real-life situations, which is shown in a field study by Wu et al. (2018) who found that most realistic SNRs were between 2 and 14 dB. We chose a lower value of this range such that it is sufficiently challenging for both the listener and the NR algorithm and in line with research by Houben et al. (2013). The nominal speech level was 70 dB Sound pressure level (SPL), and the noise level was 65 dB SPL. For the participants in the HI-mild and HI-moderate groups, the sound signals were amplified according to the National Acoustics Laboratories-Revised Profound (NAL-RP) fitting rule (Dillon, 2001). The sound signals for the left and right ear received the same NAL-RP amplification, based on the ear with the smallest hearing loss. This approach was possible because the asymmetry in the hearing loss was small; see Appendix A.

Statistical Analysis

Data-Driven Analysis of Preference

The second research question was to define the required number of distinct NR settings to be able to classify our participants into similar groups based on their NR preference. To answer this question a cluster analysis was performed. Figure 6 shows the results of the hierarchical cluster analysis The dendrogram is shown on the left. The horizontal distance on the dendrogram represents the calculated Manhattan (L1) distance between the merged clusters.

OPEN IN VIEWER

Figure 6.

Hierarchical cluster analysis of the win-count data. The different gray scales represent the win counts (white is low, black is high).

From left to right, the dendrogram shows the possible different levels with different numbers of clusters starting with two clusters in the first level and finishing with 30 separate clusters, one for each participant (see James et al., 2013 for a tutorial on hierarchical clustering and dendrograms). In the dendrogram, we can distinguish two stable clusterings: one with two clusters and one with three clusters. Stable clusterings are those that persist over a large horizontal range in the dendrogram. The clustering with two clusters has one group with six participants that prefer no NR and one group with 24 participants that prefer some non-zero level of NR. This clustering supports the choice that some hearing aid manufacturers make: NR is either on or off. The other stable clustering is one with three clusters. This clustering consists of the same group of six participants that prefer no NR and separates the remaining 24 participants into 9 and 15 participants that differ in the strength of NR they prefer. The next (unstable) clustering with five clusters does not provide us with additional insight. We focus on the clustering with three clustering for the rest of this manuscript. For each of these three clusters, the middle part of Figure 6 shows the win-count data for each participant. The different gray scales represent the win counts (white is low, black is high). Participant identity is indicated at the right of the figure. The clusters can be interpreted as one group that prefers zero NR (coded orange in Figure 6), one group that prefers an intermediate level of NR (light blue in Figure 6), and one group that prefers a high level of NR (green in Figure 6).

The average win-count data for each of the identified relevant clusters with respect to G_min gives us information on the differences in preferences between the different clusters throughout the range of G_min. Figure 7 shows the average win-count data (with standard deviation). Note that these data were derived directly from the participants’ answers, without information on the hearing loss or hearing loss group.

OPEN IN VIEWER

Figure 7.

Cluster analysis: win count for each of the three clusters is calculated by averaging the win counts of each of the participants in a cluster. Error bars denote the standard deviation across participants.

The preference in Cluster 1 (shown in orange) is for low G_min (with mean G min opt = 0 dB), the preference in Cluster 2 (shown in blue) is for intermediate G_min (mean G min opt = 10.8 dB), and the preference in Cluster 3 (shown in green) is for high G_min (mean G min opt = 18 dB).

The last question addressed is how many paired comparisons are needed to place a participant in one of the three preference clusters. As 180 paired comparisons (the number we used in this study) is too many for clinical practice, only G_min levels of 0, 7, 8, 9, 10, and 18 dB were used in a simulation study. These G_min levels coincide approximately with the G min opt from each of the three clusters. Response data were used from paired comparisons between G_min = 0 dB and each of G_min = 7, 8, 9, or 10 dB (eight pairs, as each could be the first or second of a pair), between G_min = 18 dB and each of G_min = 7, 8, 9, or 10 dB (eight pairs) and G_min = 0 dB and G_min = 18 dB (two pairs). To increase the weight of these last pairs, the data were used twice for a total of 8 + 8 + 2 × 2 = 20 pairs. The data contained only two pairs of comparisons between G_min = 0 dB and G_min = 18 dB and hence only two out of 18 pairs compare preference for Cluster 1 versus 3 (cluster numbering as in Figure 7). As this is an important comparison, we increased the weight of these comparisons by including them twice, increasing the fraction from 2/18 to 4/20. As there were two repetitions of all paired comparisons per participant, the total number of paired comparisons per participant was 40.

