Abstract
The extent to which active noise cancelation (ANC), when combined with hearing assistance, can improve speech intelligibility in noise is not well understood. One possible source of benefit is ANC's ability to reduce the sound level of the direct (i.e., vent-transmitted) path. This reduction lowers the “floor” imposed by the direct path, thereby allowing any increases to the signal-to-noise ratio (SNR) created in the amplified path to be “realized” at the eardrum. Here we used a modeling approach to estimate this benefit. We compared pairs of simulated hearing aids that differ only in terms of their ability to provide ANC and computed intelligibility metrics on their outputs. The difference in metric scores between simulated devices is termed the “ANC Benefit.” These simulations show that ANC Benefit increases as (1) the environmental sound level increases, (2) the ability of the hearing aid to improve SNR increases, (3) the strength of the ANC increases, and (4) the hearing loss severity decreases. The predicted size of the ANC Benefit can be substantial. For a moderate hearing loss, the model predicts improvement in intelligibility metrics of >30% when environments are moderately loud (>70 dB SPL) and devices are moderately capable of increasing SNR (by >4 dB). It appears that ANC can be a critical ingredient in hearing devices that attempt to improve SNR in loud environments. ANC will become more and more important as advanced SNR-improving algorithms (e.g., artificial intelligence speech enhancement) are included in hearing devices.
Introduction
Wearable audio devices with active noise cancelation (ANC)—the ability to reduce the level of environmental sounds via destructive interference—have reached widespread adoption in the consumer electronics marketplace. Consumers with normal hearing primarily use this technology to reduce the intensity of the sound reaching their ears from the environment, thereby enabling them to listen to their preferred content at a comfortable listening level. Additionally, most modern ANC systems incorporate a canal-facing microphone which is used to cancel both external and body-conducted sounds (e.g., Mejia et al., 2008; Sabin, 2020). This ability to cancel the occlusion effect (e.g., Mueller et al., 1996) is particularly valuable in hearing aids because it allows the user to have the benefit of a sealed earbud (e.g., bass output) without the typical drawback (e.g., “boomy” own voice). While these benefits of ANC are widely appreciated, less is known about how ANC can improve the primary use case of hearing aids—speech understanding. Modeling such improvements are the focus of this article.
For the sake of clarity, we note that there are various algorithms in the hearing aid literature with similar names to ANC but very different functions (e.g., “adaptive noise reduction” attempts to reduce the amplification applied to noise; Chung, 2004). Throughout this report, we use ANC to refer only to cancelation of sound via destructive interference. None of these similarly named hearing aid algorithms are ever considered in this article. For further clarity, we also note that despite having “noise” in its name, ANC reduces the level of the entire incoming signal (i.e., both speech and noise).
In recent years, ANC has been integrated into several hearing assistance devices (including products from Bose, Apple, and Nuheara). Despite this integration, there has been little research into how ANC interacts with the traditional hearing aid signal processing to improve speech intelligibility in noise. This is a particularly exciting possibility because the ability to improve hearing in noise is regularly rated as the most important feature to hearing aid consumers (Jorgensen & Novak, 2020; Kochkin, 2012; Manchaiah et al., 2021; Palmer et al., 2009; Picou, 2020).
To understand the influence of ANC on speech understanding, it is first important to recognize that the sound reaching the user's eardrum is the superposition of two signals: (1) the amplified path transmitted from the hearing aid transducer and (2) the direct path, primarily transmitted through vents in the ear tip (Figure 1A). For sloping hearing losses, the amplified path boosts high frequencies (e.g., Figure 1B, solid line) and typically has a latency on the order of a few milliseconds that reflects the digital signal processing in the hearing aid (Alexander, 2016; Launer et al., 2016). The direct path is primarily the propagation of sound through the vent; it has a low pass shape (e.g., Figure 1B, dashed line) and a latency on the order of a few microseconds. When the two paths are similar in level but different in latency, they will interact to create unpleasant artifacts such as comb filtering (Figure 1C). The hearing aid system designer might even choose to apply a high pass filter to the amplified path to avoid these artifacts (e.g., Bramsløw, 2010). In contrast, when ANC is integrated, the hearing aid will substantially reduce the level of sound transmitted through the direct path (Figure 1D) by producing an antiphase signal in a fast path with microsecond-scale latency. This reduction to the direct path from ANC increases the level difference between paths (Figure 1E), thereby considerably reducing comb filtering (Figure 1F). In addition, this increased difference between paths can lead to improved signal-to-noise ratio (SNR) at the user's eardrum. In a standard (non-ANC) hearing aid, the level of the direct path can limit the extent to which the noise in the amplified path can be reduced. This is especially true in mid-to-low frequency bands where the direct and amplified paths often have similar levels. When ANC lowers the level of the direct path, it effectively lowers the “floor” imposed by the direct path, thereby allowing the noise in the aided path to be reduced further. We use the term “SNR Realization” to describe this phenomenon because increases to SNR in the amplified path (decreases in noise level) that would not be realized at the eardrum without ANC can become “realized” there when ANC is integrated.
