Physiological and subjective evaluation of music processing strategies for hearing-impaired listeners: a pilot study

1 Introduction

Hearing loss affects a growing proportion of the global population and is expected to become a major public health challenge over the coming decades. The World Health Organization estimates that by 2050, more than 2.5 billion people will be living with some degree of hearing loss (World Health Organization, 2024). In the United Kingdom alone, approximately one in six adults are affected (Royal National Institute for Deaf People, 2024), with prevalence rising sharply as populations age and increased exposure to noise. Beyond difficulties in understanding speech, hearing loss has profound consequences for music perception and enjoyment - a central component of cultural participation and wellbeing.

Music contributes to emotional regulation, social cohesion, and quality of life (Coffman, 2022; MacDonald et al., 2012). Engaging with music activates reward and affective circuits (Salimpoor et al., 2009) and provides both direct benefits (e.g., emotional pleasure) and indirect benefits (e.g., strengthening social bonds and combating loneliness). However, people with hearing loss frequently report that music becomes less enjoyable, with melodic structure, instrumental balance, and lyrics often becoming harder to follow (Madsen and Moore, 2014). As high-frequency information is degraded, music can sound duller and less engaging, and lyrics can become less intelligible. In turn, this can reduce motivation to participate in musical activities and lead to wider impacts on mental health and social wellbeing.

1.1 Limitations of current hearing aid technologies

The challenge of music appreciation is particularly acute for individuals using hearing devices, including hearing aids - the focus of this study - and cochlear implants. These devices are primarily optimised for speech intelligibility, and their performance for music reproduction remains limited (Greasley et al., 2020). Processing strategies such as multi-band compression and noise reduction, designed to enhance speech clarity, often distort the dynamic range, harmonic structure, and timbre of music. Consequently, some hearing-aid users find listening to music unsatisfying and may even avoid music entirely, or choose to remove their hearing aids during music listening (thus losing the benefit of amplification).

Despite advances, a significant proportion of recipients fail to wear their devices; systematic reviews report that non-use rates among fitted users range from approximately 5%–24% across different studies (McCormack and Fortnum, 2013). The standard speech-optimised processing in most modern hearing aids can introduce undesirable artefacts, such as modulation, spectral smearing, and high levels of distortion, which significantly degrade the perceived quality of music (Madsen and Moore, 2014). Improving the musical experience of hearing-aid users is therefore a critical goal for increasing device adoption and improving quality of life. To overcome the limitations of relying solely on self-report, researchers increasingly combine subjective quality ratings with physiological measures to capture listening effort and engagement. Electrodermal Activity (EDA) and Photoplethysmography (PPG) are reliable measures of autonomic arousal and emotional state, which have been used to quantify the effort required to process speech through hearing aids and to track the emotional experience of music listening (Salimpoor et al., 2009). In the context of DSP to assist hearing-impairment, a model that truly improves music quality should ideally lead not only to higher subjective ratings but also to reduced listening effort (e.g., lower tonic EDA) or a stronger, positive emotional response (e.g., higher phasic EDA). Our study is among the first to apply this objective physiological methodology to the evaluation of the latest generation of machine learning-optimised DSP models from the CADENZA challenge.

1.2 Advances in DSP and the CADENZA challenge

Recent advances in machine learning and digital signal processing (DSP) have driven interest in personalising audio processing such that it might enhance music listening for people with hearing loss. The CADENZA Challenge was launched to accelerate innovation in this space, focusing on how machine learning DSP can improve hearing aid performance for music perception. Independent research groups contributed DSP models designed to improve music listenability for hearing-impaired listeners. These models were evaluated against shared datasets and listener tests, making it possible to compare enhancements in a standardised way. The primary objective was to maximize the Hearing Aid Audio Quality Index (HAAQI), an objective measure designed to predict human judgments of music quality for hearing-impaired listeners. The DSP models evaluated in this study were selected because they represent submissions that successfully leveraged metrics like HAAQI to achieve high predicted music quality and intelligibility relative to the unprocessed baseline.