Subsequently, a resampling procedure was used on these paired comparisons, simulating a listening experiment where pairs were presented in random order. In resampling, each paired comparison is equally likely to be picked and each pair can be picked only once: uniform sampling without replacement. Note that this resampling is an average-case scenario: one could easily improve on this by adaptive sampling, where one uses previous responses to update the sampling strategy. For example, after a participant has indicated 6 times a preference for G_min = 0 dB over G_min = 18 dB one is unlikely to get new information by asking the participant a seventh or eighth time (as is done in the resampling procedure.

Getting back to the resampling procedure, 5–40 pairs were sampled (in Step 5) from the 40 paired comparisons selected previously and the Euclidian distance was calculated between the win counts of the sample and the mean win counts of each cluster as shown in Figure 7. Lastly, the cluster with the shortest Euclidean distance from the sample was picked. This simulation was run 500 times to estimate how often the sample ended up in the correct cluster (i.e., the cluster assigned to the participant based on all 180 paired comparisons).

The results of this simulation are shown in Figure 8. For nearly all participants in the first and third cluster, after 20 paired comparisons a probability of >80% of choosing the correct cluster was reached. In the middle cluster participants NH4, NH5, NH6, HI-mild2, HI-mild9, HI-mod3, HI-mod7, and HI-mod8 needed more paired comparisons to reach a probability of >80% or did not reach this percentage at all. Apparently, a significant fraction of participants in this cluster needs a wider range in G_min to adequately estimate the actual preference. This results in more comparisons than the G_min range in our simulation. We have no explanation for why the probability of choosing the correct cluster of HI-mild7, which belongs to the third cluster, was significantly lower in comparison to the other participants in this cluster.

OPEN IN VIEWER

Figure 8.

Probability of selecting the correct cluster (green, blue, or orange) for each participant based on a random subset of the paired-comparison data. Data shown represent resampling from paired comparisons between three distinct groups based on G_min (0 dB, 7, 8, 9, 10 dB, and 18 dB). The color coding represents the cluster the participants were assigned to by the cluster analysis (see Figure 5). The horizontal blue lines indicate a probability of 0.5.

Q1: Does Preference for NR Strength Differ Between NH and HI Listeners?

In line with our hypothesis, there was a significant difference in mean preference between NH participants and HI participants with moderate hearing loss. On average, the participants with moderate hearing loss preferred stronger NR than the NH participants did. Also, there was a moderate positive correlation between PTA_{1, 2, 4 kHz} and G min opt , and between SRT and G min opt . There was no significant difference in average NR strength preference between NH participants and HI participants with mild hearing loss and between HI participants with mild or moderate hearing loss.

The effect of age was explored by calculating the Spearman correlation between age and G min opt , and between age and PTA_{1, 2, 4 kHz}. Age and PTA_{1, 2, 4 kHz} were significantly correlated. This was expected because age was not a selection criterion for participant inclusion and hearing loss generally increases with age (Hoffman et al., 2017). G min opt was not significantly correlated with age, which suggests that the correlation between hearing loss and G min opt was not primarily driven by age. These results imply that a person with more severe hearing loss is likely to prefer stronger NR, irrespective of age.

Houben et al. (2011b) studied preferences for NR strength in a similar experiment. The average value of G min opt was only 0.3 dB larger for HI participants than for NH participants and this difference was not significant. They used three different background noises and two different NR algorithms. They tested 10 NH participants and 7 HI participants. The current study included more HI participants and divided them into groups that differed in hearing loss. Additionally, the current design included more NR strength levels which resulted in more pairwise comparisons for the single NR algorithm. These methodological differences may explain why the effect of hearing loss on NR strength preference was significant here whereas it was not in Houben et al.’s (2011b) study.

The direction of preference (those with more severe hearing loss are likely to prefer stronger NR) is in line with the hypothesis that HI listeners are less sensitive to signal distortions (Brons, Dreschler, et al., 2014). Neher and Wagener (2016) also found that HI listeners preferred stronger NR. The average G min opt was 8.2 dB for the NH group, 11.6 dB for HI-mild, and 15.7 dB for HI-moderate. For the same NR algorithm and noise type, the average G min opt was 4.3 dB in Houben et al. (2011b) for 17 participants (NH and HI participants combined). The cause of this difference is not clear. Perhaps the presence of different types of background noise in the previous study influenced the preferred NR strength.