ANC in hearing aid use. (A) Conceptual rendering of the two paths reaching the users’ eardrum with a standard hearing aid. The direct path is transmitted through a vent. The amplified path (a “slow path” with millisecond-scale of latency) is transmitted from the device. (B) The insertion gain is associated with each path. The amplified path displays the NAL-NL2 (Keidser et al., 2011) target gains for a standard moderate hearing loss (N3; Bisgaard et al., 2010) in an 80 dB SPL environment. The direct path displays the average occluded insertion loss observed for an instant-fit double-dome device (Cubick et al., 2022). (C) The insertion gain of the combined paths. Note the high degree of comb filtering below 1 kHz. (D) Conceptual rendering of the paths in an ANC hearing aid. Note that the ANC signal (a “fast path” with microsecond-scale latency) is produced by the device but alters only the direct path. (E) As in B, but for the ANC hearing aid. Note the reduction to the direct path. (F) As in C, but for an ANC hearing aid. Note the reduction to comb filtering. All graphs are shown as insertion gain—the difference in level at the eardrum between the open ear and aided ear.
While aspects of this phenomenon have been appreciated for some time (e.g., Dillon, 2012; Winkler et al., 2016) we are not aware of any attempts to model when those benefits will occur in typical hearing aid usage or how big they will be, but several trends could be hypothesized. For example, it would be hypothesized that the ability for ANC to increase the “realized” SNR would increase with either less amplification in the hearing aid path or with increased capability of the hearing aid signal processing to improve SNR (because both reduce the level of the noise in the amplified path). Furthermore, nearly all modern hearing aids use wide dynamic range compression (WDRC) to reduce the hearing aid's gain as the environmental level increases. Therefore, it would be hypothesized that any benefits of ANC would be larger as the sound level of the environment increases. In this article, we attempt to model these trends. We do so by simulating the combination of amplified and direct paths and computing speech intelligibility metrics on those combinations.
Method
We examined the potential of ANC to improve speech intelligibility metrics in noise by comparing outputs from pairs of simulated hearing aids that differ only in their ability to provide ANC. For each hearing aid output, we compute two intelligibility metrics: eSTOI (Jensen & Taal, 2016) and HASPIv2 (Kates & Arehart, 2021). The difference in those metrics between pairs of simulated devices (ANC vs. non-ANC) is termed the “ANC Benefit.”
Stimuli
Standard speech and noise stimuli were presented across a wide range of input sound levels and at typical environmental SNRs for those sound levels. For the noise signal we used a stationary noise matched to the long-term average spectrum of speech (Byrne et al., 1994). That noise is scaled to simulate environmental sound levels from 50 to 90 dB SPL in 10 dB steps. For the speech, we use the Internation Speech Test Signal (ISTS) signal (Holube et al., 2010) which is a 60 s compilation of female speech across multiple languages. For each noise level, the speech was scaled to achieve the average SNR for that level, as observed by Wu et al. (2018) in individuals with mild-to-moderate hearing loss. The result is that the SNR of the input signal decreases with increasing noise level.
Hearing Loss
As described in the Introduction, the gain applied by the hearing aid (at least in the frequency bands <1 kHz) is expected to influence the amount of ANC Benefit. With this in mind, we tested across a range of standard hearing losses because increased hearing loss will call for higher target gains in the prescription. Specifically, we examined the standard sloping hearing losses from Bisgaard et al. (2010) ranging from N1 (very mild) to N5 (severe) (Figure 2, left). For each audiogram, we computed the prescribed target insertion gains at 50, 65, and 80 dB SPL input levels using the NAL-NL2 (dll v2.15) (Keidser et al., 2011) fitting formula using the corrections for binaural fitting, new users, unspecified gender, and non-tonal language. Then, using our hearing aid simulator (described below) we adjusted the gain and compression parameters to match those targets. The NAL-NL2 targets (dashed lines) and our matches (solid lines) are shown for each hearing loss on the right side of Figure 2. We were able to achieve high-quality fits for all hearing losses, with all fits within <4 dB of target at all frequencies and input levels.
