Age-related hearing loss affects 466 million people worldwide, a number projected to reach 630 million by 2030, and represents the fourth leading cause of disability, as well as one of the largest modifiable risk factors for dementia [1-5]. A primary challenge for individuals with hearing impairment is understanding speech in noisy environments, which demands a complex interplay between sensory input and cognitive processing [6,7]. While hearing aids restore audibility, they inadequately address the cognitive demands of listening in noise [8], highlighting a critical need for auditory training interventions that enhance selective auditory attention—the ability to focus on target sounds amidst distractors, exemplified by the “cocktail party effect” [9]. Neuroscientific research using functional magnetic resonance imaging and electroencephalogram confirms that specific neural networks dedicated to selective attention are critical for both sound localization and speech segregation, and that targeted training can improve these functions in hearing-impaired individuals [10-14]. Systematic reviews demonstrate that auditory training paradigms can enhance auditory and cognitive skills, including speech perception, working memory, attention, and processing speed [15-17]. However, traditional programs face significant barriers including poor user adherence and limited accessibility [18]. Digital platforms like LACE and Amptify have improved accessibility through at-home delivery [19,20]. However, an opportunity to incorporate spatial audio cues was identified—a dimension absent from current programs but fundamental to inherently 3D, real-world listening. Real-world listening is inherently 3D, with the brain leveraging binaural cues to localize and segregate sound sources, primarily interaural time differences (microsecond delays between ears) and interaural level differences (intensity differences between ears) [21].

Beyond acoustic properties, the listener’s state of engagement critically modulates auditory processing and learning [22]. Research demonstrates that “active listening”—the cognitive engagement with sound—can enhance auditory performance, with studies showing that active listening postures improve 3D sound localization and that embodied training activities can benefit related perceptual skills [23,24]. This engagement often manifests through spontaneous head movements, which dynamically alter interaural time difference and interaural level difference cues to improve localization and resolve spatial ambiguities [25]. Studies in virtual reality (VR) environments suggest that listeners increase head movements when facing challenging listening conditions, reflecting this as an instinctual learning strategy [25,26]. Building on this, research by Valzolgher et al [27] supports the use of active, sensorimotor training to improve spatial auditory performance. While the evidence for immediate transfer to speech-in-noise remains an area of ongoing investigation, a closed-loop audio-motor game requiring continuous sensorimotor prediction and correction has been shown to yield a 25% improvement in speech-in-noise perception [28]. This suggests that such paradigms may achieve generalized learning that transfers beyond trained stimuli, potentially addressing the specificity limits common in traditional training [28]. Therefore, our system design integrates active listening tasks, natural head movements via 6-degree-of-freedom tracking, and goal-directed audio-motor interaction to explore whether these mechanisms can support robust and generalizable auditory learning.

Translating embodied audio-motor training from the laboratory to daily life requires technology that is both environmentally aware and accessible. Modern augmented reality (AR) frameworks are a promising modality for this challenge, having become increasingly accessible through advances in mobile hardware and software [29]. While previous work has explored audio-motor feedback on tablets or in VR [27-30], AR provides the critical advantage of integrating virtual elements into the user’s real environment rather than replacing it. We developed ARIA (Augmented Reality Immersive Auditory training), a handheld mobile intervention that enables in situ training, where the system acoustically adapts to the participant’s actual room geometry. By tracking smartphone position and using earbud-embedded gyroscopes for head rotation, ARIA delivers spatial audio that facilitates natural, embodied training within users’ physical spaces. While serious games show mixed results, well-designed gamification incorporating clear goals, immediate feedback, and structured progression can address adherence challenges [31,32]. Following these principles, our system was refined from an earlier prototype that informed the initial design [33], synthesizing evidence-based audio-motor training with AR’s environmental mapping to create an engaging, acoustically realistic training framework.

The primary aims of this pilot study were to (1) evaluate the feasibility and acceptability of delivering AR-based auditory training to a foundational cohort of middle-aged adults (aged 50‐65 y) to establish platform viability, (2) assess usability through standardized scales and qualitative feedback, and (3) conduct exploratory analyses of speech-in-noise outcomes to inform future efficacy trials.

