Effects of Source-Specific Dynamic Range Compression on Sound Quality for Individuals With Hearing Loss

Introduction

Imagine sitting at a busy restaurant where a friend is telling a story across the table, nearby dishes are clattering, people are talking at adjacent tables, and loud music is played in the background. These situations are challenging for hearing aid users, at least in part, owing to the manner in which nonlinear hearing aid processing, such as dynamic range compression (DRC) is implemented in most commercial hearing aids. In most devices, the same compression parameters are applied to all sounds in a mixture of signals coming through microphones, typically after some form of noise reduction or directional processing (Dillon, 2012; Ricketts et al., 2019). Consequently, the compression parameters may not be optimized for individual sound sources. Moreover, compressing signals jointly can introduce undesirable nonlinear distortions and reduce speech intelligibility and sound quality (Corey & Singer, 2021; Stone & Moore, 2004, 2008). This study considers a different approach to dynamic range control, where each individual sound signal in the mixture is compressed independently (e.g., Corey & Singer, 2017; Kowalewski et al., 2020; May et al., 2018; Overby et al., 2023; Zhang et al., 2025), possibly using different parameters, much as different vocal and instrument tracks are processed separately before mixing in music production. This source-specific (independent) compression strategy could help overcome the disadvantages of the joint compression of signals (Corey & Singer, 2021; Overby et al., 2023), particularly if combined with future advances in real-time source separation and wireless connectivity.

DRC is a fundamental feature of modern hearing aids that provides level-dependent gain to accommodate the reduced dynamic range associated with sensorineural hearing loss (Dillon, 2012; Kates, 2005). Soft sounds are amplified to maintain audibility and loud sounds receive less gain to ensure listening comfort, typically over the input range of 40-90 dB SPL. In hearing aids, DRC is typically implemented across multiple frequency channels to provide frequency-specific and level-dependent gain according to prescriptive targets such as NAL-NL2 (Keidser et al., 2012) or DSL v5.0 (Scollie et al., 2005) for an individual audiogram. The behavior of the compressor in each channel is controlled by several parameters: compression ratio, compression threshold, and time constants. The compression threshold is the input level above which compression is applied, and lower compression thresholds result in compression over a wider input dynamic range. The compression ratio determines the amount of gain reduction applied for levels above the threshold. For example, a 2:1 compression ratio means that a 20 dB increase in input level produces only a 10 dB increase in output level. Higher compression ratios are usually required to accommodate the narrower dynamic range for more severe hearing losses, although clinically the recommended limit is 3:1. Finally, the time constants include attack time, which determines how quickly the gain-control circuit reduces gain after a sudden increase in input level, and release time, which determines how quickly gain is restored following a decrease in input level. Together, these time constants determine the extent to which the short-term amplitude fluctuations carrying critical speech information (e.g. manner of articulation, voicing, cues for recognition of vowels) (Fogerty, 2013; Rosen, 1992) are preserved or distorted by the compressor (Jenstad & Souza, 2005; Kates, 2010).

DRC parameters that increase audibility can also increase signal distortion (see Moore, 2008; Souza, 2002 for review). For example, higher compression ratios are often required to accommodate a narrower dynamic range for more severe hearing losses, but they increase temporal envelope distortion and other forms of nonlinear distortion (Jenstad & Souza, 2007; Souza, 2002). Similarly, fast release times (<200-250 ms) restore audibility for brief, low-intensity phonemes but can also distort short-term amplitude fluctuations (Jenstad & Souza, 2005; Kates, 2010), thereby reducing the temporal envelope modulation depths (Alexander & Rallapalli, 2017; Jenstad & Souza, 2007; Stone & Moore, 2007) and spectral contrasts for multichannel compression (Bor et al., 2008; Stone & Moore, 2008).

The parameters optimized for one signal type (e.g., speech) may not be optimal for other signal types (e.g., music or noise) because the effects of compression parameters depend on the signal characteristics and the measured outcomes. Several studies have shown that fast-acting compression improves the intelligibility of soft speech either in quiet (Davies-Venn et al., 2009; Souza & Turner, 1998, 1999) or when speech is present in dips of background noise (Kowalewski et al., 2018). However, as the signal input level increases, the intelligibility benefits with fast-acting compression diminish. This decrease in benefit has been associated with the trade-off between audibility and distortion, where temporal envelope distortion dominates perception at louder levels once optimum audibility has been achieved (Davies-Venn et al., 2009). Similarly, compression has shown modest benefits in improving sound quality compared to linear amplification (Davies-Venn et al., 2007), likely due to better loudness, comfort, and pleasantness for loud inputs. However, fast-acting compression results in poorer sound quality ratings than slow-acting compression for both speech (Arehart et al., 2007, 2010, 2011, 2022; Neuman et al., 1995, 1998) and music signals across different genres (Croghan et al., 2012, 2014; Neuman et al., 1998). Depending on the genre and production techniques (Vickers, 2010), music can have a wider dynamic range and a higher peak-to-root-mean-square (RMS) ratio compared to speech, making it more vulnerable to distortions introduced by compression parameters such as higher compression ratios and fast release times (Chasin & Russo, 2004). Thus, slow-acting compression or even linear amplification may be preferred when music is the signal of interest (Croghan et al., 2014; Kirchberger & Russo, 2016).