Hearing losses and prescribed gains. Audiograms (left) are shown for five standard sloping hearing losses (Bisgaard et al., 2010) increasing in severity from very mild (N1) to severe (N5). The corresponding plots on the right show target NAL-NL2 (Keidser et al., 2011) gains (dashed lines) for 50, 65, and 80 dB SPL (upper to lower lines) speech-shaped inputs as well as the fits to target achieved in our hearing aid simulator (solid lines). High-quality fits (<4 dB error) were achieved for all hearing losses.
Audio Simulation
We simulated pairs of hearing aids that differ only in their ability to provide ANC. The schematic describing our audio simulations is shown in Figure 3. All processing for this simulation was done at a sampling rate of 16 kHz.
Hearing aid simulation. Schematic shows the major processing blocks used to simulate pairs of standard and ANC hearing aids. Both hearing aids use the same amplified path (right side) that is comprised of an SNR increase followed by WDRC. The direct path (left) side applies the REOIG associated with the simulated ear tip. For the ANC (but not standard) hearing aids an additional filter is applied (see Figure 4, solid lines). The final outputs are the combinations of the amplified path with the appropriate direct path.
First, the speech and noise are scaled to the target SPL and SNR. They are then split into two parallel paths designed to simulate the direct and amplified paths. The direct path (Figure 3, left side) is simulated by first passing the combined input signal through a filter designed to match the real ear occluded insertion gain (REOIG) caused by the passive attenuation from the ear tip. We fit a zero-delay FIR filter to match the across-subject average REOIG for an instant-fit double dome (Cubick et al., 2022) (Figure 4, dashed line). This REOIG is similar to that obtainable with a 3-mm vent. This venting is similar to the most occluding instant-fit ear tips, and the least occluding custom earmolds. Next, if the device being simulated is ANC-capable, the REOIG filter output is passed through an additional static IIR filter in the time domain. The response of the filter was designed to simulate the attenuation (expressed in Insertion Gain) provided by a simplified active cancelation system with both feedforward and feedback ANC. We repeat the test under four hypothetical ANC responses (Figure 4, solid lines) with peak attenuations (i.e., “strength”) ranging from 10 to 40 dB in 10 dB steps. We refer to these filters by their peak strength simply for convenience. We do not intend to imply that the band where the ANC peak occurs (100–500 Hz) is the most relevant one for speech intelligibility. We also include a fifth condition with no direct path whatsoever (Inf dB) to test the theoretical limit of the influence of ANC. Because ANC systems must operate with extremely low latency (microsecond scale) to be effective, the output of each ANC filter is adjusted to add no delay.
Direct path filters. The REOIG is shown for the standard hearing aid (dashed line) with a double-dome instant-fit ear tip (Cubick et al., 2022) as well as for systems that use the same ear tip together with increasing levels of hypothesized peak ANC strength (solid lines). Inset numbers indicate the maximum level of cancelation simulated by each filter.
The amplified path (Figure 3, right side) is simulated by first applying an SNR increase. This stage is designed to represent the “signal cleaning” portions of typical hearing aid signal processing paths. This cleaning could come from directional processing, SNR-improving algorithms, or even remote microphones. This stage is intended to be method-agnostic and is therefore implemented as a simple broadband reduction to the noise. The amount of reduction to the noise is termed the “HA SNR Increase.” Across testing, we vary this value from 0 to 10 dB SNR in steps of 2 dB. After the HA SNR Increase is applied, the signal is passed through an offline hearing aid simulator. This simulator is a custom set of standard algorithms designed to replicate the most common aspects of hearing aid signal processing. The only enabled algorithms were those that support WDRC. No adaptive algorithms were used. WDRC processing was done via a 12-channel multiband compressor with a single “knee point” and compression ratio per channel. The output of the simulated hearing aid was then adjusted to a target latency (a delay relative to the direct path). This latency is important to model because the degree of comb filtering between direct and aided paths will increase as the latency increases. We set this latency to 5 ms, within the 1–10 ms range of typical hearing aids (Alexander, 2016; Launer et al., 2016). For this simulation, the ANC and amplified paths see the same input, which would be the case if they received input from the same microphone(s). This is often the case in real products but is not required.