The present study leverages stimuli provided by the CADENZA Challenge, including three enhancement models (E001, E021, E022) and renderings generated via the CAD1 ² workflow, which involves generating the processed stimulus based on the listener’s individual audiogram and prescribed hearing aid gain (Roa Dabike, 2023). E021 remixed music stems (vocals, drums, bass, other) using hearing-loss–aware processing to improve overall perceived quality and clarity for listeners with audiometric loss. E022 applied a different hearing-loss–compensation strategy to remix the same stems, aiming for better intelligibility and balance but trading off some objective quality relative to E021. E001 is a baseline-style internal system used in CAD1 analyses, built on the Hybrid Demucs demixing pipeline, that provides a reference remix HAAQI score against which other entries were compared.

1.3 Physiological measurement techniques

While subjective ratings provide crucial insights into perceived quality and enjoyment, advances in wearable sensing now allow researchers to probe listener engagement using real-time physiological measurement. Physiological responses are closely linked to the hedonic and affective dimensions of music listening (Salimpoor et al., 2009). We used the Empatica Embrace which is a medical-grade device combining compact sensors for electrodermal activity (EDA) and photoplethysmography (PPG), enabling continuous measurement of arousal and cardiovascular responses during music listening (Almarshad et al., 2022; Böttcher, 2022). Peaks in EDA and heart-rate variability (HRV) often correspond to moments of pleasure and emotional intensity (Benedek and Kaernbach, 2011; Tervaniemi et al., 2021). Integrating these measures with subjective quality ratings can therefore provide a richer, multi-dimensional view of listener engagement. If successful, in the future, this type of research may offerpotential pathways toward real-time adaptive personalisation for hearing technologies.

1.4 Aims of this study

This study evaluates how listeners with and without hearing loss respond to music processed using DSP models from the CADENZA Challenge. We combine subjective ratings of Basic Audio Quality and likeability with wearable-derived physiology to investigate:

Do physiological responses from wearables reflect subjective preferences?

Do different DSP models change the perceived quality and enjoyment of music in self-reported BAQ and likeability?

Which combinations of subjective–physiological markers could inform future listener-aware, adaptive frameworks for hearing aid DSP?

By integrating subjective appraisal with autonomic measures, this pilot experiment explores a potential paradigm for next-generation hearing technologies that could eventually move beyond static, audiogram-based tuning toward dynamic, personalised optimisation of music listening experiences. While tentative, these preliminary findings suggest that variability-based physiological markers (EDA, PPG) warrant further investigation as candidates for further exploration alongside adaptive personalisation in hearing technologies. We tentatively position this work within the growing shift toward listener-aware solutions, where future hearing aids and consumer devices might hypothetically adapt to optimise music enjoyment (as well as speech perception).

2 Methods

Initially, fourteen adults took part in the study, divided into two groups:

• Normal-hearing group (NH): N = 9 tested (see definition below), of whom 1 was excluded due to device recording failure, giving a final NH used in our analysis N = 8 (mean age = 25.0 years, SD = 9.65, 87.5% male).
• Hearing-aid user group (HA): N = 5 tested; 1 excluded due to device failure and 1 excluded because personalised stimuli for their audiometric profile were not available, giving a final HA in our analysis N = 3 (mean age = 54.0 years, SD = 15.95, 66% male).

Normal-hearing participants were classified according to World Health Organization (WHO) and Global Burden of Disease (GBD) criteria, based on audiograms obtained using a Kamplex r27a audiometer (calibrated 21 December 2023). For hearing-aid users, audiograms were obtained via the CADENZA Challenge database. For the NH group, audiograms were typically not available prior to participation and were measured on site. Thresholds showed <20 dB HL across standard frequencies (250–8,000 Hz). The HA group presented with mild to moderate sensorineural hearing loss, consistent with typical hearing-aid candidacy profiles reported in the CADENZA listener database.

All participants gave informed consent. The protocol received approval from the University of ******* Research Ethics Committee, Project ID 480, Review reference 20240480249.