Because no single compression strategy can simultaneously optimize audibility and signal quality across all acoustic environments, hearing aids have managed this challenge with different approaches. Several hearing aids have defaulted to either slow-acting or fast-acting DRC depending on the manufacturer’s philosophy. Slow-acting DRC is a logical choice to minimize the aforementioned distortion effects of fast-acting DRC, but it can still result in poorer audibility, especially for soft speech levels (Souza, 2002) or at lower SBRs where listeners with hearing loss struggle the most (Kowalewski et al., 2018; Moore, 2026). Next, the strong rationale for applying different compression parameters to different sound classes has motivated the development of algorithms in modern hearing aids that adaptively adjust compression parameters based on the detected sound environment such as overall level, signal onset time or speech-in-quiet vs speech-in-noise vs music scene type, although mixed benefits are reported across studies (e.g., Balling et al., 2022; Chen et al., 2021; Moore et al., 2004; Pittman et al., 2014; Rallapalli & Alexander, 2019). Although automatic scene detection preceding compression parameter selection offers some degree of adaptation to the most dominant sound source (Yellamsetty et al., 2021), once a scene is detected, the same set of compression parameters is applied to the input signal regardless of the components in the mixture. This method has a major drawback. When compression acts on a mixture of signals, nonlinear distortion effects can be produced because each signal affects the time-varying gain applied to others. Multiple time-varying signals modulate one another, inducing a negative correlation between the envelopes of otherwise unrelated sounds (Alexander & Masterson, 2015; Corey & Singer, 2021; Shen & Lentz, 2010; Stone & Moore, 2004, 2007, 2008; Walaszek, 2008). Because the gain is primarily driven by the loudest signal, conventional DRC tends to reduce positive long-term signal-to-noise ratios (SNRs; Alexander & Masterson, 2015; Hagerman & Olofsson, 2004; Rhebergen et al., 2009; Souza et al., 2006). Across-source distortion effects also reduce effective compression ratios (ECRs), especially at low SNRs, meaning that quiet signals are not mapped to the intended range (Braida et al., 1982; Souza et al., 2006; Stone & Moore, 1992). Corey and Singer (2021) mathematically modeled these three distortion effects and proved that they are unavoidable consequences of compressing sources as a mixture, regardless of implementation or parameter choices. However, the severity of distortion depends on the compression ratio and on the dynamic ranges of the signal envelopes, which differ for different sound types, channel structures, and attack and release times.

Recent studies have shown that some of the drawbacks of conventional DRC can be mitigated by applying different compression parameters to each signal within a mixture based on the type of sound source. This approach resembles media production, in which separate effects are applied to each instrument or vocal track before mixing. Currently, in commercial hearing aids this effect has been examined by combining DRC with directional microphone or beamformer outputs. Corey and Singer (2017) proposed a beamforming-based separator followed by independent compression of sources, resulting in improved acoustic measures of distortion. Rallapalli et al., (2021) reported reduced temporal envelope distortions when speech in multi-talker babble was processed with the combination of fast release time and directional microphones, relative to omnidirectional microphones, particularly at high SNRs. One manufacturer applies beamformer technology in the hearing aid to independently process signals from the front (presumably speech) with linear gain and signals from the rear with more compression and noise reduction. Studies have reported that this split processing with the beamformer output resulted in lower speech recognition thresholds (better) in a restaurant noise scenario (Jensen et al., 2021) and reduced listening effort based on electroencephalogram responses (Slugocki & Korhonen, 2022), relative to conventional DRC applied to the mixed input from directional microphones. Folkeard et al. (2024) also reported positive subjective ratings on the domains of speech intelligibility, sound quality, loudness, listening effort, attentional focus, source localization, background noise reduction, and overall satisfaction with this split processing technology in older adult hearing aid users during real-world group conversations in a noisy food court environment. However, the effectiveness of source separation with microphones is limited by the spatial configuration of sources in the environment and number of microphones in the array.

Other researchers have proposed monophonic source separation techniques, including time-frequency masks and deep neural networks, to apply different compression parameters to different sound sources without knowledge of the signal directions. One approach, known as scene-aware compression, applies compression in individual time-frequency (T-F) units depending on the presence of the target signal in background noise or reverberation (May et al., 2018). Fast-acting compression is applied to T-F units dominated by speech, and a slower release time is applied to units dominated by background noise or reverberation. Compared to the conventional fast-acting DRC, the scene-aware system resulted in similar effective compression ratios and speech modulation depths while improving the broadband SNR and reducing spatial distortions (May et al., 2018). Overby et al. (2023) further showed that scene-aware compressors exhibited more acoustic similarity to ideal reference systems compared to fast and slow compressors, especially under adverse conditions such as low SNRs and reverberant environments. Moreover, behavioral studies have shown better speech intelligibility and subjective preference with scene-aware strategies compared to conventional DRC systems with either slow- or fast-acting compression alone (Kowalewski et al., 2020). Additionally, when combined with deep neural network (DNN)-based single-channel noise reduction, scene-aware compression was rated similarly to an ideal reference system (Overby et al., 2025). More recently, Zhang et al. (2025) developed a Neural-WDRC strategy that used DNNs to estimate speech and noise individually. Subsequently, fast-acting compression was applied to the estimated speech alone and slow-acting compression was applied to the estimated noisy portions of the signal, while maintaining a certain noise level for environmental sound awareness. Their study showed that speech and noise signals processed with Neural-WDRC had higher signal fidelity, exhibited reduced cross-modulation and spectral distortions, and were preferred over conventional DRC methods by listeners with normal hearing and relatively flat 40 dB hearing loss. Most recently, Rallapalli et al., (2021) demonstrated that independent compression of speech and music sources yielded better sound quality ratings in listeners with audiometrically normal hearing under ideal conditions where ground-truth sound sources were available to the compressor. Specifically, independent DRC improved overall sound quality and speech clarity over conventional DRC when speech was softer than the background music, without negatively affecting the speech and music naturalness ratings.

Across studies there are some notable gaps. First, while most of these studies focused on speech intelligibility or pairwise listener preferences in noise, fewer studies have focused on the effects of independent DRC on sound quality. Sound quality is an important outcome measure for a few reasons. Sound quality is directly associated with hearing aid benefit and satisfaction ratings (Bannon et al., 2023; Picou, 2022) and more recently has been shown to be associated with listening-related fatigue and mood (Picou et al., 2026). Furthermore, with regard to hearing aid signal processing such as DRC, poorer sound quality ratings are associated with increasing levels of distortion (i.e., due to higher compression ratios and faster compression time constants which may be considered optimal for improved speech audibility) (e.g., Arehart et al., 2022; Bannister et al., 2024; Souza et al., 2015), and sound quality ratings have been shown to be sensitive to subtle changes in low-level distortions even when speech intelligibility is at a maximum (Arehart et al., 2022). Thus, sound quality may provide a more sensitive and ecologically valid outcome measure for distortions introduced by DRC than speech intelligibility alone.