The final step was combining the paths to create pairs of output signals. For the standard hearing aid signal, the amplified path was combined with the direct path that did not include the ANC filter. For the ANC hearing aid signal, the direct path included the ANC filter.
Metrics
All simulated hearing aids were analyzed using two metrics that have been shown to be related to speech intelligibility: the Hearing Aid Speech Perception Index version 2 (HASPIv2; Kates & Arehart, 2021) and Extended Short-Time Objective Intelligibility (eSTOI; Jensen & Taal, 2016). The difference in metric scores between the standard versus ANC hearing aids is termed the “ANC Benefit.” HASPIv2 is an intrusive metric (i.e., it requires a clear reference) that compares the clean speech and hearing aid signals using a model of the impaired auditory system. The inputs are the audiogram, the clean signal, and the aided signal. The final stage of this model uses an ensemble of neural networks trained on human perceptual data. The overall output of these networks is a value between 0 and 1 that corresponds to the estimated proportion of sentences accurately recognized (using the experiments as the training data). This metric allows us to estimate the size of the ANC Benefit in units that are more similar to “intelligibility,” and is sensitive to improvements from both audibility and SNR. As with human intelligibility data, HASPIv2 values saturate at high SNRs (Kates & Arehart, 2022) when the value approaches the asymptote value of 1.0. In this scenario, large changes to SNR will only yield small metric changes, potentially obscuring very noticeable effects. With this limitation in mind, we also include eSTOI as a second intelligibility metric. eSTOI compares the clean speech and processed signals using a spectrogram-like representation with features selected to match performance of listeners with normal hearing. The output is a scalar value between 0 and 1 that has a monotonically increasing relation with the speech intelligibility but does not directly model percent correct speech intelligibility. We’ve observed that, unlike HASPIv2, eSTOI has a near-linear (non-saturating) relationship to SNR for the tested conditions. Thus, including this metric allows for an additional lens to examine the influence of ANC.
Results
Effect of Environment Level and HA SNR Increase
It is first important to describe the plots that we use throughout this section to visualize the degree of ANC Benefit (e.g., Figure 5). The ANC Benefit is expected to increase as the gain applied to the noise decreases (i.e., as the noise levels in the direct and amplified paths become more similar). For the conditions considered here, that gain will decrease if either (1) the environment level increases (causing WDRC to reduce overall gain) or (2) the HA SNR Increase gets larger (allowing the device to further attenuate the noise). Both factors would be expected to contribute largely independently of each other. With this in mind, it is helpful to visualize the intelligibility metrics simultaneously as a function of both environmental level and HA SNR Increase. The trends influencing the ANC Benefit are shown in Figure 5 for the 30 dB ANC level and the N3 audiogram. The top row (panels A–D) shows the trends for HASPIv2 metric. Figure 5A shows the HASPIv2 scores for the unaided condition. The contours (indicating HASPI value) in this plot are vertical here because there is no hearing aid to apply an SNR increase (the vertical dimension). The metric values decrease toward the right because the environmental SNR decreases, reducing intelligibility. Figure 5B shows the values of the same metric but for a standard hearing aid gain fit to prescriptive targets and combined with a direct path of a passive double-dome ear tip (Figure 4, dashed line). Notice that the overall HASPI values have increased (at least for quiet inputs) relative to the unaided condition. This indicates that the standard aid does provide some speech intelligibility benefit. There is no improvement and some worsening at higher environmental levels when HA SNR Increase is low. Figure 5C shows the same metric, but for an ANC hearing aid designed to be identical to the standard one except for a direct path that uses an ANC filter with 30 dB max attenuation (see below for an examination of ANC level). The HASPIv2 scores here have increased relative to panel B indicating the ANC aid does provide more benefit than the standard device, especially in loud environments. The difference between Figures 5B and 5C (i.e., the “ANC Benefit”) is shown in panel D. Figure 5D shows how the size of the ANC Benefit changes as a function of environment level and HA SNR Increase. The size of the ANC Benefit increases toward the top right corner of the panel. This trend indicates that the size of the ANC Benefit increases with both the environmental level and the hearing aid's ability to provide an SNR increase (HA SNR Increase). The size of the ANC benefit is substantial. Each contour is a change of 0.1 in the HASPv2 metric. For example, the HASPIv2 score is >30% (0.3 contour) higher than a standard aid when a user is in environments more intense than about 70 dB SPL and wearing devices that have performance that is similar to a HA SNR Increase value of at least 4 dB. There is some positive ANC Bfor virtually all measured conditions.