2.1 Apparatus and test environment

Testing was conducted in a semi-anechoic chamber to minimise outside acoustic interference. Audio playback was via Sennheiser SC60 circumaural headphones, which included an in-line volume attenuator and an integrated USB digital-to-analogue converter for playback–ensuring repeatability with the CADENZA Challenge which used the same model. The wrist-wearable device collected electrodermal activity (EDA) and photoplethysmography (PPG) data continuously. Stimulus playback was controlled via a desktop computer running MATLAB-based scripts. Participants had control over their own playback level using the in-line attenuator on the headphone connection and as part of the familiarisation exercise were instructed to adjust the playback level to a level they found personally comfortable.

2.2 Stimuli and DSP models

Fifteen musical excerpts were supplied by the CADENZA Challenge, each processed under the following digital signal processing (DSP) conditions:

• Baseline: Personalised for each listener using their audiogram. A listeners. json configuration file - configuration file—a standardized data format used in the CADENZA workflow to map individual audiometric thresholds to gain compensation parameters - was generated per participant following the Cadenza CAD1 recipe to produce individualised baseline stimuli.
• E001, E021, E022: Three proprietary enhancement models provided by Cadenza, each applying different DSP strategies designed to improve music listenability for people with hearing loss. Full details of the processing involved in each can be seen in (Roa-Dabike et al., 2025).

The musical excerpts were played back as. flac lossless audio files - APlaceForUs BigDummyShake CurAnLongAgSeol Directions IllFate Mallory OhNo SchoolboyFascination Semantics SingWithMe TheEnglishActor Tiring TooMuch WeFeelAlright. They were all edited to a 15s excerpt, and include a variety of “pop” and “rock” genres.

The baseline condition was used as a reference for physiological normalization; however, to minimize participant fatigue in this pilot study, subjective ratings for the baseline were only collected for a subset of participants and thus excluded from group-level statistical comparisons The presentation order of excerpts and models was fully counterbalanced to minimise sequence effects.

2.3 Procedure

Participants were seated comfortably within the semi-anechoic chamber, fitted with the measurement device and headphones, and given a short familiarisation period, during which the wearable device also has time to acclimatise to the wearer. EDA was sampled at 4 Hz and PPG at 64 Hz using the Empatica Embrace device. For each trial, a musical excerpt was presented in one of the four DSP model conditions. Personalised versions of each processed excerpt were generated using listeners’ audiograms, enabling a realistic simulation of hearing-aid fitting and ensuring that comparisons reflected genuine model-level differences rather than mismatched amplification. After each excerpt, participants rated Basic Audio Quality (BAQ): Subjective rating of the perceived technical quality of the sound, following ITU-R BS.1534, (International Telecommunication Union, 2015), which asks listeners to rate the technical quality of audio reproduction on a standardised scale, and Likeability, as a subjective rating of enjoyment of the excerpt (music listening is highly subjective, and Quality may be completely orthogonal to Likeability). At the end of the session, participants were debriefed and given the opportunity to provide open-ended feedback.

2.4 Physiological data acquisition and preprocessing

The wearable device provides two types of physiological activity data: Electrodermal Activity (sometimes referred to as Galvanic Skin Response, or sweat), and Photoplethysmography (heart-rate). EDA and PPG are validated indicators of autonomic arousal linked to music emotion, reward and engagement. Previous research has linked music-induced pleasure and immersion to changes in tonic and phasic EDA and to variations in inter-beat interval (IBI), e.g., (Salimpoor et al., 2009). These measures allow continuous, unobtrusive monitoring during music listening. A baseline period allows for tonic EDA and resting IBI in each measure. We used randomised trial order controls for fatigue and order effects, with overall trial lengths fixed according to the 15s stimulus edited durations. These two metrics were acquired and pre-processed before analysis as follows.