Next most studies have focused on noise as the background signal, probably because noisy environments are the most challenging for hearing aid users. However, this also presents an important gap because signals other than speech in noise that are important for the listener pose challenges for hearing aid compression. As noted previously, music often has a wider dynamic range and a higher peak-to-RMS ratio compared to speech and noise (Chasin & Russo, 2004). Thus, compression parameters such as fast release time that are optimal for speech audibility could result in greater gain fluctuations, temporal envelope distortions, variations in timbre and spectral density for music. Moreover, in contrast to noisy backgrounds that are generally considered unwanted signals, music when present along with speech may often be a signal of interest. In fact, music listening is an important activity for many hearing aid users, yet it is often cited as source of dissatisfaction (Greasley et al., 2020; Leek et al., 2008; Looi et al., 2019). Thus, understanding how independent DRC affects speech in music mixtures is important for optimizing hearing aid performance.

Third, prior studies examining the effects of independent DRC in listeners with hearing loss have largely been restricted to controlled audiometric configurations, such as flat 40 dB HL or audiograms matched to a specific profile (e.g., moderate to severe loss). The one prior study systematically examining sound quality as an outcome with independent DRC was conducted on listeners with audiometrically normal hearing (Rallapalli et al., 2021). However, it is of clinical interest to evaluate independent DRC in listeners with the range of hearing losses typically encountered in clinical practice, and to determine whether the sound quality benefits observed in previous work extend to this population. Studies have shown that listeners with hearing loss rate sound quality differently than listeners with normal hearing across a range of signal processing conditions, suggesting that findings from normal hearing populations may not generalize directly. For example, Hansen (2002) reported that listeners with hearing loss preferred lower compression thresholds and a slightly shorter release time compared to listeners with normal hearing for subjective intelligibility of speech signals. Moreover, there were significant differences in preference across hearing aid DRC conditions for nonspeech signals (such as music and noise) for listeners with hearing loss, but not for listeners with normal hearing. Similarly, Arehart et al. (2011) found that listeners with hearing loss rated speech quality lower across a wide range of hearing aid processing conditions and were more sensitive to compression and additive noise compared to listeners with normal hearing. The differences in sound quality ratings between the groups likely reflect the underlying auditory processing changes with sensorineural hearing loss. Compared to listeners with normal hearing, listeners with hearing loss experience reduced audibility, lower temporal and spectral resolution (Ponsot et al., 2021; Zaar et al., 2023), altered loudness growth (Marozeau & Florentine, 2007; Moore & Glasberg, 1997), and greater susceptibility to temporal envelope distortions (Souza et al., 2015). These differences may alter the perceptual consequences of compression on sound quality relative to listeners with normal hearing (Blamey & Martin, 2009; Moore & Sęk, 2016; Shi et al., 2007) and the magnitude of benefit from independent DRC, warranting direct empirical investigation in listeners with hearing loss.

The present study addresses these gaps in the following ways. First, the study examines the effects of independent DRC on sound quality as the primary outcome measure for speech mixed with both music and noise backgrounds, extending the scope of prior work beyond speech intelligibility and noise-only conditions. Second, the study was conducted on listeners with a range of hearing losses typical of hearing aid candidates, with all listeners receiving individually prescribed frequency shaping. Third, the study used ground-truth source signals, such as the type that may be available from a remote microphone or streaming accessory. This approach decouples the benefits of independent DRC from the limitations of any specific source separation algorithm. Therefore, the results could be generalized to any current or future source separation technology. Using a within-subjects design, this study compared the sound quality ratings between conventional DRC and independent DRC conditions for speech mixed with music or noise backgrounds. We hypothesized that independent DRC would result in better sound quality ratings than conventional DRC for speech mixed with music or background noise, particularly when speech was softer than the background. Linear amplification was included as a control condition to isolate the effect of compression itself, and it was expected that mixing signals before and after linear amplification would yield no differences in sound quality ratings.

Methods

Participants

Twenty individuals (7 females, 13 males) with mild to moderately severe bilateral sensorineural hearing loss, aged 24–84 years (M = 69.05, SD = 14.54) participated in the study. Air-conduction thresholds were recorded at octave frequencies ranging from 250 Hz to 8000 Hz. Bone conduction thresholds were recorded at octave frequencies ranging from 500 Hz to 4000 Hz. All participants had symmetric audiograms for both ears. Asymmetry was defined as a difference of > 15 dB HL at any frequency between 250 and 4000 Hz. The average four frequency pure tone average (PTA) for the right ear was 43.31 dB HL and for the left ear was 43.76 dB HL. A paired t-test indicated no significant difference in PTA between the ears (t=-0.47, p=0.64). Figure 1 shows the air conduction thresholds of the participants. Note that the left ear 3000 Hz air conduction threshold was not obtained for one participant; however, since this frequency was not used to set hearing aid simulator gains, it did not affect that participant’s results.

OPEN IN VIEWER

Figure 1.

Air-conduction thresholds for participants in the right and left (test) ears. The solid line indicates average thresholds across participants for each ear

The participants had no evidence of middle ear pathology based on air-bone gaps within 15 dB HL across test frequencies and tympanometry. All participants spoke English as their primary language, reported no neurological disorders, and had normal or corrected-to-normal vision based on their self-reports. Participants completed the Montreal Cognitive Assessment (MoCA; Nasreddine et al., 2005) within 12 months of recruitment and received a passing score of >22 (Luis et al., 2009; Pugh et al., 2018; Rossetti et al., 2011). All study procedures were reviewed and approved by the Institutional Review Board, and all participants signed an informed consent form for participation.

Stimuli

Stimuli included speech mixed with music or noise. The inclusion of two different background sounds with varied modulation properties allowed us to evaluate the effects of DRC strategies in different acoustic contexts (Lundberg et al., 2020; Moore & Tan, 2003). Moreover, the two background signals are different in that, music may be considered a signal of interest along with speech, whereas noise is generally an unwanted signal. Speech and music signals were identical to those reported by Rallapalli et al., (2021). Briefly, the speech signal consisted of a 10-second segment of spontaneous speech produced by a female speaker from the ALLSTAR database (Bradlow, 2026). The music signal was a classical music excerpt from Beethoven’s 7^th Symphony. The noise signal was speech-shaped noise matched in the long-term average speech spectrum to the speech signal. Both the music and noise signals were matched in duration with the speech signal. Figure 2 shows the long-term average spectrum for speech, music, and noise signals. To minimize variability in speech intelligibility, a single speech sample was used across all listening conditions. Because the effects of compression are known to vary with relative signal levels in a mixture (Alexander & Masterson, 2015; Naylor & Johannesson, 2009; Rhebergen et al., 2009), we tested three speech-to-background ratios (SBRs). The speech level was held constant at 65 dB SPL, and the background levels were varied to achieve -10, 0, and +10 dB SBRs.