We found a significant effect of hearing loss on preferred NR strength, although the effect size was only moderate (ρ = 0.46). This could be explained by the large individual differences in preferences (see Figure 3). Thus, there is substantial variation in preference irrespective of hearing status. Other individual traits might underlie this large spread of preferences for NR strength. Houben et al. (2011b) also found a large spread of preferences for NR strength for both NH participants and HI participants, with standard deviations of 3.6 dB and 4.0 dB, respectively. Several studies have investigated individual factors in relation to personal preferences for signal processing features in hearing aids (Neher & Wagener, 2016; Perry et al., 2019; Recker et al., 2020; Sugiyama et al., 2022). Unfortunately, for NR strength preference most of the investigated personal traits or factors were not predictive, including the acceptable noise level (ANL; Neher & Wagener, 2016; Recker et al., 2020), self-reported sound personality traits (Neher & Wagener, 2016), and the detection threshold for signal distortions (Brons, Dreschler et al., 2014; Neher & Wagener, 2016). On the other hand, Neher (2014) cautiously concluded that lower working memory might be related to preference for stronger NR. Individual preferences for NR settings, and thus NR strength, are complex and not easily predicted by subjective or objective measures.

A possible and promising explanation for the spread of preference is the individual trade-off between noise tolerance and distortion tolerance. Several researchers have used this trade-off theory to explain individual differences in preferences for NR settings (Brons, Houben, et al., 2014; Houben et al., 2013; Luts et al., 2010; Neher & Wagener, 2016; Reinten et al., 2019; Rohdenburg et al., 2005; Völker et al., 2018). Sugiyama et al. (2022) assessed individual differences in tolerance for signal distortion and residual noise using a single-channel speech enhancement algorithm. They found that their participants (N = 32) could be divided into two equal groups: those who are sensitive to distortions and those who are sensitive to noise. Kubiak et al. (2022) also found stable responses for participants (N = 30) who were classified either as “noise-haters” or “distortion-haters” in complex listening situations with different maskers. In the results shown in Figure 2, there are some participants of whom we can expect an individual trade-off to be made. For instance, the preference curve of NH1 peaks in the middle of our range of NR strengths suggesting this participant prefers an equal balance between speech distortion and noise attenuation. The preference curve of NH8 strongly suggests a preference for as much noise attenuation as possible in spite of the inevitable speech distortions, and vice versa for NH2. HI-mild9, however, which shares a similar G min opt as NH1, does not seem to show such a clear trade-off in preference as the preference curve is much flatter. Perhaps this participant is more indifferent in preferring a certain NR strength and is bothered less by noise as well as distortion effects. In future research, we aim to study this individual trade-off for noise and distortion tolerance in NR strength preferences.

Q2: How Many Distinct Settings Are Required to Classify Participants Into Similar Groups of NR Strength Preference?

The results of the cluster analysis showed that for our NR algorithm, the participants could be divided into three groups with similar preferences. The G min opt values for the three clusters were 0 dB (no NR), 10.8 dB, and 18 dB; see Figure 6. Figure 7 shows the mean win counts for each of the three clusters. This figure clearly illustrates the differences in individual preference which can be categorized as no NR, medium NR, and strong NR. The results imply that it seems possible to limit the number of choices of NR strengths for the clinician. Specifically, for a NR algorithm similar to the one used in this study and with a similar range of NR strengths, three levels of NR strength might suffice. We do not want to imply, however, that the three settings of NR strengths suffice in all other NR algorithms, which can differ in signal processing strategies as well as in the range of NR strengths. For instance, should we have included even a stronger threshold for G_min, we might have concluded that four settings were required to be able to accommodate all preferences for NR strength. The results also imply that we can assume that simply offering an NR-on or NR-off option in a hearing aid is too limited, as there are listeners who do not prefer the maximum, nor do they prefer the minimal amount of gain reduction. The actual amount of three clusters is only a starting point based on a limited data set.

In the results of the cluster analysis, we can see that there was a trend for those with more severe hearing loss to belong to a cluster with a higher preferred NR strength, which is in line with our results for Q1. Most NH participants fell in Cluster 1 (low NR strength) or 2 (intermediate NR strength). HI-mild and HI-moderate participants fell more in Clusters 2 or 3 (high strength) than in Cluster 1. This suggests that most HI participants should be fitted with a device with NR active. However, the categorization of a participant in a cluster was not clear-cut because all three clusters contained one or more participants for the HI-moderate group. So there are, at least some, moderately HI participants who prefer no NR. These results reinforce earlier findings that the amount of hearing loss cannot reliably predict the preference for NR strength for an individual (Arehart et al., 2015; Brons, Dreschler et al., 2014). Therefore, instead of selecting the NR strength based on the audiogram, one could attempt to measure the preference with a short preference measurement.