2.5 Electrodermal activity (EDA)

EDA was recorded continuously during each trial. Preprocessing included: Decomposition of tonic and phasic components using Continuous Decomposition Analysis (CDA) in Ledalab (Benedek and Kaernbach, 2011). Further artefact removal via manual inspection for movement-related spikes, where filtering alone was insufficient, following (Park et al., 2022).

2.6 Photoplethysmography (PPG)

PPG signals were recorded using integrated light sensors. Processing began with beat detection, using the MATLAB PPG Beats Toolbox (Charlton et al., 2022). This was followed by applying a 4th-order Chebyshev high-pass filter (cut-off = 0.15 Hz), following (Allen and Murray, 2004; Joseph et al., 2014), to reduce low-frequency movement artefacts. This was used instead of the alternative, 3rd-order Butterworth low-pass filter, which is less robust to the type of movement artefacts which wrist-based sensors are susceptible to when a participant moves their hands.

2.7 Analysis metrics

EDA data were decomposed into tonic and phasic components using Continuous Decomposition Analysis in Ledalab, which is widely used and validated for music and emotion research. PPG data were processed using the PPG-Beats toolbox, selected for its strong benchmarking performance against other open-source algorithms. Motion artefacts were reduced using established high-pass filtering approaches for PPG waveform correction. These choices ensured reproducible and physiologically interpretable measures across participants.

2.8 EDA metrics

ΔTonic EDA describes the change in tonic electrodermal activity from excerpt start to end. Phasic EDA variability describes the magnitude of the cumulative sum of phasic responses per trial, reflecting within-trial autonomic arousal. Correlations and regressions were used to explore relationships between ΔTonic EDA, phasic variability, BAQ ratings, and likeability ratings.

2.9 PPG/IBI metrics

ΔIBI describes the change in inter-beat interval from excerpt start to end (ms). IBI variability was computed as the cumulative sum magnitude of trial-wise variability, mirroring phasic EDA variability. Baseline-corrected maximum IBI describes the maximum IBI during excerpt minus mean baseline IBI (exploratory only).

2.9.1 Physiological data methods

We use three targeted metrics: phasic EDA variability, IBI variability, and baseline-corrected maximum IBI.

Conventional electrodermal analysis relies heavily on discrete skin conductance response (SCR) peak counting. However, short-duration stimulus windows (15 s) paired with the low sampling rate (4 Hz) of wearable sensors often fail to capture a sufficient frequency of discrete SCR peaks. We therefore integrated the cumulative sum of the continuous decomposed phasic signal over each trial. This acts as an uninterrupted integration of sympathetic arousal, tracking cumulative neural output rather than isolated threshold crossings.

Traditional Heart Rate Variability (HRV) metrics, such as RMSSD, SDNN, or spectral powers (LF/HF), require stable, continuous data streams typically lasting at least 2–5 min to be physiologically valid. Computing standard frequency or time-domain HRV over brief 15-s windows yields highly unstable estimates that are easily biased by transient respiratory sinus arrhythmia cycle gaps. To evaluate rapid, trial-by-trial autonomic changes during short musical excerpts, we calculated the cumulative sum of trial-wise absolute changes in the inter-beat interval (IBI). This custom windowed variability serves as a directional index of transient vagal and sympathetic shifts specifically tailored to short-duration music processing tasks.

Baseline-corrected maximum IBI was used to reflect transient decelerations in heart rate which have been linked in psychological literature to deep listening states. Correcting against a pre-stimulus baseline isolates this reaction from slower drifts in heart rate.

2.10 Statistical analyses

Analyses were performed in Python and MATLAB. Three approaches were used:

• Descriptive summaries: Mean ± SD for likeability, BAQ, ΔTonic EDA, phasic variability, and IBI variability per model and group.
• Correlation analyses: Pearson correlations between BAQ, likeability, ΔTonic EDA, phasic variability, and PPG metrics.
• Exploratory mixed-effects regressions: DV: Likeability or BAQ. Mixed-effects regression allows us to model repeated measurements from the same listeners while accounting for individual differences.