This study was a single-arm, pre-post–follow-up pilot feasibility trial conducted over a period of 8 weeks, evaluating an at-home training protocol for ARIA. All participants provided written informed consent.

The study recruited middle-aged adults aged 50 to 65 years, an at-risk demographic for age-related sensorineural hearing loss, via university and company email lists supplemented by snowball sampling. Inclusion criteria were as follows: (1) aged 50 to 65 years, (2) self-reported functional hearing not requiring amplification, (3) ability to commit to an 8-week protocol, and (4) Korean fluency. Exclusion criteria were as follows: (1) diagnosed hearing impairment requiring hearing aids, (2) history of ear surgery or chronic ear disease, (3) neurological conditions affecting auditory or cognitive processing, and (4) inability to complete touchscreen-based tasks. Following screening, 11 participants (8 males and 3 females) enrolled. This cohort was specifically selected to evaluate the technical feasibility of the ARIA interface in a relatively high-functioning group before extending the intervention to clinical populations with advanced hearing loss.

This study was reviewed and approved by the Institutional Review Board of the Korea Advanced Institute of Science and Technology (KAISTIRB-2025‐32). All participants provided written informed consent prior to enrollment. Participants were informed of the study’s purpose, procedures, potential risks, and their right to withdraw at any time without penalty. Personal data were anonymized using participant identification codes, and all data were stored on password-protected, encrypted servers accessible only to the research team. Participants received financial compensation (up to ₩300,000, approximately US $216) contingent on protocol adherence.

The ARIA intervention was delivered as a handheld mobile app running on iOS devices. To ensure standardization across participants, all were provided with an iPhone 14 Pro and Apple AirPods Pro 2 earbuds. The app was developed using Unity 2022.3 LTS with ARKit 4.0 for spatial tracking and used the Wwise audio engine for 3D sound rendering [34,35].

The study was conducted over 8 weeks and consisted of a baseline assessment, a 4-week training period, and a follow-up retention assessment. At baseline (week 0), participants attended an in-person session where they received study devices (iPhone 14 Pro and AirPods Pro 2) and completed a 30-minute tutorial with research staff guidance. Baseline speech-in-noise perception was assessed using the Korean Matrix Sentence Test (KMST) at 3 fixed SNR conditions (0 dB, −6 dB, and −9 dB) delivered diotically through Sennheiser HD 650 headphones at each participant’s most comfortable listening level in a sound-treated room [37]. Hearing thresholds were screened using the mobile Mimi Hearing Test app, which has previously been found to produce results comparable to standard audiograms [38]. Additionally, to account for potential baseline differences in auditory processing skills, participants’ musical experience was quantified using the Goldsmiths Musical Sophistication Index [39,40].

During the 4-week training period (weeks 1‐4), participants completed training and evaluation sessions at home on self-selected schedules. The research team sent scheduled reminders about remaining sessions and current progress. Participants returned for posttraining assessment at week 4, where the KMST was readministered using identical procedures, and participants completed the System Usability Scale (SUS) and Player Experience of Need Satisfaction (PENS) questionnaire, followed by a 20- to 30-minute semistructured exit interview [41,42]. A final follow-up at week 8 consisted of a KMST reassessment to evaluate retention of training effects.

Speech-in-noise perception was measured using the KMST, with the percentage of correctly identified words at 3 SNRs (0 dB, −6 dB, and −9 dB) as outcomes. The ARIA system logged trial-by-trial performance, including localization accuracy (identifying sound source position), overall accuracy (both source position and discrimination), response time, and distance error. These metrics were analyzed separately for training sessions and evaluation sessions. User experience was evaluated through the SUS (0‐100 scale) and PENS, which assessed game engagement through Competence, Autonomy, and Presence subscales (1‐7 scale), with the Relatedness subscale omitted as ARIA lacks social features and nonplayable characters. Semistructured exit interviews were conducted to explore participants’ experiences with the AR technology, their perceived benefits and challenges, and their suggestions for improvement (Multimedia Appendix 2). Interview data were analyzed using reflexive thematic analysis. The lead author (PK) performed initial line-by-line inductive coding, followed by a process of collaborative refinement with the research team (SYK, HJL, and IYC). This peer debriefing allowed for a multiperspective reading of the data to establish the credibility and dependability of the thematic structure. All themes remained grounded in specific participant excerpts to maintain confirmability, while transferability was supported through the detailed reporting of our recruitment context and demographics [43,44].