OPEN IN VIEWER

Figure 2.

Long-term average spectra for speech, music, and steady noise signals in the study at 65 dB input level

Stimuli were presented through Sennheiser HD 25 supraaural headphones and sounds were routed from the computer through the U24 1&2 ESI Audio Device (WDM). A 30-second long concatenated speech in music or speech in noise signal at 0 dB SBR, digitally scaled to a peak amplitude of +/- 1, was used for calibration with a Larson and Davis Model 831 sound-level meter and an artificial ear simulator (AEC201-A). All reported signal levels were A-weighted. The signals were digitally scaled down to an input of 65 dB SPL relative to a maximum output of 93.3 dB SPL. The final presentation level of the signals in the experiment was based on individualized gain functions computed from the left ear air-conduction thresholds as described under ‘Hearing aid processing’. Processed signals were presented monaurally to the left ear to minimize the potential binaural effects on sound quality perception, such as binaural loudness summation (Balfour & Hawkins, 1992) and to allow for comparison with previous research (Rallapalli et al., 2021).

Hearing Aid Processing

A hearing aid simulator implemented in MATLAB was used to apply the multichannel DRC across 12 channels (for complete descriptions, see Arehart et al., 2022; Kates et al., 2019). The channel center frequencies were 125, 500, 375, 500, 750, 1000, 1500, 2000, 3000, 4000, 5000, and 6000 Hz. Prior to applying the compression, an envelope peak detector was used to estimate the overall level of each channel. Level-dependent gain was applied above a knee point of 45 dB SPL. Unlike Rallapalli et al., (2021), where compression ratios and gain were applied for a simulated mild-to-moderate hearing loss for the DRC strategies, in this study, the individualized frequency shaping was provided with NAL-R to compensate for audibility loss and the compression ratio was fixed at 2:1. This allowed us to maintain the same gain across amplification types for a given listener. Gains were applied based on left ear air conduction thresholds at 250 Hz, 500 Hz, 1000 Hz, 2000 Hz, 4000 Hz, and 6000 Hz. The attack time was fixed at 5 ms and the release time varied depending on the combination of the signal mixture and DRC strategy. No vent effects were simulated to dissociate the influence of open-ear acoustics on the sound quality ratings (Winkler et al., 2016).

The two DRC strategies differed in the stage at which the compression was applied, as shown in Figure 3. For independent DRC conditions, the speech and background signals were first separately compressed using different parameters and then mixed. Release time parameters were selected to maximize the presumed benefit for each signal type. A fast release time of 50 ms was applied to the speech signal to maximize audibility, whereas a slow release time of 2000 ms was applied to the music or noise signals as slower release times are more appropriate for relatively steady-state background signals and/or to minimize gain fluctuations (Moore, 2008). For the conventional DRC conditions, the speech and background (music or noise) were mixed, and then compression was applied to the joint signals. A fast release time of 50 ms was applied to maximize the audibility of speech signals. Moreover, a fast release time is more susceptible to cross-modulation distortion than slow release time (Stone & Moore, 2007, 2008) and therefore represents the condition under which maximum differences between independent and conventional DRC, and in turn, sound quality outcomes would be expected. Within each mixing strategy, linear amplification (or a compression ratio of 1) with NAL-R gains was included as a control condition. Linear processing produces identical signals whether applied before or after mixing.

OPEN IN VIEWER

Figure 3.

Schematic depicting the stages at which compression parameters are applied for independent and conventional dynamic range compression (DRC)

Results

Acoustic Metrics

Table 1 presents the results of the acoustic analyses across the three metrics and signal types. The overall patterns remained the same regardless of the background signal type. As expected, the ECRs for a given signal were lower with the conventional DRC than with the independent DRC. There was some dependency on the SBR such that with conventional DRC, the ECRs for speech increased, whereas the ECRs for the background signal decreased as the SBR increased from -10 to +10 dB. This effect was expected because compression gains tend to be driven by the louder signal in a mixture. On the other hand, with independent DRC, the speech ECR remained consistent across the input SBRs because the speech input level was fixed at 65 dB SPL. The music ECR varied as its input level changed between 55- and 75-dBSPL, which is the expected behavior of a knee-shaped compression curve. Similar to speech, background music was compressed more consistently with independent DRC than with conventional DRC. Because stationary noise has a negligible input dynamic range, it would not have been meaningful to measure its ECR. The effects on the long-term output SBR also varied with the DRC strategy and input SBR. At -10 dB SBR, the output SBR was ∼2-3 dB higher with the independent DRC than with the conventional DRC, whereas at +10 dB SBR, the output SBR was ∼1 dB lower with the independent DRC than with the conventional DRC. At 0 dB input SBR, the differences in the output SBR between the two DRC strategies were minimized. With regard to ASMC, the speech and background output signals had stronger negative cross-correlations with the conventional DRC, whereas the correlations were negligible (∼0) with the independent DRC. This across-source distortion effect was strongest with speech-shaped noise at higher SBRs, as the loud parts of the speech signal drove modulations in otherwise-steady noise. Finally, the signal mixing type had no effect on any of the acoustic metrics when linear gain was applied to the signals. That is, the signals were acoustically identical regardless of whether the gain was applied before or after mixing, as expected.

OPEN IN VIEWER

Table 1.

