The research team, led by Shujiro Minami of the National Hospital Organization Tokyo Medical Center and collaborators at Keio University and eleven other Japanese institutions, set out to address a glaring gap in digital audiology. Most validated online hearing tests have been built for English speakers, leaving the estimated 125 million Japanese speakers with few culturally and linguistically appropriate tools. The team’s platform runs entirely in a web browser on any smartphone, tablet, or computer, requiring no special equipment and no visit to a clinic.

Between April 2023 and March 2026, the platform recorded 8,022 unique users and 86,959 test executions across four full-length modules: a music perception task based on the Montreal Battery of Evaluation of Musical Abilities, speech perception in quiet using standardized Japanese monosyllables, speech-in-noise tests at three signal-to-noise ratios, and a consonant identification task. Completion rates ranged from 43.7 percent for the longest module to 85.8 percent for the shortest, suggesting that test duration rather than difficulty drives whether users finish.

The score distributions revealed a striking pattern. Speech perception in quiet showed pronounced upper-end clustering, with 41 percent of completed attempts and nearly 58 percent of users scoring at or above 90 percent correct. This ceiling effect was expected, as monosyllabic word recognition in quiet is known to saturate even in cochlear implant recipients. In contrast, music perception scores were broadly distributed, with only about 21 percent reaching the 90 percent threshold.

The speech-in-noise modules behaved exactly as the scientific literature would predict. Mean accuracy fell progressively from 72 percent at a favorable +10 dB signal-to-noise ratio to 57 percent at the most challenging 0 dB condition, where no attempt reached 90 percent. These patterns persisted when the analysis was restricted to one attempt per user, confirming that they were not artifacts of a small number of enthusiastic repeat testers.

Music perception correlated only modestly with all speech-related outcomes, with Spearman correlation coefficients between 0.31 and 0.39, meaning it shared roughly 9 to 15 percent of variance with speech measures. This finding suggests that musical and speech listening abilities, while related, tap into substantially different auditory skills and cannot substitute for one another in assessment.

When users repeated the same module, the intraclass correlation coefficients ranged from 0.58 to 0.85, indicating moderate to good reliability. However, the smallest detectable change between two administrations ranged from 17.6 to 35.8 percentage points, far exceeding the differences that would typically matter after a hearing aid adjustment or cochlear implant remapping. A second attempt was consistently slightly higher than the first, consistent with a small practice effect.

The most clinically consequential part of the study compared the platform’s 20-item online speech check against the conventional 20-item 67-S monosyllabic word recognition test administered in a calibrated audiology booth. Forty-one participants completed both tests within days to weeks of each other. Agreement was moderate, with a Pearson correlation of 0.65 and an intraclass correlation coefficient of 0.62, but the limits of agreement spanned an enormous range from 34.7 percentage points below to 22.2 percentage points above the clinical score.

The online scores showed a small negative bias of 6.2 percentage points, which the authors attribute to a structural feature of the comparison rather than a defect of the test. The clinical reference represents a best-performance measure obtained under optimized conditions with professional calibration, while the online test was self-administered on personal hardware in an environment of the participant’s choosing. Only 63 percent of participants scored within 10 percentage points across the two methods.

The authors are explicit about what their data do and do not establish. No pure-tone audiometry reference standard was available, no referral threshold was defined, and neither sensitivity nor specificity for hearing-loss screening was estimated. The platform is best characterized as a scalable instrument for describing listening performance at the group level and engaging users with hearing self-assessment, not as a diagnostic or monitoring tool. Reducing measurement error through longer tests, standardized transducers, or averaged repeated administrations will be essential before individual-level clinical applications become viable.

The distinction between correlation and agreement is central to understanding why the study’s moderate correlation coefficients coexist with limits of agreement that span nearly fifty-seven percentage points. A correlation coefficient describes how closely two sets of measurements track each other across a group of people: if someone who scores high on the clinical test also tends to score high online, the correlation will be strong even if the two values differ substantially for each individual. Agreement, by contrast, asks whether the two measurements can be used interchangeably for a single person. The Bland-Altman framework, which the authors applied to their paired data, makes this distinction explicit by plotting the difference between methods against their average. The resulting limits of agreement, from minus 34.7 to plus 22.2 percentage points, mean that for roughly 95 percent of individuals the online score could fall anywhere within that band relative to the booth-based score. A patient whose word recognition drops from 80 to 55 percent on the online check might, in truth, have changed very little, while another showing an apparently stable score could have deteriorated meaningfully. This is why the authors insist that the platform cannot yet support between-visit monitoring of hearing-aid fittings or cochlear-implant mappings, despite that being one of the most frequently cited potential uses of such tools.