All analyses were conducted using R version 4.3.1 (R Foundation for Statistical Computing) with an α level of .05. KMST improvements were assessed using paired 2-tailed t tests with Bonferroni correction for multiple comparisons across 3 SNR conditions. Effect sizes were calculated using Cohen d. Spearman correlations examined relationships between in-game performance improvements and KMST gains. Linear mixed-effects models assessed session-by-session learning trajectories for in-game metrics during evaluation sessions.

A total of 11 participants (8 males and 3 females) with a mean age of 53.0 (SD 3.0) years completed the 8-week study protocol. Demographic details, including education level, living conditions, usage of smartphones and video games, are summarized in Table 2. All participants who enrolled completed all phases of the study, resulting in a retention rate of 100%. Initial performance on the KMST at baseline showed a mean score of 92.5% (SD 2.3) at 0 dB SNR, 72.9% (SD 8.7) at −6 dB SNR, and 46.6% (SD 15.2) at –9 dB SNR. These 3 conditions represent graded levels of difficulty for the same speech-in-noise perception outcome, with lower SNRs imposing progressively greater perceptual and cognitive demands. Baseline scores confirmed that the more challenging conditions provided sufficient room for improvement while the near-ceiling performance at 0 dB SNR reflected participants’ functional hearing status.

To evaluate the preliminary efficacy of the ARIA training program, exploratory analyses of speech-in-noise perception across all tested conditions were conducted. Paired 2-tailed t tests with Bonferroni correction yielded significant gains from pretraining to posttraining at 0 dB SNR (t₁₀=3.43; P=.02; d=1.20),−6 dB SNR (t₁₀=5.34; P<.001; d=1.14), and −9 dB SNR (t₁₀=4.34; P=.004; d=0.58; Figure 3). Improvements were largest at the more challenging listening conditions, where baseline performance left greater room for change. All effect sizes met or exceeded the Cohen convention for medium effect (d>0.5); however, given the single-arm design and small sample size, these should be interpreted as preliminary efficacy signals. Individual participant analysis revealed consistent improvement patterns: 9 of 11 (81.8%) participants improved at 0 dB SNR, all participants improved at −6 dB SNR, and 10 of 11 (90.9%) participants improved at −9 dB SNR. All participants improved in at least 2 conditions. Follow-up assessment at 8 weeks showed no significant decline from posttraining scores at either −6 dB (P=.05) or −9 dB (P=.77), suggesting short-term retention. However, repeated exposure to the same test and the absence of a control group limit the interpretation.

Participants demonstrated significant learning within the ARIA game during weekly evaluation sessions. Level-matched analysis comparing the first and last evaluation sessions revealed robust improvements in foundational spatial hearing component skills: localization accuracy improved by 12.4% (t₁₀=3.85; P=.003; d=1.16) and distance error reduced by 20.1% (t₁₀=6.50; P<.001; d=1.96; Figure 4). However, overall accuracy—requiring simultaneous correct identification of both sound type and location—showed only a nonsignificant trend (6.9% gain; t₁₀=1.58; P=.145; d=0.48). This dissociation suggests that while component spatial skills improved reliably, integrating these skills under the attentional demands of the full task remained more challenging. Linear mixed-effects models across all 4 evaluation sessions confirmed these patterns, with significant session-by-session improvements for localization (β=.038; P=.02) and distance (β=−.067; P<.001) but not overall accuracy (β=.020; P=.21). Presession self-report measures of sleep, fatigue, and valence did not show significant associations with in-game performance or learning trajectories, with individual response patterns suggesting limited variability across sessions.