Acoustic Metrics Across Test Signals: Effective Compression Ratio (ECR), Across Source Modulation Correlation (ASMC), and Long-Term Output (O/P) Speech-To-Background Ratio (SBR). The Processing Conditions Include Amplification Type: Dynamic Range Compression (DRC) or Linear Amplification (Lin.); Mixing Strategy: Conventional (Con.: Signals Mixed before Amplification) or Independent (Ind.: Signals Mixed after Amplification); SBRs: -10, 0, +10 dB

Background signal	Amplification type	SBR (dB)	Mixing strategy	Speech ECR	Background ECR	ASMC (Power)	O/P SBR (dB)
Music	DRC	-10	Con.	1.09	1.70	-0.30	-9.16
Music	DRC	0	Con.	1.16	1.32	-0.31	-2.00
Music	DRC	+10	Con.	1.20	1.05	-0.29	5.48
Music	DRC	-10	Ind.	1.22	1.44	0.01	-6.99
Music	DRC	0	Ind.	1.22	1.39	0.01	-1.69
Music	DRC	+10	Ind.	1.22	1.25	0.01	4.55
Music	Lin.	-10	Con.	1.00	0.99	0.06	-10.92
Music	Lin.	0	Con.	1.00	0.99	0.06	-0.62
Music	Lin.	+10	Con.	1.00	0.99	0.06	9.38
Music	Lin.	-10	Ind.	1.00	0.99	0.06	-10.92
Music	Lin.	0	Ind.	1.00	0.99	0.06	-0.62
Music	Lin.	+10	Ind.	1.00	0.99	0.06	9.38
Noise	DRC	-10	Con.	1.03	1.26	-0.27	-11.27
Noise	DRC	0	Con.	1.10	0.75	-0.59	-3.28
Noise	DRC	+10	Con.	1.20	0.41	-0.67	4.67
Noise	DRC	-10	Ind.	1.22	1.00	-0.01	-7.88
Noise	DRC	0	Ind.	1.22	1.00	0.00	-2.76
Noise	DRC	+10	Ind.	1.22	1.00	0.00	3.66
Noise	Lin.	-10	Con.	1.00	0.99	0.02	-10.69
Noise	Lin.	0	Con.	1.00	0.99	0.02	-0.69
Noise	Lin.	+10	Con.	1.00	0.99	0.02	9.31
Noise	Lin.	-10	Ind.	1.00	0.99	0.02	-10.69
Noise	Lin.	0	Ind.	1.00	0.99	0.02	-0.69
Noise	Lin.	+10	Ind.	1.00	0.99	0.02	9.31

Sound Quality

Figures A & B (Appendix B) show the full distribution of raw sound quality ratings for speech in music and speech in noise signals, across the three subscales. The observed trends showed that the ratings were generally higher as the SBR increased from -10 to +10 dB. Ratings on all three subscales were relatively higher with the DRC conditions than with linear amplification, especially at -10 dB and 0 dB SBRs. Ratings with linear amplification were observed to be slightly better than or similar to those with DRC at +10 dB SBR. Similarly, the observed ratings with independent DRC were higher than those with conventional DRC for all subscales, particularly at -10 dB and 0 dB SBR. However, there were no observable differences in ratings between the conventional and independent mixing of signals with linear amplification, as expected.

Across all LME models, initial analysis showed that the main effect of repetition was not significant (p > 0.05). Thus, repetition was dropped and the final models present results from ratings averaged across the two presentations within a given condition. Tables 2 and 3 show the Type III tests for the fixed effects in the final LME models for the speech in music and speech in noise conditions, respectively. The models showed significant effects of SBR, amplification type, and mixing strategy on sound quality ratings within each subscale for speech in music signals. The LME models revealed significant main effects of SBR and mixing strategy across all sub-scale ratings for speech in noise signals. Additionally, the noisiness scale showed a significant main effect of amplification type. However, because two-way and/or three-way interactions between these factors were significant within each model, the description of the results focuses on the post-hoc analysis conducted to explore the nature of these interactions. PTA and Age did not have a significant effect on the ratings for any of the sound-quality subscales.

OPEN IN VIEWER

Table 2.

Type III Tests of Fixed Effects From the Linear Mixed Effects Model Examining the Effects of Amplification Type (Ref = Linear), Mixing Strategy (Ref = Conventional) and Speech-To-Background Ratio (SBR; Ref = -10 dB) on Overall Sound Quality, Speech Clarity, and Music Pleasantness Ratings for the Speech In Music Signals. Each Scale was Modeled Separately. Num DF = Numerator Degrees of Freedom, Den DF = Denominator Degrees of Freedom, PTA = Four-Frequency Pure Tone Average in Test Ear

Scale	Effect	Num DF	Den DF	F-value	p-value
Overall Sound Quality	SBR	2	449	311.65***	<.001
	Amplification type	1	449	23.13***	<.001
	Mixing strategy	1	449	21.35***	<.001
	SBR * Amplification type	2	449	35.63***	<.001
	SBR * Mixing strategy	1	449	1.14	0.322
	Amplification type * Mixing strategy	1	449	31.70***	<.001
	SBR * Amplification type * Mixing strategy	2	449	2.86	0.059
	PTA	1	449	1.85	0.175
	Age	1	449	3.35	0.068
Speech Clarity	SBR	2	449	744.41***	<.001
	Amplification type	1	449	17.64***	<.001
	Mixing strategy	1	449	89.14***	<.001
	SBR * Amplification type	2	449	39.27***	<.001
	SBR * Mixing strategy	2	449	14.99***	<.001
	Amplification type * Mixing strategy	1	449	71.44***	<.001
	SBR * Amplification type * Mixing strategy	2	449	8.32***	<.001
	PTA	1	449	0.08	0.771
	Age	1	449	2.35	0.126
Music Pleasantness	SBR	2	449	198.72***	<.001
	Amplification type	1	449	187.12***	<.001
	Mixing strategy	1	449	5.38*	0.021
	SBR * Amplification type	2	449	66.72***	<.001
	SBR * Mixing strategy	2	449	3.31*	0.037
	Amplification type * Mixing strategy	1	449	22.88***	<0.001
	SBR * Amplification type * Mixing strategy	2	449	2.0	0.137
	PTA	1	449	0.05	0.816
	Age	1	449	0.01	0.927

***p<0.001, **p<0.01, *p<0.05.

OPEN IN VIEWER

Table 3.

