Oldenburg Hearing Health Repository (OHHR)
- OHHR is an open-access hearing-health dataset that integrates calibrated absolute-level measurements and supra-threshold parameters to enable detailed auditory profiling.
- It supports mobile hearing test calibration by employing Bayesian regression and nearest-neighbor models to estimate and correct device calibration offsets from ACALOS measurements.
- It serves as a benchmarking substrate for comparing eight auditory profiling frameworks using standardized preprocessing, intrinsic metrics, and manifold learning techniques.
The Oldenburg Hearing Health Repository (OHHR), also described as the Oldenburg Hearing Health Record, is an open-access hearing-health dataset used for the analysis of auditory profiles and for modeling supra-threshold hearing measures under controlled calibration conditions. In the studies considered here, OHHR serves two distinct but related roles: as a calibrated source of adaptive categorical loudness scaling (ACALOS) measurements for estimating device calibration offsets in mobile hearing tests, and as an extended multimodal dataset for systematic comparison of auditory profiling frameworks under identical preprocessing and evaluation conditions (Xu et al., 20 Aug 2025, Xu et al., 7 Jan 2026).
1. Repository definition and research role
In the auditory profiling study, the extended OHHR is the common, open-access dataset used to apply and compare eight auditory profiling frameworks under identical conditions, enabling direct and fair comparisons across approaches. The analyses used an extended OHHR comprising 1,127 participants with mean age $67.2$ years (); were male and were female. The data supporting that study are openly available on Zenodo at 10.5281/zenodo.14177903, and the underlying dataset is described by Jafri et al. (2025) as “OHHR – The Oldenburg Hearing Health Record Dataset” at 10.5281/zenodo.16919812 (Xu et al., 7 Jan 2026).
In the calibration-offset study, OHHR is used as a large, fully calibrated, and richly annotated dataset containing CLS outcomes at two frequencies and spanning diverse hearing profiles. That study used a subset comprising participants with mean age $70.0$ years (), including 556 male and 291 female listeners. All participants had dB HL. Because calibration offsets can be simulated precisely on top of calibrated measurements, OHHR is presented there as an appropriate basis for training and validating algorithms that infer calibration offsets from level-independent CLS features (Xu et al., 20 Aug 2025).
Taken together, these uses position OHHR as a repository for both population-level auditory characterization and method development. A plausible implication is that the repository’s value lies not only in scale, but also in the coexistence of calibrated absolute-level measurements and supra-threshold descriptors that can support downstream statistical modeling.
2. Data composition and measurement modalities
The extended OHHR used for auditory profiling contains eight tests yielding 37 parameters. These tests comprise a questionnaire, two cognitive assessments—verbal intelligence and DemTect—an SF-12 health survey, adaptive categorical loudness scaling, two speech-in-noise tests, namely the Göttingen Sentence Test (GÖSA) and the Digit Triplet Test (DTT), and pure-tone audiometry (Xu et al., 7 Jan 2026).
The pure-tone audiometry component consists of air-conduction thresholds at 11 frequencies from $0.25$ to $8$ kHz for both ears. Instrumentation was a Unity II audiometer with HDA200 headphones, measured according to IEC 60645-1 (2002) standards in sound-treated booths. Explicit inclusion and exclusion criteria are not stated in the auditory profiling paper, although its limitations section notes that OHHR primarily includes older patients with hearing impairments, which may constrain generalizability (Xu et al., 7 Jan 2026).
The ACALOS subset used in the mobile-calibration work is more narrowly specified. CLS was performed with calibrated Sennheiser HDA200 headphones using narrowband noises of 2 s at 1500 Hz and 4000 Hz in an ISO 16832-compliant procedure. The sample spanned several Bisgaard hearing-profile classes: 0, 1, 2, 3, 4, and 5. The study states that this spread covers a wide range of audiometric configurations from near-normal to sloping and flat sensorineural loss (Xu et al., 20 Aug 2025).
A central measurement construct in the calibration study is categorical loudness scaling. ACALOS uses an 11-point scale with 7 labeled categories and 4 intermediate steps, mapping perceived loudness to categorical units (CU), typically from 0 (“not heard”) up to 50 CU (“too loud”). Per ear and frequency, a loudness growth function is fitted, yielding 6, 7, 8, 9 (0), 1 (2), and 3 (4) (Xu et al., 20 Aug 2025).
3. Level-independent CLS descriptors and calibration modeling
The calibration-offset study uses OHHR to exploit the distinction between level-dependent and level-independent CLS descriptors. The loudness growth slope in dB per CU is written as 5, with 6 denoting categorical loudness units. Under an additive calibration offset 7, slopes are invariant because
8
Similarly, the dynamic range is defined as
9
and remains invariant under additive offset because
0
The study therefore treats 1, 2, and 3 as calibration-invariant descriptors (Xu et al., 20 Aug 2025).