Estimation of ANC Benefit. Contour plots show the values of speech intelligibility metrics (top row: HASPIv2; bottom row: eSTOI) as a function of the environmental level (horizontal axes) and the capability of the hearing aid to increase SNR (HA SNR Increase; vertical axis). The second row of the x-axis labels displays the SNR associated with each environmental level (Wu et al., 2018). The left panels (A and E) show the metric for an unaided condition. Those values are then shown for a standard hearing aid (panels B and F) and an ANC hearing aid (panels C and G). The difference between standard and ANC hearing aids (the ANC Benefit) is shown in the rightmost panels (D and H).
The eSTOI values for the same sequence of conditions are shown in the bottom row of Figure 5, panels E–H. The overall trends are similar to the HASPIv2 values. However, since eSTOI was not designed to account for hearing loss, it does not show a benefit of a standard hearing aid over the unaided condition. The pattern of ANC Benefit for this metric (Figure 5H) is similar to that of HASPIv2 (Figure 5D), both showing increasing benefit toward the top right. The shape of the ANC Benefit plot for eSTOI is slightly different from the one for HASPIv2 likely due to differently shaped relationships between each metric and SNR. While HASPIv2 is sigmoidal with respect to SNR, eSTOI is largely linear with respect to SNR (at least over the tested conditions). This linear relationship (0.0225 eSTOI/dB SNR for the conditions here) can be used to estimate the SNR equivalent of the ANC Benefit. Each contour is intended to approximate an ANC Benefit equivalent of an additional 1 dB SNR. Again, this metric shows a sizable ANC Benefit. For example, the eSTOI increase of an ANC versus standard hearing aid is 0.0675 (3 dB SNR equivalent) when a user is in environments that are more intense than about 70 dB SPL and wearing devices with ability to increase SNR by about 4 dB.
Effect of ANC Strength
We next examined how the size of the ANC Benefit varies with the strength of ANC. We repeated the same analysis as above but varied the ANC filter to five different strengths (Figure 4, solid lines). The ANC Benefit across these conditions is shown in Figure 6 where the top row is the HASPIv2 metric and the bottom row is eSTOI. The peak strength of ANC, noted in each panel's title, increases from left to right across the panels. Additionally, in the rightmost panels (Figure 6F and 6L), the average ANC Benefit (across all SPL and HA SNR Increase values) for each ANC strength is plotted. At the lowest peak ANC strength (10 dB peak; Figure 6A and 6G) there is only a small benefit of ANC (maximum ANC Benefit about 0.2). However, the size of the ANC Benefit increases substantially at the 20 dB peak level (Figure 6B and 6H; maximum ANC Benefit is about 0.5). The increase to the average sizes of those benefits with increasing ANC strength can be clearly seen in Figure 6F and 6L. At peak ANC strengths higher than 20 dB, the ANC Benefit continues to rise, but at a slower rate. There is virtually no difference between the 40 dB peak ANC strength and infinite ANC conditions indicating there would be no additional ANC Benefit beyond peak 40 dB of cancelation.
Influence of ANC strength. The ANC Benefit (as in Figure 5, right) is shown for five different “strengths” of ANC ranging from 10 dB (A and G) to Infinite (E and K). The top row shows values of HASPIv2 and the bottom row shows eSTOI. ANC strength is labeled in each panel's title. Also shown is the average ANC Benefit (across environmental level and HA SNR Increase) for each strength (panels F and L, for HASPIv2 and eSTOI, respectively).