Type III Tests of Fixed Effects From the Linear Mixed Effects Model Examining the Effects of Amplification Strategy (Ref = Linear), Mixing Strategy (Ref = Conventional) and Speech-To-Background Ratio (SBR; Ref = -10 dB) on Overall Sound Quality, Speech Clarity, and Noisiness Ratings for the Speech In Noise Signals. Each Scale was Modeled Separately. Num DF = Numerator Degrees of Freedom, Den DF = Denominator Degrees of Freedom, PTA = Four-Frequency Pure Tone Average in Test Ear

Scale	Effect	Num DF	Den DF	F-value	p-value
Overall Sound Quality	SBR	2	449	436.00***	<.001
	Amplification type	1	449	0.26	0.610
	Mixing strategy	1	449	8.83**	0.003
	SBR * Amplification type	2	449	42.55***	<.001
	SBR * Mixing strategy	2	449	2.54	0.080
	Amplification type * Mixing strategy	1	449	23.20***	<0.001
	SBR * Amplification type * Mixing strategy	2	449	0.20	0.817
	PTA	1	449	0.58	0.448
	Age	1	449	1.13	0.288
Speech Clarity	SBR	2	449	893.33***	<.001
	Amplification type	1	449	3.31	0.070
	Mixing strategy	1	449	58.05***	<.001
	SBR * Amplification type	2	449	71.58***	<.001
	SBR * Mixing strategy	2	449	6.83**	<.001
	Amplification type * Mixing strategy	1	449	62.94***	<.001
	SBR * Amplification type * Mixing strategy	2	449	3.46*	0.032
	PTA	1	449	0.69	0.407
	Age	1	449	0.20	0.654
Noisiness	SBR	2	449	607.31***	<.001
	Amplification type	1	449	56.65***	<.001
	Mixing strategy	1	449	25.11***	<.001
	SBR * Amplification type	2	449	73.95***	<.001
	SBR * Mixing strategy	2	449	1.13	0.325
	Amplification type * Mixing strategy	1	449	26.23***	<.001
	SBR * Amplification type * Mixing strategy	2	449	1.24	0.292
	PTA	1	449	0.12	0.733
	Age	1	449	0.08	0.781

***p<0.001, **p<0.01, *p<0.05.

Figures 4–7 depict marginal means and 95% confidence intervals from the LME models for the key interactions of interest (i.e., interactions involving mixing strategy and amplification type) for each signal mixture. Figures C-E (Appendix C) show the marginal means and 95% confidence intervals for the remaining significant interaction effects.

OPEN IN VIEWER

Figure 4.

Marginal means from a linear mixed-effects model illustrating the two-way interaction between amplification type (Dynamic range compression or DRC, Linear; X-axis) and mixing strategy (independent, conventional; see legend) on sound quality ratings for speech in music signals. Each panel represents a different sound quality sub-scale: Overall sound quality (Overall), Music pleasantness. Significant (*** p < 0.001) and non-significant (N.S.) pairwise comparisons between mixing strategies within each amplification type are shown

OPEN IN VIEWER

Figure 5.

Marginal means from a linear mixed-effects model illustrating the three-way interaction between amplification type (Dynamic range compression or DRC, Linear; X-axis), mixing strategy (independent, conventional; see legend), and input signal-to-background ratio (SBR; -10 dB, 0 dB, +10 dB; see different panels) on speech clarity ratings for speech in music signals. Significant (*** p < 0.001) and non-significant (N.S.) pairwise comparisons between mixing strategies within each amplification type and SBR are shown

OPEN IN VIEWER

Figure 6.

Marginal means from a linear mixed-effects model illustrating the two-way interaction between amplification type (Dynamic range compression or DRC, Linear; X-axis) and mixing strategy (independent, conventional; see legend) on sound quality ratings for speech in noise signals. Each panel represents a different sound quality sub-scale: Overall sound quality (Overall), Noisiness. Significant (*** p < 0.001) and non-significant (N.S.) pairwise comparisons between mixing strategy within each amplification type are shown

OPEN IN VIEWER

Figure 7.

Marginal means from a linear mixed-effects model illustrating the three-way interaction between amplification type (Dynamic range compression or DRC, Linear; X-axis), mixing strategy (independent, conventional; see legend), and input signal-to-background ratio (SBR; -10 dB, 0 dB, +10 dB; see different panels) on speech clarity ratings for speech in noise signals. Significant (*** p < 0.001) and non-significant (N.S.) pairwise comparisons between mixing strategy within each amplification type and SBR are shown

Discussion

The purpose of this study was to compare the effects of independent and conventional DRC on sound quality ratings among listeners with hearing loss. Sound quality ratings were obtained for two types of signal mixtures, including speech in music and speech in noise, at three SBRs. The study found that independent DRC consistently resulted in higher sound quality ratings than conventional DRC for both types of signal mixtures. Specifically, when compression was applied independently to the speech and background signals, listeners assigned higher ratings for overall sound quality, speech clarity, music pleasantness, and noisiness (higher ratings = less noisy) than when the signals were compressed jointly after mixing. As expected, this pattern of results did not emerge with linear amplification because the signals were acoustically identical when the gain was applied before (independent) and after (conventional) mixing. The sound quality results also agree with the acoustic metrics showing a reduced ASMC, indicating reduced cross-modulation distortions between independently compressed signals. Generally, higher ECRs and higher output SBRs were obtained for the independently compressed signals, indicating improved audibility of the softer input signal in the mixture relative to the conventional DRC.

