The research drew its stimuli from the CADENZA Challenge, an international machine-learning competition launched to accelerate innovation in hearing-aid music processing. Independent research groups submitted digital signal processing (DSP) models designed to improve music listenability for hearing-impaired listeners, and the Salford team selected three enhancement models, E001, E021, and E022, that had achieved high scores on the Hearing Aid Audio Quality Index (HAAQI), an objective metric that predicts human judgments of music quality for hearing-impaired listeners. E021 remixed music into stems, vocals, drums, bass, and other instruments, using hearing-loss-aware processing to improve perceived quality and clarity. E022 applied a different compensation strategy to the same stems, aiming for better intelligibility and balance but trading off some objective quality. E001, built on the Hybrid Demucs demixing pipeline, served as a baseline-style reference system against which the other entries were compared.

Eleven participants, eight normal-hearing adults and three hearing-aid users, sat in a semi-anechoic chamber wearing Sennheiser SC60 circumaural open-back headphones, the same model used in the CADENZA Challenge to ensure repeatability. Fifteen pop and rock excerpts, each edited to fifteen seconds, were played back as lossless FLAC files under four processing conditions: the three enhancement models plus a personalised baseline tailored to each listener’s audiogram. A standardized configuration file mapped individual audiometric thresholds to gain compensation parameters following the CADENZA CAD1 recipe, so every participant heard versions of each excerpt simulated for their own hearing profile. Presentation order was fully counterbalanced, and participants controlled their own playback level through an in-line volume attenuator to a personally comfortable setting.

After each excerpt, participants rated Basic Audio Quality (BAQ), a subjective measure of perceived technical sound quality following the ITU-R BS.1534 standard, and likeability, a separate rating of enjoyment. The distinction matters because music listening is highly subjective, and technical quality can be completely orthogonal to preference. Throughout the session, a medical-grade Empatica Embrace wrist wearable recorded electrodermal activity (EDA) at 4 Hz and photoplethysmography (PPG) at 64 Hz, capturing autonomic arousal and cardiovascular responses continuously. EDA, sometimes called galvanic skin response, reflects sympathetic nervous system output through sweat gland activity, while PPG tracks heart rate through integrated light sensors that detect blood volume changes in the wrist.

The physiological preprocessing was deliberately tailored to the constraints of short stimuli and wrist-based sensing. EDA signals were decomposed into tonic and phasic components using Continuous Decomposition Analysis in Ledalab, a widely validated tool for music and emotion research, with manual artefact removal for movement-related spikes. Because conventional analysis relies on counting discrete skin conductance response peaks, which short fifteen-second windows at 4 Hz often fail to capture in sufficient frequency, the team instead integrated the cumulative sum of the continuous phasic signal over each trial, tracking cumulative sympathetic output rather than isolated threshold crossings. PPG data underwent beat detection using the MATLAB PPG-Beats Toolbox, followed by a fourth-order Chebyshev high-pass filter with a 0.15 Hz cut-off, chosen over a Butterworth low-pass alternative for its robustness to the movement artefacts that plague wrist sensors.

The results were striking in their consistency. Model E021 achieved the highest BAQ and likeability ratings across both participant groups. Normal-hearing listeners gave E021 a mean BAQ of 5.83 and a mean likeability of 5.83, compared with 4.42 and 3.29 for E022 and 4.05 and 3.91 for E001. Hearing-aid users showed the same pattern, rating E021 at 5.95 for BAQ and 5.24 for likeability, though with greater variability than the normal-hearing group. Across all participants, BAQ and likeability shared a moderate positive association, with Spearman’s rank correlation around 0.48, suggesting that perceived technical quality is a major determinant of enjoyment for listeners with hearing loss, even though other factors clearly contribute as well.

The physiological data told a subtler but converging story. Absolute tonic EDA changes and inter-beat interval (IBI) shifts across stimuli showed no consistent group-level effects, with substantial individual variability. But variability-based measures proved far more sensitive. Phasic EDA variability, the cumulative magnitude of phasic responses per trial, correlated positively with both BAQ (r = 0.31, p = 0.03) and likeability (r = 0.28, p = 0.04). IBI variability, computed as the cumulative sum of trial-wise absolute changes in inter-beat interval, correlated even more strongly with BAQ (r = 0.42, p = 0.02) and likeability (r = 0.39, p = 0.03). Participants who rated E021 highest tended to show elevated phasic EDA variability and IBI variability during that model, suggesting greater autonomic engagement for the processing strategy they subjectively preferred.

The authors are careful about interpretation. Autonomic activation is fundamentally non-valence-specific, meaning elevated phasic EDA and IBI variability do not automatically equate to positive enjoyment. The heightened variability observed during the preferred E021 model could reflect increased aesthetic engagement or emotional intensity, but it could equally signify stimulus novelty, heightened attentional allocation, or even increased cognitive effort required to parse the remixed stems. The pilot sample size, particularly the three hearing-aid users, limits deep isolation of these mechanisms, and the physiological metrics are best interpreted as signs of general autonomic reactivity and processing demands rather than direct proxies for musical pleasure. The team also notes that the CADENZA models were provided as pre-processed files without internal algorithmic details, making E021’s performance a closed black box that prevents isolating fundamental signal-processing principles.

Genre preferences present another confound. The normal-hearing group averaged around 25 years old while the hearing-aid group averaged around 54, and the stimulus pool of contemporary pop and rock excerpts could systematically skew both subjective likeability ratings and autonomic indicators through generational differences in genre familiarity. An older listener experiencing lower baseline preference for heavy rock tracks may display altered phasic EDA variability driven by aesthetic friction rather than the technical processing of the DSP models. Future larger-scale studies, the authors argue, must incorporate systematically balanced genre selection, perhaps classical, jazz, and folk, alongside explicit baseline genre-affinity controls to isolate true model-driven processing benefits.

Despite these caveats, the study positions itself within a growing shift toward listener-aware hearing technologies. Linear mixed-effects regression, which treats participant-level variation as a random effect to handle the nested structure of repeated trials, revealed low to moderate explained variance, with R-squared values typically below 0.15, indicating that a substantial portion of BAQ variance remains unaccounted for by physiological measures alone. A truly effective real-time adaptive system would need a more sophisticated multi-modal approach, integrating a weighted combination of physiological measures with behavioral, acoustic, or cognitive features. But the proof of concept is clear: wearable-derived variability markers can track subjective listener responses even in small cohorts, and future hearing aids or consumer devices might hypothetically adapt DSP parameters in real time to optimise music enjoyment as well as speech clarity. For a population where music enjoyment is a central component of cultural participation and wellbeing, and where non-use rates among fitted hearing-aid users range from roughly 5 to 24 percent, that possibility represents a major unmet need finally coming within reach.

Subject of Research: Hearing loss and music processing

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

Article References: Williams, D. A. H., Waterhouse-Stocks, R., Vos, R., & Cox, T. (2026). Physiological and subjective evaluation of music processing strategies for hearing-impaired listeners: a pilot study. Frontiers in Signal Processing, 6, Article 1876242. https://doi.org/10.3389/frsip.2026.1876242

Image Credits: AI Generated

DOI: 10.3389/frsip.2026.1876242

Keywords: hearing loss, hearing aids, music perception, digital signal processing, electrodermal activity, photoplethysmography, CADENZA challenge, wearable sensors, Physiological, subjective, evaluation, music