The six predictors entering both models are the level-independent features at two frequencies: 4, 5, 6, 7, 8, and 9. Calibration offsets were simulated from Gaussian distributions,
0
with 1 dB and 2 dB, to mimic uncalibrated smartphones. The additive offset was applied only to level-dependent CLS parameters—3, 4, 5, and 6—while the invariant features remained unchanged by design. A common 7 across frequencies was assumed for each listener, reflecting device-level calibration bias (Xu et al., 20 Aug 2025).
The Bayesian regression model does not predict 8 directly. Instead, it predicts 9 from invariant features in calibrated OHHR data. For participant $70.0$0 at frequency $70.0$1,
$70.0$2
The paper reports that the model was fit with brms/Stan to the calibrated OHHR dataset ($70.0$3). On uncalibrated measurements, the observed $70.0$4 is shifted by $70.0$5, so the offset estimate is formed from the difference between observed and model-predicted $70.0$6, averaged over frequencies: $70.0$7 An additional self-consistency check using the same framework strongly predicted $70.0$8 with $70.0$9 at 1500 Hz and 0 at 4000 Hz, with median absolute error approximately 1–2 dB depending on Bisgaard profile and frequency (Xu et al., 20 Aug 2025).
The nearest-neighbor comparator used the same six invariant features. Distance was defined by minimizing RMSE between feature vectors, equivalently Euclidean distance on standardized features,
3
with 4 and uniform weighting. The offset estimate was then derived from 5 differences relative to the nearest calibrated OHHR neighbor (Xu et al., 20 Aug 2025).
4. OHHR as a basis for auditory profiling comparisons
In the profiling study, OHHR is used to compare eight profiling frameworks under a common data regime. These include six audiogram-only approaches—Baseline (PTA4), WHO HI grades, WARHICS levels, Bisgaard profiles, audiometric phenotypes, and general phenotypes—and two comprehensive approaches that incorporate supra-threshold information, namely BEAR profiles and Hearing4All profiles (Xu et al., 7 Jan 2026).
The framework definitions differ in both data requirements and assignment behavior. Fixed numbers of classes are reported as follows: Baseline 6, WHO 7, WARHICS 8, Bisgaard 9, audiometric phenotypes 0, general phenotypes 1, BEAR 2, and Hearing4All 3. When applied to OHHR, some frameworks left participants unassigned: 704 for audiometric phenotype, 545 for general phenotype, and 87 for BEAR. Hearing4All assigned all participants (Xu et al., 7 Jan 2026).
The study reports that both clustering method and number of profiles substantially influenced the resulting auditory profiles. A direct comparison of vector quantization (VQ) and Gaussian mixture modeling (GMM) at 4 showed that VQ had significantly lower Davies–Bouldin scores than GMM (5). When GMM was evaluated across 6 to 7, the Davies–Bouldin index was lowest at 8, while excluding that trivial two-group case, the lowest value occurred at 9; the worst value occurred at $0.25$0, and differences across $0.25$1 were significant by ANOVA ($0.25$2) (Xu et al., 7 Jan 2026).
Across intrinsic measures normalized by $0.25$3, Bisgaard profiles achieved the lowest Davies–Bouldin index among the compared frameworks, while audiometric phenotypes performed worst by that criterion. WHO HI grades had the highest Calinski–Harabasz index. Bisgaard was closest to $0.25$4 on the Silhouette Index, whereas audiometric phenotypes were farthest. Among comprehensive frameworks, BEAR outperformed Hearing4All on two of three measures—Calinski–Harabasz and Silhouette—but left 87 participants unidentified, whereas Hearing4All achieved the lowest Davies–Bouldin index among comprehensive profiles and assigned all participants (Xu et al., 7 Jan 2026).
These findings make OHHR a benchmarking substrate for comparing profile systems that are otherwise difficult to evaluate side by side. The repository’s combination of audiometric and supra-threshold variables is especially consequential here because some frameworks depend only on thresholds, whereas others integrate loudness scaling and speech-in-noise performance.
5. Preprocessing, manifold learning, and intrinsic evaluation
Before principal component analysis, OHHR data were standardized to unit variance. PCA was conducted using FactoMineR, with five retained dimensions and visualization focused on PC1 and PC2. PC1 explained $0.25$5 of the variance and PC2 explained $0.25$6. The paper reports that PC1 was predominantly associated with audiometric thresholds, whereas PC2 was driven by supra-threshold loudness parameters (Xu et al., 7 Jan 2026).