To explore potential mechanisms of improvement, correlations between in-game skill development and KMST gains were analyzed. Total training time showed a moderate positive correlation with KMST improvement at −6 dB SNR (ρ=0.547; P=.082), which approached but did not reach statistical significance.

Component skill analysis revealed stronger relationships for localization accuracy. Localization improvements during training correlated significantly with KMST gains at −6 dB SNR (ρ=0.639; P=.03) and −9 dB SNR (ρ=0.612; P=.045). Distance error reduction showed a weaker relationship with KMST gains at −6 dB (ρ=0.547; P=.08). These skill-specific correlations suggest a potential association between trained spatial hearing abilities and speech-in-noise performance, though the direction and underlying mechanisms of this relationship require confirmation in controlled trials.

The acceptability and usability of the AR-based intervention for the target population were assessed, as these factors are critical for determining real-world viability. The ARIA app demonstrated acceptable usability for the target population. The mean SUS score was 70.2 (SD 19.6), just above the industry definition of “average” usability [45]. More notable than the mean, however, was the wide range (47.5‐92.5), indicating substantial individual differences in usability perception that warranted further investigation through qualitative analysis. The PENS scale indicated that the core gameplay loop was engaging, with participants reporting high levels of Competence (mean 5.2, SD 0.89; range=3‐7) but more moderate levels of Autonomy (mean 4.5, SD 1.62; range=2‐5) and Presence (mean 4.8, SD 1.08; range=3‐6; Table 3). This pattern of high competence with moderate autonomy may reflect the structured training paradigm, where progression is prescribed rather than self-directed, and suggests that introducing greater player agency in session structure or difficulty selection could improve engagement.

One limitation of this study is the absence of a control group and the small sample size (N=11), which together preclude definitive causal attribution. Observed KMST improvements may reflect practice effects, maturation, or placebo effects rather than training-specific benefits. While the short-term retention of gains at 8 weeks and the skill-specific correlations suggest a preliminary efficacy signal, only a randomized controlled trial with an active control can establish causality.

The assessment battery’s scope presents another key constraint. Our reliance solely on the KMST prevents conclusions about the cognitive breadth of training effects. Without targeted cognitive assessments, it is not possible to distinguish whether improvements reflect enhanced spatial selective attention—our proposed mechanism—or more general cognitive gains. Furthermore, the training occurred in varied home environments, where factors such as room geometry, lighting, and surface reflectivity were not modeled analytically, potentially influencing both user experience and performance outcomes.

Additionally, the financial compensation provided (up to ₩300,000, approximately US $216) may have influenced the 100% adherence rate and the reported “homework” sentiment. These results may not generalize to nonincentivized, real-world deployment contexts. Finally, sample homogeneity restricts generalizability. Participants were highly educated, technologically proficient, middle-aged adults (mean age 53.0, SD 3.0 y) with functional hearing. Findings may not generalize to older adult populations (aged ≥65 y) who may present with more advanced presbycusis, greater cognitive decline, and different technology adoption patterns. The current sample served as a feasibility testbed, but testing with older adult populations remains an essential next step. Additionally, participants had functional hearing not requiring amplification at baseline, whereas individuals with diagnosed hearing loss may show different training responses, engagement patterns, and usability challenges.

This study demonstrated the feasibility of delivering a gamified, AR-based audio-motor training paradigm to middle-aged adults in their home environments. Beyond feasibility, several findings carry broader implications. The dissociation between successful core engagement and peripheral usability friction suggests that, for AR-delivered health interventions more generally, streamlining environmental setup is as critical as refining the intervention content itself. The exploratory correlations between spatial skill development and speech-in-noise gains provide a preliminary rationale for investigating audio-motor training as a mechanism for auditory intervention and potential rehabilitation. This hypothesis warrants further testing in controlled trials involving diverse clinical populations. Finally, the observed potential for performance-based training to promote hearing health awareness suggests a dual function for gamified auditory interventions—as both training tools and experiential screening platforms. These findings provide a foundation for refining the ARIA platform and designing a larger randomized controlled trial.