Effect of Hearing Loss
Our next analysis focused on how the size of the ANC Benefit changes across hearing loss severity. We repeated the same analysis as above but for five different standard hearing losses increasing in severity from N1 = very mild to N5 = severe (Figure 2). The peak ANC strength was fixed at 30 dB across these analyses. The pattern of ANC Benefit is shown in Figure 7 for HASPI (top row) and eSTOI (bottom row) where the examined hearing loss is in each panel's title. As shown in Figure 6, we report the average ANC Benefit per hearing loss in the rightmost panels. The HASPIv2 and eSTOI scores have different patterns of ANC Benefit with respect to hearing loss level. The HASPIv2 scores show a small increase in ANC Benefit from N1 to N3 and then a strong decline for N5. However, the eSTOI scores show a monotonic decrease in ANC Benefit as hearing loss increase from N1 to N5.
Influence of hearing loss. As shown in Figure 6, but for different degrees of hearing loss (indicated in each panel's title).
This difference likely arises from fundamental differences in how the metrics are computed. eSTOI does not consider the impact of hearing loss, but has a near-linear relationship to SNR (see Method, Metrics). ANC Benefit with eSTOI decreases with increasing hearing loss because hearing aid gain increases with increasing hearing loss, and ANC becomes gradually less impactful with increasing hearing aid gain. The near-linear relationship can be used to estimate the effects of processing on SNR which also may be a helpful lens to understand the impact of the ANC Benefit. Consider the 0.0675 eSTOI (approximately 3 dB SNR) ANC Benefit. For the N1 (very mild) loss this level of ANC Benefit (or greater) would be predicted for nearly all environmental levels and for most hearing aid capabilities. For the N5 (severe) loss, this level of benefit would only occur in the loudest environments (>80 dB SPL) and the most capable devices (>6 dB HA SNR Increase). In contrast, HASPIv2 shows a different pattern across hearing loss because (a) it reflects the combined improvements from audibility and SNR and (b) it has a non-linear, saturating, relationship to SNR (see Discussion, ANC Benefits Increase for Less Severe Hearing Losses).
Discussion
ANC is being incorporated into an increasing number of hearing devices, yet there has been little research into its influence on aided speech intelligibility in noise. Here we presented a model showing how ANC and hearing aid signal processing can interact to influence speech intelligibility metrics. This model predicts that ANC can provide a substantial benefit to intelligibility metrics especially in the most challenging conditions and on the devices that can provide the largest increase to SNR. This pattern of benefit is consistent with the mechanism described in the introduction (“SNR Realization”) where ANC lowers the “floor” on the minimum level of the noise at the eardrum. This lowered floor allows any increases to SNR that were created by the hearing aid to be “realized” at the eardrum. The size of this ANC Benefit is highly variable, depending on the properties of the environment, listener, and device. The model presented here makes several predictions about that variability that can be tested in the real world.
ANC Benefit Increases in More Challenging Environments
The predicted benefit of ANC increased as the environmental sound level increased. This trend can be seen in Figure 5D and 5H where the size of the ANC Benefit increases toward the right side of each contour plot. This trend is largely caused by WDRC. For sensorineural hearing losses (including the ones tested here), WDRC is used to match prescribed gain targets that decrease as the environmental SPL increases (e.g., Figure 2). As that gain decreases, the noise in the direct path will increasingly contribute to the resulting speech intelligibility—effectively placing a “floor” on the minimum level of the noise at the eardrum especially in moderate-to-loud environments and in mid-to-low frequency bands. ANC can move that floor much lower which allows noise to be reduced further, thereby improving intelligibility, and creating the ANC Benefit.
This pattern of benefit is particularly valuable because users regularly cite improving speech intelligibility in loud environments as one of the most important attributes of a hearing aid as well as a major source of current dissatisfaction (Jorgensen & Novak, 2020; Kochkin, 2012; Manchaiah et al., 2021; Palmer et al., 2009; Picou, 2020). The limits imposed by the noise in the direct path of standard hearing aids could be a major contributor to that dissatisfaction. If so, ANC-enabled hearing aids have the potential to increase satisfaction in this critical use case. The increase to SNR at the user's ear from ANC can be substantial. The added SNR from ANC can be estimated from the eSTOI benefit plots (Figure 5H) where each contour approximates a 1 dB SNR change. The estimated just noticeable difference in SNR is 3 dB (McShefferty et al., 2015) which corresponds to the 0.0675 eSTOI contour. This implies that the additional increase to SNR due to ANC would be noticeable for a wide range of conditions (>0.0675) and will be meaningful (>0.135; 6 dB SNR; McShefferty et al., 2016) in some of the more extreme conditions. Furthermore, it is possible that the metrics used here underestimate the value of improving SNR in lower frequency bands. There is some evidence (e.g., Calandruccio & Doherty, 2008; Hogan & Turner, 1998) that the traditional frequency importance functions may change for individuals with hearing loss. This could be a fruitful topic to consider in future work.