Because each participant provided repeated ratings and physiological responses across multiple DSP models, observations were not independent. We therefore used linear mixed-effects regression, which is standard for repeated-measures perceptual and physiological data. Mixed-effects models allow participant-level variation to be treated as a random effect, providing unbiased estimates of fixed effects such as DSP model and hearing group while retaining all available data.

2.11 Linear mixed-effects model specification

To account for the nested structure of our data (multiple trials nested within individual participants), we constructed two separate linear mixed-effects models (LME). Model selection was performed systematically, starting from a baseline null model (random intercept only) and incrementally adding fixed effects (DSP Model and Participant Group), comparing models via Likelihood Ratio Tests (LRT). The final models for both subjective outcomes (BAQ and Likeabilit) are specified as:where Y _ij represents the subjective rating (BAQ or Likeability) for trial i from participant j. β₀ represents the fixed intercept, β₁ and β₂ are the fixed effect coefficients for the DSP Model and Participant Group, respectively. U _0j represents the random intercept for participant j (accounting for baseline individual variation), and e _ij represents the residual error. Model residual normality and homoscedasticity were verified visually using Q–Q plots and residual-versus-fitted plots to ensure robustness.

Y i j = β 0 + β 1 D S P M o d e l i j + β 2 P a r t i c i p a n t G r o u p j + u 0 j + e i j

3 Results

3.1 Data overview

This resulted in 11 valid datasets (NH = 8, HA = 3). Analyses below report the full tested sample size where appropriate but note valid participant counts for specific analyses. The subjective ratings (BAQ, Likeability) and physiological metrics were analyzed using Linear Mixed-Effects Models (LMEMs). The fixed effects (DSP Model and Participant Group) were chosen using a systematic model selection process. We started with a null model and then added terms sequentially, confirming that the inclusion of each new fixed effect term resulted in a significantly better fit to the data using the Likelihood Ratio Test.

3.2 Subjective ratings

Descriptive statistics of subjective ratings of BAQ and Likeability are presented in this section.

3.2.1 Basic Audio Quality (BAQ) and likeability

Participants rated both Basic Audio Quality (BAQ) and music likeability after each trial as shown in Figure 1.

E021 consistently produced the highest BAQ and likeability ratings across both groups. HA participants showed greater variability than NH participants, particularly for E001 and E022. Across all participants, BAQ was positively associated with (ρ ≈ 0.48). While the Pearson correlation suggests a strong linear trend, the lower Spearman rank correlation indicates that the relationship is likely influenced by the skewed distribution of ratings or influential data points. We therefore rely on the more conservative Spearman estimate for interpretation. The Baseline condition was included as a reference in the model set but was not rated independently by all participants, so it is not included in the group-level comparison. EDA and PPG often deviate from normal distributions, so non-parametric Spearman rank correlations were computed in addition to Pearson correlations. Spearman analyses revealed a small but significant association between BAQ and mean tonic EDA (ρ = 0.12, p = 0.003), no significant association between Likeability and tonic EDA (ρ = 0.07, p = 0.10), and a moderate positive correlation between BAQ and Likeability (ρ = 0.46, p < 0.001). The discrepancy between these coefficients suggests that distributional assumptions (e.g., normality) may not be fully met, and the Spearman rank correlation provides a more accurate reflection of the monotonic relationship between these variables Table 1 shows a comparison between BAQ and likeability.

TABLE 1. Comparison of BAQ and likeability between groups. N denotes total observations across participants and trials.

3.3 Electrodermal activity (EDA)

EDA analyses were based on ΔTonic EDA and phasic variability metrics. There were no consistent group-level differences in ΔTonic EDA between DSP models. However, individual responses varied substantially, with some participants showing tonic decreases during preferred DSP models, while others showed increases. Phasic variability, calculated as the magnitude of the cumulative sum of phasic activity per trial, tended to be higher for E021 than for E001 or E022 in both groups. This suggests greater autonomic engagement for the DSP model most positively rated subjectively.