For non-linear manifold learning, the study used t-SNE via Rtsne with perplexity $0.25$7, $0.25$8, no initial PCA step, and other hyperparameters at defaults. The role of PCA was interpretability and linear visualization; the role of t-SNE was preservation of local neighborhood relations and clearer two-dimensional separation. In the reported results, t-SNE yielded slightly clearer separations than PCA across frameworks; in both views, normal-hearing groups appeared at the extreme right and greater hearing loss at the left (Xu et al., 7 Jan 2026).
The intrinsic measures used for framework comparison were the Davies–Bouldin index, the Calinski–Harabasz index, and the Silhouette Index, each normalized by $0.25$9 to account for differing numbers of profile classes. Bootstrapping was performed with 1,000 resamples, each containing $8$0 points, using scikit-learn implementations (Xu et al., 7 Jan 2026).
The study provides the standard definitions. For $8$1 clusters, the Davies–Bouldin index is
$8$2
where lower values are better. For a point $8$3, the Silhouette coefficient is
$8$4
with range $8$5 to $8$6, where higher values are better. The Calinski–Harabasz index is
$8$7
where higher values are better. A Dunn index formula is also given for completeness in the study details, but it was not used in the paper (Xu et al., 7 Jan 2026).
A methodological point emerges from the OHHR-based analysis: intrinsic metrics and manifold visualizations address different aspects of profile validity. The paper explicitly notes that intrinsic measures provide benchmarking but do not ensure clinical usefulness, and that 2D PCA or t-SNE projections may obscure separability present in higher dimensions (Xu et al., 7 Jan 2026).
6. Applications, limitations, and future directions
In the smartphone-hearing context, OHHR supports a workflow in which ACALOS is run on the smartphone, invariant features are extracted at multiple frequencies, calibrated $8$8 values are predicted from OHHR-trained models, and an estimated device calibration offset $8$9 is used to correct subsequent presentation levels. The study states that apps can use this mechanism to correct stimulus levels on the fly, enabling trustworthy supra-threshold assessments and potentially improving threshold-related estimates without physical calibration tools. It further suggests integration into OHHR workflows through periodic retraining of Bayesian regressors on the growing CLS subset, expansion to more frequencies, inclusion of device metadata such as make, model, operating system, and audio-path specifics, and incorporation of real-world smartphone CLS recordings linked to OHHR profiles for external validation and domain adaptation (Xu et al., 20 Aug 2025).
Quantitatively, the Bayesian regression model achieved correlations of up to 00 between estimated and true calibration offsets when 01 dB, and approximately 02 when 03 dB. The nearest-neighbor model achieved approximately 04–05 for 06 dB and approximately 07 for 08 dB. All reported correlations were significant at 09. Uncertainty reduction was defined as
10
with regression-model values 11, 12, 13, and 14 for 15, and nearest-neighbor values 16, 17, 18, and 19 for the same conditions. For stratified analysis at mean 20 dB and SD 21 dB, overall MAE was 22 dB for regression and 23 dB for nearest neighbor; single-frequency estimates at 4000 Hz were worse, with overall MAE 24 dB (Xu et al., 20 Aug 2025).
In the auditory profiling context, OHHR supports objective comparison of auditory-profile generation strategies through consistent preprocessing, normalization of intrinsic measures, and common visualization procedures. The study concludes that Bisgaard profiles are strongest among audiogram-only approaches, while Hearing4All is promising for comprehensive profiling because it combines a near-optimal number of classes (25) with high clustering quality and full assignment coverage (Xu et al., 7 Jan 2026).
Both lines of work also delimit the repository’s current boundaries. For the calibration study, offsets were simulated as Gaussian and frequency-constant, whereas real devices may exhibit frequency-dependent and context-dependent biases; the older-adult cohort with hearing loss limits tested generalizability; and external validation on truly uncalibrated smartphone measurements remains necessary (Xu et al., 20 Aug 2025). For the profiling study, OHHR primarily represents older patients with hearing impairments; correspondence among profiling frameworks remains incomplete; sample-size effects, additional auditory features, and alternative algorithms such as decision trees and spectral clustering remain to be explored; and intrinsic clustering measures do not by themselves establish clinical utility (Xu et al., 7 Jan 2026).
A plausible implication of these studies is that OHHR is most consequential when used as an extensible reference infrastructure rather than as a closed benchmark. The calibration work points toward OHHR-linked mobile data collection and hierarchical modeling of device heterogeneity, while the profiling work points toward framework comparison on common data with explicit attention to coverage, separability, and interpretability.