The negative bias of 6.2 percentage points deserves particular attention because it runs in a direction that might seem counterintuitive. One might expect unsupervised home testing to produce worse performance than testing in a sound-treated booth with a trained clinician, and indeed the online scores were lower on average. But the authors frame this as a structural consequence of comparing a real-world condition against a best-performance reference rather than as evidence of a flawed instrument. Clinical speech audiometry is administered under optimized conditions: calibrated equipment, controlled ambient noise, standardized presentation levels, and professional guidance. The online test inherits whatever acoustic environment, playback hardware, and volume settings the participant happens to have. That the average deficit is only about six points, despite these uncontrolled conditions, suggests the speech-based paradigm is comparatively robust to environmental variability, consistent with the broader literature showing that speech-in-noise and speech recognition tasks tolerate uncontrolled listening environments better than pure-tone detection does.

The reliability findings add an important layer to this picture. Intraclass correlation coefficients between 0.58 and 0.85 across modules indicate that repeated self-administration produces reasonably stable group-level rankings, and the upper end of that range approaches the reliability of some clinical instruments. Yet the smallest detectable change values, ranging from 17.6 to 35.8 percentage points, reveal the practical cost of this variability. The smallest detectable change represents the minimum difference between two administrations that can be attributed to real change rather than measurement noise with reasonable statistical confidence. For a patient monitoring their own hearing between clinic visits, a change smaller than roughly eighteen to thirty-six points on the relevant module cannot be distinguished from random variation. Clinical changes in word recognition that matter for decisions about device programming or medical management are often far smaller than this, which is precisely why the authors identify measurement error, rather than validity or engagement, as the central obstacle to individual-level use.

The ceiling effect observed in the quiet speech condition also carries implications for how such platforms should be designed. When 41 percent of completed attempts and nearly 58 percent of users score at or above 90 percent, the upper portion of the measurement scale compresses into a narrow range where differences between high performers become difficult to detect. This saturation is not unique to the Japanese materials; monosyllabic word recognition in quiet is known to plateau in listeners with normal or near-normal hearing across languages, and the phenomenon has been documented even among cochlear-implant recipients, which historically motivated the development of implant-specific speech batteries. The practical consequence is that tasks administered in quiet are poorly suited to distinguishing among people with good hearing or tracking change in those with mild losses. The speech-in-noise modules, by contrast, showed no such clustering: at the 0 dB signal-to-noise ratio, no attempt reached 90 percent, and accuracy declined in an orderly fashion as the listening condition became harder. This graded difficulty gives noise-based tasks a wider effective measurement range and helps explain why speech-in-noise paradigms have become the preferred foundation for many digital hearing assessment efforts worldwide.

The modest correlation between music perception and speech outcomes, with Spearman coefficients between 0.31 and 0.39, aligns with a growing body of evidence that musical and speech listening abilities, while sharing some underlying auditory resources, are substantially dissociable. The broad distribution of music scores, with only about 21 percent of users reaching the 90 percent threshold, suggests the Montreal Battery of Evaluation of Musical Abilities-based module avoids the ceiling compression that limits the quiet speech task. Whether music-based measures can add independent information about real-world listening, particularly in noisy social settings, remains an open question that the authors’ data raise but cannot resolve.

Finally, the study’s scale offers a rare window into how the public actually engages with online hearing assessment. Completion rates spanning 43.7 to 85.8 percent across modules, with shorter modules completed more often, indicate that test burden shapes real-world participation more than task difficulty does. The persistence of score distributions when analysis was restricted to one attempt per user strengthens confidence that the observed patterns reflect genuine properties of the tasks and the population rather than the behavior of a small cadre of repeat testers. For developers of digital health tools, these usage data provide a template for evaluating not just whether an instrument works, but whether people will actually use it long enough for it to matter.

Subject of Research: Real-world performance and clinical concordance of a Japanese browser-based hearing test platform

Article Title: Real-world performance of a Japanese browser-based hearing test: large-scale online usage data and concordance with clinical speech audiometry

Article References: Minami, S., Shimanuki, M., Nishiyama, T., Nishimura, H., Tateda, M., Kato, K., Yoshida, H., Sasaki, H., Shimizu, S., Wakisaka, N., Masuda, M., Marunaka, H., & Masuda, S. (2026). Real-world performance of a Japanese browser-based hearing test: large-scale online usage data and concordance with clinical speech audiometry. Frontiers in Audiology and Otology, 4, Article 1921721. https://doi.org/10.3389/fauot.2026.1921721

Image Credits: AI Generated

DOI: 10.3389/fauot.2026.1921721

Keywords: hearing test, digital health, speech audiometry, speech in noise, music perception, test-retest reliability, measurement error, hearing loss screening, teleaudiology, Japanese language, browser-based assessment, cochlear implant