Table 2 shows r and p values correlating EDA and the subjective measures. Whilst the tonic EDA changes were inconsistent, phasic variability tracked subjective quality and enjoyment moderately Figure 2 shows phasic EDA and BAQ.

TABLE 2. Relationships between EDA and subjective measures.

3.4 Photoplethysmography (PPG)

PPG analyses explored three candidate inter-beat interval (IBI) metrics:

• ΔIBI, IBI variability, and baseline-corrected maximum IBI.
• ΔIBI (Change Over Stimulus) showed no consistent directional effect across DSP models at the group level. However, within-participant analyses revealed some alignment with subjective preferences: participants who rated E021 highest tended to show small decreases in IBI (increased heart rate). IBI variability -measured analogously to phasic EDA variability -was elevated during E021 for a subset of participants. This suggests greater physiological engagement for the DSP model most strongly associated with positive BAQ and likeability ratings. The Baseline-Corrected Maximum IBI–an exploratory measure (max IBI during trial −mean baseline IBI) showed inconsistent results and was thus excluded from inferential analyses.

Of the PPG markers, this suggests that IBI variabilitycovaries most closely with both BAQ and likeability, particularly for E021 as shown in Figure 3.

Across EDA and PPG measures E021 elicited the strongest combination of subjective and physiological engagement. While tonic EDA and ΔIBI alone showed inconsistent effects, phasic EDA variability and IBI variability demonstrated moderate correlations with both BAQ and likeability. E021 outperformed other DSP models on both Basic Audio Quality and likeability ratings within this sample. Across trials, BAQ and likeability ratings shared a moderate positive association (Spearman’s ρ ∼ 0.48, p < 0.001). Phasic EDA variability and IBI variability were moderately correlated with both metrics, suggesting increased engagement for preferred models. Absolute tonic EDA and ΔIBI metrics were less sensitive markers of music-related response. These findings suggest variability-based measures may better capture music-related arousal than absolute level changes Table 3 shows relationships between PPG and subjective measures.

TABLE 3. Relationships between PPG and subjective measures. The Baseline condition served as a physiological reference; subjective ratings were collected for only a subset of trials to reduce participant fatigue and were thus excluded from group-level comparisons.

Table 4 shows a summary of subjective (BAQ, Likeability) and physiological (ΔTonic EDA, Phasic EDA Variability, IBI Variability) metrics for each DSP model. Values represent means ± standard error where applicable -EDA is measured in microSiemens and reflects the mean change relative to the pre-stimulus baseline Table 5 compares subjective responses between both groups.

TABLE 4. Summary statistics for DSP models.

TABLE 5. Comparison of subjective scores between NH and HA groups.

Mean scores and Standard Error of the Mean (SEM) for Basic Audio Quality (BAQ) and Likeability ratings, contrasting Normal-Hearing Participants (NH) and Hearing-Aid (HA) Users across the three Digital Signal Processing (DSP) Models (E001, E021, E022). Scores ranged from 1 (Worst/Least Likeable) to 10 (Best/Most Likeable). Model E021 consistently received the highest mean ratings for both metrics and groups, while the HA User group showed greater score variability (larger SEM) overall.

3.4.1 Participant-level distributions

Individual participant trajectories (Figure 4) and distribution plots featuring boxplots overlayed with individual data points (Figure 5).

4 Discussion

Across both groups, model E021 consistently achieved the highest ratings for BAQ and likeability, indicating that this DSP approach yielded music reproduction perceived as both technically superior and more enjoyable. The correlation between BAQ and likeability. Suggests that perceived technical quality is a major determinant of enjoyment for listeners with hearing loss, although the moderate strength of the rank-based correlation implies that other factors also contribute to overall enjoyment. While likeability is technically an assessment of preference, its strong association with BAQ and its use as a measure of The moderate association between BAQ and likeability (Spearman’s (ρ ≈ 0.48) indicates that perceived technical quality is a contributing factor to enjoyment for this cohort of listeners with hearing loss. While likeability is an assessment of preference, its association with BAQ in this exploratory dataset allows us to interpret it as a useful proxy for tracking relative aesthetic satisfaction as a reliable proxy for music enjoyment in this context. In other words, when a model improved clarity and richness, participants not only rated the sound quality higher but also experienced greater aesthetic satisfaction. This aligns with prior findings from the Hearing Aids for Music project (Greasley et al., 2020) and other studies on hearing technology (Madsen and Moore, 2014; McCormack and Fortnum, 2013), which have highlighted the importance of developing DSP strategies optimised for music quality, not just speech intelligibility.