The sound quality ratings with the different mixing strategies for DRC depended on the SBR for certain subscales. Primarily, a three-way interaction between mixing strategy, SBR, and amplification type was observed for speech clarity ratings. Independent DRC had significantly higher speech clarity ratings than conventional DRC at -10 dB and 0 dB SBRs when the speech was softer or equal in level to the background signal. Because these are conditions where speech audibility is low, independent DRC was more effective than conventional DRC in increasing the gain for the speech signal with the same release time of 50 ms, as evidenced by higher speech ECRs and higher output SBR measurements. Notably, the size of the effect decreased as the SBR increased from -10 dB to 0 dB SBR, and there was no significant difference in speech clarity ratings between independent DRC and conventional DRC at +10 dB SBR, where the speech was louder than the background signal. This pattern of results at higher SBRs suggests a ceiling effect such that the benefits of independent DRC on speech clarity can diminish once the speech is sufficiently audible (Davies-Venn et al., 2009). These results align with the output SBR measurements showing a difference of ∼ 1 dB between independent and conventional DRC, indicating no measurable advantage for speech audibility with independent DRC at the highest input SBR. In fact, the overall quality ratings with DRC (regardless of the mixing strategy) were significantly lower relative to linear amplification at +10 dB SBR. There was also an observed trend towards lower speech clarity with DRC compared to linear amplification at +10 dB SBR (Appendix B – Figures A & B). With both DRC strategies, there was likely distortion of the temporal envelope of the speech with a fast release time, which may have affected sound quality perception relative to linear amplification at positive SBRs (Arehart et al., 2007, 2011, 2022). Moreover, at +10 dB input SBR, softer background signals were provided with more gain relative to the speech signal with the DRC strategies, and the output SBR measures were ∼ 4-6 dB lower, contributing to the reduced speech clarity compared to linear amplification. It is worth pointing out that the increase in audibility of the softer background signal may be an intended effect of compression in some listening situations and not necessarily a drawback if the signal of interest for the listener is background music.

The independent mixing of signals also resulted in higher ratings for music pleasantness for speech in music signals at the lowest SBR (-10 dB). It is likely that this result was primarily driven by DRC because the linear amplification conditions were acoustically identical. These findings are only partially explained by acoustic metrics. At this SBR, music was louder than speech, and independent DRC applied less compression (lower ECR) to the music signal compared to conventional DRC. Based on previous research and the higher ratings for both DRC strategies relative to linear amplification, one might expect that less compression at high input levels would be unfavorable for music pleasantness (e.g., Davies-Venn et al., 2009). However, listeners may also have been sensitive to cross-modulation distortions (Stone & Moore, 2004), which were negligibly low with independent DRC, potentially explaining the higher ratings for music pleasantness compared to conventional DRC at -10 dB SBR. Moreover, an added advantage of independent DRC was the use of a slow release time for music signals, which would have preserved the temporal envelope and sound quality relative to conventional DRC with a fast release time (Moore & Sęk, 2016). At higher SBRs, neither DRC strategy affected pleasantness ratings, likely because the music was at a comfortable level (55 and 65 dB SPL). Notably, the reduced ratings of music pleasantness with linear amplification at -10 dB and 0 dB SBR compared with both DRC strategies appear contradictory to these results. One possible explanation is that because DRC and linear amplification conditions were evaluated within the same experimental block, listeners may have formed judgments relative to other conditions presented. In this context, the relative loudness or dynamic characteristics of music with linear amplification may have sounded comparatively unpleasant alongside compressed alternatives (e.g, Croghan et al., 2012).

Finally, listeners rated the speech in noise signals as less noisy with independent DRC than with conventional DRC, regardless of the SBR. This result corresponded with the higher output SBRs measured at -10 dB and 0 dB SBRs with independent DRC relative to the conventional DRC. When compression parameters were applied to the signals before mixing, more gain was applied to the softer speech signal, and less gain was applied to the louder noise signal at -10 dB SBR. When speech and noise levels were equal at 0 dB SBR, conventional DRC amplified the noise in the dips of the speech signal (Naylor & Johannesson, 2009; Souza et al., 2006), whereas this effect did not occur with independent DRC, creating the perception of reduced noisiness with the latter. At a positive input SBR (+10 dB), although the output SBR with independent DRC was lower (owing to more gain being applied to the softer noise signal relative to the louder speech signal), the reduced cross-modulation distortions (ASMC values ∼0) combined with slow release time for noise may have reduced the perception of noisiness compared to conventional DRC.

The secondary analysis provided additional insight into the sub-scales (perceptual dimensions) underlying the sound quality benefits observed across both signal mixtures (Appendix D). For speech in music signals, pleasantness was a significant predictor of overall sound quality, though the contribution was smaller than that of speech clarity. Thus, it is possible that while the benefits of music pleasantness at low SBRs contributed to overall sound quality, the advantage of independent DRC was likely driven by improvements in speech clarity (corresponding to increases in speech audibility). On the other hand, for the speech in noise signals noisiness and speech clarity contributed equally to overall sound quality, indicating that the benefit of independent DRC in background noise was driven by lower noisiness as well as improved speech clarity rather than one dimension alone. The differential weighting of subscales across signal mixtures suggests that multiple dimensions can provide insights into sound quality benefits with independent DRC rather than relying solely on overall sound quality ratings or paired comparisons.

The present study included listeners with a clinically representative range of hearing losses to evaluate whether sound quality benefits observed in prior work extend to this population. The pattern of results showing higher overall sound quality and speech clarity ratings at the lowest SBRs with independent DRC of speech in music signals was broadly consistent with the normal-hearing group tested in Rallapalli et al., 2021. The present study additionally indicated that these benefits may extend to higher SBRs, speech in noise signals, music pleasantness ratings, and reduced noisiness. The broader range of significant effects, especially at higher SBRs, could be for two reasons. The use of a fixed compression ratio of 2:1 across channels in the present study compared with frequency-specific compression ratios based on NAL-NL2 (often < 2:1, especially in the low-frequency channels) in Rallapalli et al., 2021 may have contributed to the favorable effects at higher SBRs. Additionally, it is possible that listeners with hearing loss were more sensitive to the improved speech audibility and reduced distortion with independent DRC compared to the listeners with normal hearing (e.g., Arehart et al., 2011; Hansen, 2002) for whom audibility may have been less of a limiting factor. Overall, the favorable results with these settings in the present study suggest that listeners with hearing loss may tolerate greater compression, possibly due to the reduced nonlinear distortion effects (e.g. Rallapalli et al., 2021), and may take advantage of the audibility improvements with independent DRC.