Q3: How Many Paired Comparisons Are Required to Find the Optimal Setting for an Individual?

If one could a priori categorize a participant into one of the three clusters in a reliable way, one could optimize the NR for that individual. Paired comparisons have been used for many decades to evaluate preferences for hearing aid features (Byrne, 1994; Neuman et al., 1995; Zerlin, 1962), and are often used for measurement of user preference for NR settings (Brons, Houben, et al., 2014; Marzinzik & Kollmeier, 2003; Smeds et al., 2010). In this study, we have used a complete set of two alternative forced-choice paired comparisons analyzed with the QUL model. There are other methods of paired comparison testing such as offering a “tie” option or offering more alternatives in one comparison, as are there other statistical methods and models (i.e., the Bradley–Terry–Luce model or the Elimination by Aspects model) to analyze paired comparison data (Cattelan, 2012; Tsukida & Gupta, 2011). In clinical practice, however, paired comparisons are not routinely used for hearing aid fitting and fine-tuning (Amlani & Schafer, 2009). In a survey of 251 audiologists, Anderson et al. (2018) found that the vast majority of the respondents used default settings of the manufacturer (58%) or their own expertise (38%) for fitting an NR feature in a hearing aid. An understandable reason for not using paired comparisons is that they take considerable time, which is not feasible in clinical practice. Therefore, we investigated whether the optimal level of NR strength can be found with a small subset of paired comparisons.

Figure 8 shows for each participant the probability of choosing the correct cluster against the number of paired comparisons. For five participants (e.g., NH4), the probability did not reach 50%. This implies that for these participants a higher probability, if possible, would require more comparisons, or a wider range of included NR strengths. The majority of these participants are in the middle cluster. However, for most of the participants in the first and third clusters, the probability of choosing the correct cluster was close to 100% after approximately 10–20 paired comparisons. Such a measurement should take about 5–10 min. For a clinical setting, we suggest using a limited set (e.g., 15 comparisons, or only the comparisons of the G m i n o p t values of the three clusters). If the listener cannot complete the task or if the result is inconclusive (i.e., no G m i n o p t can be found using the QUL model), the middle setting should be selected. A shortened preference measurement should be applicable for many hearing aid users in clinical practice, or alternatively could be incorporated in a hearing aid using machine learning, as suggested by Søgaard Jensen et al. (2019).

Although the NR algorithm used in this study is comparable to NR systems used in modern hearing aids, the optimal NR strength levels from this experiment apply only to the limited conditions tested in this study. Moreover, this study did not account for the effects of other non-linear processing, such as amplitude compression. We do not know if preferences for NR strengths differ at other input SNRs. However, since the chosen input SNR of +5 dB is representative of real-life scenario this limitation does not influence the conclusions of this work.

It is important to discuss other possible effects of our signal processing that might contribute to the differences in preference for NR strength. It is currently not known if or how the chosen hearing aid fitting rule (i.e., the amplification strategy to compensate for hearing loss) influences the preference for NR. To achieve results that are relevant to H users, one needs to use fitting rules. Fitting rules are used in hearing aids to prescribe individual, frequency-dependent amplification to compensate for the personal hearing loss. Research on preference with HI participants is thus complicated by the individual frequency-dependent hearing loss. The combination of the hearing loss with the chosen frequency-dependent fitting rule will determine the amount of spectral coloring relative to participants without hearing loss. Note that even without the use of frequency-dependent amplification there is some spectral coloring relative to NH (e.g., high frequencies can be attenuated by age-related hearing loss). We chose to apply the linear NAL-RP filter because it resembles the commonly used NAL fitting rule and it avoids the known complicated interactions with non-linear amplification (Brons et al., 2015). To what extent our choice of fitting rule (NAL-RP) has influenced the preference results is unknown. More specifically it is known that listeners might prefer less gain than prescribed by the NAL-RP fitting rule (e.g., Humes et al., 2000, 2001). This “too harsh” sound might have influenced the preference. However, due to the RMS correction that effect seems limited because overall the NR acts on all frequencies (the noise was spectrally matched to the speech).