A challenge of this study is that the performance of model E021 is a closed, ‘black-box’ to us. Because the internal DSP architecture, multi-band parameters, and exact allocation of demixing thresholds for the CADENZA models (E001, E021, and E022) were provided as pre-processed files without open algorithmic structures, notably around the acoustic features, such as specific frequency bands, compression rates, or spatial reconstruction choices. While our findings demonstrate a preference for E021’s overall output, they cannot be used to isolate fundamental signal-processing principles. Consequently, future iterations of this psychophysiological evaluation workflow should prioritize using open-box algorithms where parameters can be systematically isolated and manipulated, enabling researchers to correlate specific acoustic treatments with autonomic and subjective responses directly.

It is important to note that absolute likeability ratings could be influenced by individual music genre preferences, particularly among the older participant group who may have less affinity for certain musical styles (e.g., heavy rock) included in the stimuli set. Crucially, this variance in stylistic alignment represents a significant confounding variable given the age difference between our two cohorts (Normal-Hearing age ∼25, HA age ∼54). Because the CADENZA stimulus pool includes contemporary pop and rock excerpts, generational differences in genre familiarity and aesthetic baseline preference could systematically skew both subjective likeability ratings and autonomic indicators. For instance, an older listener experiencing lower baseline preference or higher cognitive fatigue when exposed to heavy rock tracks may display altered phasic EDA variability or compressed cardiac patterns driven by aesthetic friction rather than the technical processing of the DSP models. Given our pilot constraints, we cannot mathematically decouple age-skewed genre preferences from audiometric status, and future larger-scale studies must incorporate a systematically balanced genre selection (e.g., classical, jazz, and folk) alongside explicit baseline genre-affinity controls to isolate true model-driven processing benefits. The physiological measurements provided an aligned pattern of data tracking alongside subjective reports during trials with E021 based on a within-subject comparison. Importantly, the type of metric mattered: Tonic EDA changes (absolute shifts in skin conductance) and ΔIBI (heart period change across trials) showed high inter-individual variability and no consistent group-level effects. Variability-based measures were more sensitive: Phasic EDA variability correlated positively with both BAQ (r ≈ 0.31) and likeability (r ≈ 0.28) while IBI variability correlated moderately with BAQ (r ≈ 0.42) and likeability (r ≈ 0.39). These findings alignwith research linking autonomic variability to pleasurable music listening (Benedek and Kaernbach, 2011; de Fleurian and Pearce, 2021), however a crucial caveat of psychophysiological markers is that autonomic activation is fundamentally non-valence-specific. Elevated phasic EDA and IBI variability do not automatically equate to positive enjoyment or hedonic pleasure. Instead, these responses reflect a generalized state of sympathetic nervous system arousal, which can be driven by a multitude of competing psychological constructs. For example, the heightened variability observed during the preferred E021 model could reflect increased aesthetic engagement or emotional intensity, but it could equally signify a response to stimulus novelty, heightened attentional allocation, or even increased cognitive effort required to parse the remixed stems. Given that our pilot sample size limits deep isolation of these mechanisms, these physiological metrics are best interpreted as signs of general autonomic reactivity and processing demands rather than direct, pure proxies for musical pleasure.