Limitations & Future Directions

As a proof of concept for sound quality assessment of independent DRC in listeners with hearing loss, the study used highly controlled compression parameters and short-duration signals. With regard to compression parameters, the focus was on fast-acting compression alone because it is more favorable for speech audibility, (Davies-Venn et al., 2009; Kowalewski et al., 2018; Souza & Turner, 1998, 1999), which is presumably the primary goal for hearing aids. However, the same parameters can result in more cross-modulation distortion with conventional DRC (Stone & Moore, 2007, 2008). For these reasons, fast-acting compression was most likely to reveal differences between independent and conventional DRC. However, this choice limited direct comparisons with slow-acting or adaptive compression strategies that may be used in certain commercial hearing aids. Linear amplification was included as a control condition representing an extreme case of slow-acting DRC, but fine-grained comparisons such as independent fast-acting vs conventional slow-acting DRC would provide a more comprehensive characterization of the effects of DRC time constants on sound quality with the different mixing strategies. Furthermore, because audibility and cross-modulation distortions are inherently confounded under fast-acting DRC, particularly at lower SBRs where speech audibility is most affected, it is difficult to isolate the specific contribution of reduced cross-modulation distortion to the observed sound quality improvements. Future studies should therefore dissociate the effects of audibility and distortion by controlling the output SBR along with other combinations of DRC time constants (e.g., May et al., 2018; Overby et al., 2023) to better understand the mechanisms underlying the benefits of independent DRC for sound quality.

On a related note, the study did not specifically examine the loudness dimension of sound quality. Independent DRC may alter the perceived loudness balance between the speech and background relative to conventional DRC. Loudness is particularly relevant for music listening and at least one study has shown a preference for DRC over no compression for perceived loudness of classical and rock music (Croghan et al., 2012). It is also possible that loudness contributed to differences in music pleasantness and noisiness ratings across conditions (Davies-Venn et al., 2007). Future research will incorporate loudness as a separate sound quality dimension to determine the relative contributions of audibility, loudness balance, and distortion reduction for sound quality judgments with independent DRC.

Previous studies on independent DRC have primarily focused on speech intelligibility outcomes, whereas the present study focused on sound quality and its perceptual dimensions. Incorporating measures of speech intelligibility alongside sound quality ratings would also help fully characterize the relative benefits of independent DRC and confirm whether sound quality and intelligibility improvements correspond with each other. Furthermore, the study used short-duration signals to maintain a high level of experimental control, which may limit generalizability. Future studies should incorporate more realistic signals such as discourse and a variety of background sound types to improve the generalizability of findings and better understand how independent DRC interacts with different acoustic environments (e.g., Lundberg et al., 2020; Overby et al., 2025). Similarly, although this study presented signals monaurally, source-specific compression may have different effects in immersive binaural presentations, where the sources are spatially separated. Since there were no spatial cues included in the signals in this study and they were presumed to be colocated, we expect that the overall pattern of results would remain the same with binaural presentation under the conditions tested. However, future work will examine sound quality with source-specific compression in more ecologically valid binaural listening conditions.

Next, given the longstanding evidence that the compressor behavior differs systematically at positive and negative SBRs (Alexander & Masterson, 2015; Naylor & Johannesson, 2009; Souza et al., 2006), three representative levels (-10, 0, +10 dB) were included to capture the range of conditions under which independent DRC would be expected to have differential effects on sound quality. However, this large spread may have introduced contrast effects between successive trials. Although conditions were randomized to minimize systematic order effects, and SBRs were included as a variable in the statistical models to parse the independent effects of compression conditions, carryover effects from large trial-to-trial SBR differences cannot be completely ruled out. This may have contributed to some of the variance observed in sound quality ratings (e.g., the contrasting effects of linear amplification and independent DRC on music pleasantness). Future studies will consider inclusion of a practice phase to stabilize listener criteria prior to testing.

Finally, we acknowledge potential limitations due to the modest sample size in the present study. The repeated-measures design provided multiple observations per participant for estimating the within-subjects effects, including the primary two-way interaction between mixing strategy and amplification type which was significant across models. However, the sample size may have had limited sensitivity to higher-order interactions. For example, it is possible that a larger sample size may have revealed variations in the benefit of independent DRC vs conventional DRC across SBRs for sound quality subscales other than speech clarity.

To evaluate the sensitivity of the sample size to detect effects of between-subjects covariates such as PTA and age, we conducted a simulation-based sensitivity analysis in SAS 9.4 using the observed repeated measures structure and variance components from each fitted model (Faul et al., 2007; Johnson et al., 2015). The analysis showed that models had reasonable sensitivity to detect relatively large age effects. Specifically, >80% power was obtained for age effects corresponding to a 0.73-point difference on the rating scale per 1 SD increase in age. However, the models had limited sensitivity to the effects of PTA. Even for the largest simulated PTA effect, corresponding to a 0.34-point difference per 1 SD increase in PTA, the estimated power remained below 50% across all models. Thus, the absence of significant PTA or age effects should be interpreted with caution as it may reflect limited sensitivity given the sample size and model structure, rather than evidence that degree of hearing loss or age had no influence on sound quality judgments. Moreover, future studies with larger sample sizes and a wider range of audiometric variability are needed to specifically evaluate whether degree of hearing loss moderates the benefit of independent DRC.

Summary

Taken together, findings suggest that independent DRC offers sound quality advantages over conventional DRC for listeners with hearing loss. These benefits were observed across both music and noise backgrounds, with the music background representing a listening context and signal type that was not previously examined in independent DRC literature for individuals with hearing loss. Moreover, sound quality and its sub-scales represent perceptual outcomes distinct from the speech intelligibility and preference measures that have been the primary focus of prior research in this area. The improved sound quality ratings supported by the acoustic metrics in this study are consistent with prior evidence that adapting compression parameters to the signal type using techniques such as scene-aware compression (Kowalewski et al., 2020; May et al., 2018; Overby et al., 2023) or DNN-based source separation (Zhang et al., 2025) can improve listener outcomes over conventional DRC on mixed signals. Critically, the present findings extend this evidence base by demonstrating that source-independent compression yields benefits across different dimensions of sound quality. The use of ground-truth-separated sources, such as those that may be available via remote microphones or streaming, allow these findings to be generalized to current and emerging source separation technologies rather than being tied to a specific source-separation algorithm, motivating further research under more representative real-world conditions.

Supplemental Material

Supplemental Material

Supplemental Material

Supplemental Material