ANC Benefit Increases in More Capable Devices
Similarly, the predicted benefit of ANC grew as the ability of a device to increase SNR (HA SNR Increase) improved. In the contour plots of ANC Benefit this effect can be seen as an increase to ANC Benefit toward the top of the plots. The likely explanation for this trend is the same as above. Specifically, if the hearing aid can increase the SNR (via beamforming or advanced signal processing; e.g., Chung, 2004), it will typically do so by reducing the level of the noise while preserving the level of the speech at gain targets. In this case, the “floor” provided by the noise in the direct path will limit the SNR that can realized at the eardrum. This phenomenon can be seen in the intelligibility metrics for the standard hearing aid in the loudest environments (Figure 5B, right side). As the hearing aid SNR increase gets larger (toward top of figure), there is little change in intelligibility metrics for the standard hearing aid—presumably due to the noise level quickly reaching the “floor” imposed by the direct path. Even for the most capable devices (HA SNR Increase = 10 dB; Figure 5B, top) the standard hearing aid is predicted to provide almost no benefit above the unaided condition in environments >70 dB SPL (compare Figure 5A and 5B, top-right corners). In contrast, the ANC hearing aid shows a steep monotonic increase in intelligibility metrics as the hearing aid becomes more capable—even in the loudest environments (Figure 5C, right side). This difference shows that ANC-capable hearing aids enable the realization of the full benefit of SNR-improving technology while standard hearing aids do not.
While this pattern of benefit is already relevant for current technology (e.g., Patel et al., 2020; Serizel et al., 2010, 2013), it will become increasingly important as more powerful algorithms such as artificial intelligence speech enhancement (Fedorov et al., 2020; Healy et al., 2013; Wang, 2017) become included in more hearing devices. Such algorithms can generate unprecedented improvements to SNR, but those improvements will not be realized at the ear when a standard hearing aid is used because the noise floor will be limited by the direct path. In contrast, the full benefit of such algorithms would be realized when ANC is used, if the strength of ANC is sufficient.
ANC Benefits Increase When the Device Can Provide More ANC
There was a monotonic increase to the size of the ANC Benefit as the peak strength of ANC increased from 10 to 40 dB (Figure 6). This trend reflects the increased ability to reduce the sound transmitted through the direct path and remove its contaminating influence on the increases to SNR created by the hearing aid. The size of the increase to ANC Benefit was largest between 10 and 20 dB and got smaller at higher strengths. This result implies that most of the ANC Benefit can be achieved in systems that can provide about 20 dB peak of cancelation, but that the full benefit requires about a 40 dB peak. At low cancelation levels, the noise in the direct path (on which the ANC operates) may still be more intense than, or similar to, that in the amplified path. As the level of cancelation increases, the noise in the direct path might be lower than amplified but still able to create distortions (e.g., comb filtering) in the signal at the ear. In order for these distortions to not reduce the intelligibility metrics, the level of the noise in the direct path must be substantially lower than that in the amplified path (by at least 10 dB).
ANC Benefits Increase for Less Severe Hearing Losses
Some ANC Benefit was present across all tested hearing losses but it tended to decrease as hearing loss severity increased. This trend was clearest in the eSTOI data (Figure 7L) where there was a monotonic decrease in ANC Benefit with increasing hearing loss. This pattern is expected because the prescribed gain in the amplified path will increase with increasing hearing loss severity. As the level of the amplified path increases, the contamination from the direct path decreases, and therefore the improvements from decreasing the level of the direct path (i.e., the ANC Benefit) also decrease. The pattern of ANC benefit for HASPIv2 (Figure 7F) was more complex and did not show the largest ANC Benefit for most mild losses. As mentioned in the Method (see Metrics subsection), HASPIv2 reflects the combined improvements from both audibility and SNR increases and that metric has a non-linear, saturating, relationship to SNR. For this metric, the N3 loss showed the biggest benefit. The benefit decreased for more mild losses (because the metric was approaching saturation). The metric also decreased for more severe loss (because the increased hearing aid gain reduces the influence of ANC). For the very mild (N1) loss, the ANC Benefit might not be experienced as an increase in intelligibility (which is approaching maximum) but rather a reduction in listening effort (Pichora-Fuller et al., 2016). Nevertheless, there is a clear prediction that ANC can provide a substantial benefit in challenging environments for all hearing losses tested here.