While E021 performed best overall, physiological responses showed substantial individual variability, some participants demonstrated increased phasic EDA and IBI variability during their preferred DSP models, indicating stronger physiological engagement or positive affective arousal., whereas others showed minimal changes. Hearing-aid users exhibited greater spread in both BAQ and likeability ratings compared to normal-hearing participants. This heterogeneity suggests that a one-size-fits-all DSP strategy may be insufficient. Instead, personalised solutions that adapt to listener profiles - incorporating audiometric data, subjective preferences, and possibly real-time physiological feedback -may better support accessible and engaging music experiences. Our preliminary results suggest that future frameworks exploring continuous engagement measures (e.g., phasic EDA variability, IBI variability) could examine whether it is feasible for hearing aids or consumer devices to dynamically adjust DSP parameters in real time. Our regression analysis revealed a low to moderate explained variance, with R2 values typically ranging below 0.15 for most individual regressions. This suggests that while a significant relationship exists, a substantial portion of the variance in BAQ remains unaccounted for by these measures alone. To develop a truly effective, real-time adaptive system, it would be necessary to use a more sophisticated multi-modal approach, integrating a weighted combination of physiological measures with established behavioral, acoustic, or cognitive features This would represent a step toward listener-aware technologies capable of optimising not only for speech intelligibility, but especially for music enjoyment -a major unmet need (McCormack and Fortnum, 2013).

5 Limitations

“A primary constraint of this study is its exploratory sample size, particularly regarding the valid hearing-aid user cohort (N = 3). Consequently, these inferential trends must be interpreted strictly as an initial proof-of-concept demonstrating the technical feasibility of tracking continuous wearable-derived autonomic indicators during auditory processing tasks, rather than a generalizable clinical assessment of the wider hearing-impaired population. Future high-powered validation trials are required to confirm these tentative psychophysiological relationships across broader demographic and audiometric strata. Even including the non-HA participants, pilot dataset issmall (final N = 11 valid participants), limiting statistical power. Specifically, the hearing-aid group size should be interpreted cautiously - wearable-derived signals are prone to motion artefacts and sampling gaps which reduced the useable dataset in our case. The laboratory context may also not generalize to everyday listening contexts, as the semi-anechoic chamber is very well insulated to exterior sound. We are unable to comment or investigate further on the enhancement models (E001, E021, E022) which were provided as pre-processed files without internal algorithmic details, limiting our ability to interpret why E021 performed best, so this should be considered a ‘black-box’ when interpreting our findings. Future work should therefore include larger and more diverse samples, consider moving from the lab to real-world listening contexts (including longitudinal, at home-testing), and compare physiological responses across DSP strategies with deeper access to model internals to facilitate such exploration.

6 Conclusion

This experiment evaluated how listeners with and without hearing loss responded to music processed under four Cadenza Challenge DSP models, combining subjective ratings of Basic Audio Quality (BAQ) and likeability with physiological measures from a wrist-based wearable device. Our aim was to understand how DSP processing strategies affect both perceived sound quality and engagement and to explore whether wearable-derived autonomic indices can complement subjective assessment in guiding future adaptive hearing-aid designs. We found that DSP processing strategies can significantly influence both perceived audio quality and physiological engagement when listening to music. Among the models tested, E021 consistently outperformed others, producing the highest ratings of BAQ and likeability. Physiological findings suggest that variability-based autonomic measures - particularly phasic EDA variability and IBI variability - provide useful, converging evidence of listener engagement, even where absolute levels (e.g., tonic EDA) show inconsistent effects. While the present study is limited by its small sample size, these pilot findings offer an initial proof-of-concept that variability-based markers represent a speculative pathway for future integration into adaptive hearing technologies. Considerably more research is required to determine whether such markers can be robustly utilized to tailor processing dynamically to optimise music enjoyment as well as speech clarity. These findings highlight a promising pathway toward accessible, personalised music listening for people with hearing loss.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frsip.2026.1876242/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Farook Sattar, University of Victoria, Canada

Reviewed by: Erica Volta, Italian National Research Council, Italy

Jyrki Alakuijala, Google Zurich, Switzerland

Footnotes

https://cadenzachallenge.org/

CAD1 was the first Cadenza signal-processing challenge, focusing on improving music listening for people with hearing loss in two everyday scenarios.