Other Mitigations of Interference From Direct Path
While this work shows the harm that the direct path can cause in non-ANC hearing aids, it is important to recognize that some of this harm can be mitigated by filtering the amplified path. Specifically, when the prescribed gain is primarily limited to high frequencies, and the vent is large, a high pass filter can be applied to the amplified path without significantly deviating from gain targets (this cannot be achieved when the prescription calls for gain in the mid-to-low frequency bands). This filtering reduces comb filtering in the low frequencies by letting the direct path dominate those bands. Supporting the value of this approach, Bramsløw (2010) observed that, for open fits, sound quality preference scores increased as the cutoff frequency of a high pass filter on the amplified path increased. Indeed at least one manufacturer (Kühnel et al., 2016) applies a dynamic filter on the amplified path to reduce low-frequency comb filtering through “direct sound compensation.”
With this in mind, we were interested to understand how filtering the aided path would change the ANC Benefit. We ran a new simulation where a high pass filter (second order, 1 kHz cutoff) was added to the amplified path of the standard hearing aid (but not the ANC hearing aid). We focused on the baseline hearing loss and ANC strength (N3 and 30 dB, respectively) and the results are shown in Figure 8. Adding this high pass filter did improve the HASPIv2 scores for the standard hearing aid on some conditions while not changing those for the ANC hearing aid, thereby reducing the ANC Benefit. This improvement can be seen in Figure 8 where the ANC Benefit is lower for the condition with the high pass filter in the standard hearing aid (Figure 8B) than the condition without that filter (Figure 8A). The difference scores in Figure 8C (panel A − panel B) show the benefit of adding the high pass filter to the standard hearing aid. The pattern of benefit shows a moderate improvement (see the 0.2 contour) for conditions with both low HA SNR Increase values (<2 dB SNR) and moderate environmental levels (60–75 dB SPL). As device capabilities improve (i.e., HA SNR Increase grows) or the environmental level increases, the benefit of the high pass filter in the standard hearing aid diminishes. This benefit eventually reaches zero for the most challenging environments and most capable devices. Overall, this post-hoc analysis shows that the ANC Benefit can decrease for some conditions if the comparison standard hearing aid attempts to reduce combing by filtering the aided path. Additionally, while comb filtering is known to harm subjective sound quality (Bramsløw, 2010), this work shows that the HASPI metric also predicts a reduction to objective intelligibility. To the best of our knowledge, this effect has not been directly tested in human perception.
Influence of a high pass filter in the standard hearing aid. HASPIv2 ANC Benefit contours for conditions without (A) and with (B) a 1 kHz high pass filter added to the amplified path of the standard hearing aid. The difference between conditions (panel A − panel B) is shown in panel C.
Limitations
This report is intended to provide an initial model-based approach to examine the interaction between ANC and hearing aid signal processing on speech intelligibility in noise. It does not consider several factors that would also be expected to influence the size of the ANC Benefit. For example, the characteristics of the speech and noise signals themselves will influence the size of the ANC Benefit. Here we used stationary speech-shaped noise as the background noise, largely matching the spectrum of the target speech. The size of the ANC Benefit would likely change for different background noises. For example, in bass-heavy noises (e.g., aircraft cabins or loud background music) the size of the ANC Benefit could grow substantially by reducing the upward spread of masking (Martin & Pickett, 1970) from low-frequency maskers to high-frequency signals. Finally, this model does not consider the spatial differences between the speech and noise signals, and how those differences would be influenced by microphone location and array design. We used the same SNRs at the input to the amplified and direct paths. In more realistic environments, these SNRs would likely be different based on the locations of the speech, noise, and microphones. Areas for future examination could include topics such as other hearing loss shapes, more negative SNRs, and other vent sizes.
Overall, we view the interaction between ANC and hearing aid signal processing as a rich (and largely untapped) area of research in hearing aid science. We hope that this report will encourage new research and better products in this space.
