HearAdvisor: Realistic Hearing Aid Analysis

• HearAdvisor is a standardized, ecologically grounded platform that evaluates hearing aids through recorded real-world acoustic scenes and listener-rated ease of speech understanding.
• It employs controlled testing with KEMAR manikin recordings and a MUSHRA-inspired blind listening paradigm to generate reliable, scene-level performance metrics.
• The system integrates traditional metrics like HASPIv2 with a novel, learned neural metric that outperforms standard methods in predicting perceived speech intelligibility.

Searching arXiv for papers on HearAdvisor and closely related hearing-aid evaluation systems. HearAdvisor is an independent test lab and public-facing platform for hearing aids that records commercial devices in realistic acoustic scenes, computes comparative performance metrics, and publishes both the metrics and the corresponding audio files. In recent work, the name also denotes a learned, data-driven metric for listener-rated ease of speech understanding, trained on a large perceptual dataset of commercial hearing-aid recordings. In related literature, the phrase “HearAdvisor-style system” is further used more broadly for standardized, scalable frameworks that evaluate, calibrate, or optimize hearables and hearing aids under realistic and individualized operating conditions (Sabin et al., 24 Jun 2026).

1. Scope, purpose, and conceptual basis

HearAdvisor is oriented toward hearing-aid consumers, clinicians, and researchers who require objective, comparable information about device behavior in realistic listening situations. Its central emphasis is speech-understanding benefit in quiet and noise, but the underlying rationale is not limited to objective intelligibility alone. The relevant perceptual construct is “ease of understanding,” which the literature describes as subjective and potentially incorporating intelligibility, listening effort, and overall sound quality. This distinction is important because HearAdvisor historically used HASPIv2, a metric designed to predict objective sentence intelligibility and validated primarily under simulated distortions, whereas its relationship to consumer-rated ease of understanding for commercial devices was uncertain (Sabin et al., 24 Jun 2026).

A common misconception is to treat HearAdvisor as merely a score leaderboard. The platform is more accurately described as an evaluation pipeline in which recordings, metrics, and human judgments are all tied to the same device corpus. For each device, the system records the hearing aid on a KEMAR acoustic manikin in many realistic environments, computes metrics and summary scores, and publishes both the metrics and the audio files. A plausible implication is that HearAdvisor occupies an intermediate position between laboratory benchmarking and consumer-facing comparison: its outputs are intended to remain technically grounded while still reflecting the kinds of listening conditions that determine practical device choice (Sabin et al., 24 Jun 2026).

2. Recording corpus and acoustic-scene design

The core HearAdvisor corpus described in recent work comprises binaural acoustic-manikin recordings of commercial hearing aids under a standardized scene design. The dataset contains 10,394 binaural acoustic-manikin recordings from 83 distinct commercial products across 72 realistic acoustic scenes, with each device tested in two fits: an initial fit and a tuned fit adjusted close to NAL-NL2 targets for the N3 audiogram. Scene-level aggregation yields 1,764 talker-pooled scene-level entries and 150 device-fit conditions across the 83 products (Sabin et al., 24 Jun 2026).

Component	Value
Products	83 distinct commercial products
Recordings	10,394 binaural acoustic-manikin recordings
Scenes	72 realistic acoustic scenes
Scene-level targets	1,764 talker-pooled scene-level entries

The acoustic scenes are based on the ARTE database of Ambisonic recordings of everyday environments. Twelve everyday background environments are decoded to eight loudspeakers arranged around the KEMAR manikin. Target talkers are rendered in six configurations: a single talker at $0^\circ$, two talkers at $\pm 45^\circ$, and three talkers at $0^\circ + \pm 45^\circ$. The talkers were recorded while listening to the corresponding background in order to induce Lombard-effect changes in level and articulation. Speech-to-noise ratios were chosen using real-world SNR distributions experienced by older adults with impaired hearing in daily life. The scenes are additionally divided into loud scenes, defined as $>70$ dB SPL, and quiet scenes, defined as $<70$ dB SPL; all recordings are diffuse-field equalized for headphone or speaker playback (Sabin et al., 24 Jun 2026).

This design makes the platform notable among hearing-aid benchmarks because it fixes the reproduction geometry and scene taxonomy across a broad commercial product set. The resulting corpus is not a simulation-only benchmark: it is a manikin-recorded corpus of actual devices under common protocols. That structure is what allows device-level and fit-level aggregation without changing the underlying acoustic reference frame (Sabin et al., 24 Jun 2026).

3. Perceptual data collection and score construction

The perceptual component is implemented through the HearAdvisor Blind Listening Challenge, embedded on device and comparison pages of the website. Participants are anonymous website visitors who self-report mild, moderate, or severe hearing loss; profound hearing loss is excluded. They are instructed to remove any personal hearing aids, listen over their best headphones or speakers, and perform a coarse calibration using a restaurant-ambience clip. Internally, $0$ dBFS is mapped to $100$ dB SPL, and listeners typically set the ambience clip to around $-30$ dBFS, approximately $70$ dB SPL, after which they are instructed not to change system volume (Sabin et al., 24 Jun 2026).

The listening paradigm is MUSHRA-inspired and blind. Each participant rates three scenes. On each scene screen, six sliders correspond to six unlabeled stimuli from the same scene. Four are real hearing-aid recordings, including the device associated with the visited page and three randomly chosen devices. Two are hidden anchors: a good anchor that meets NAL-NL2 gain targets exactly and applies a $+6$ dB SNR improvement through noise-only attenuation, and a bad anchor that is low-pass filtered at $\pm 45^\circ$0 kHz and applies no SNR improvement. Ratings use a five-point Ease of Understanding scale from $\pm 45^\circ$1 (“Really Hard”) to $\pm 45^\circ$2 (“Really Easy”) (Sabin et al., 24 Jun 2026).

The anchor design serves both screening and normalization. A session is discarded if the listener does not rate the good anchor at least one full point higher than the bad anchor on average. Ratings within each retained session are then linearly rescaled so that the participant’s mean good- and bad-anchor ratings match the global mean anchor ratings for those scenes. After filtering, 6,600 initial sessions become 4,878 after device and hearing-loss filters and 3,468 after the anchor gate. Ratings drop from 151,608 raw ratings to 104,298 after quality screening. Each individual recording is rated by a median of approximately six listeners, which motivates aggregation across the six talker configurations of a background. Agreement across talker configurations is high, with Cronbach’s $\pm 45^\circ$3, and scene-level targets are defined as rating-count-weighted means of all ratings assigned to a device $\pm 45^\circ$4 fit $\pm 45^\circ$5 background combination (Sabin et al., 24 Jun 2026).

The main epistemic consequence of this design is that HearAdvisor’s labels are neither raw clickstream preferences nor conventional laboratory MOS values. They are normalized, anchor-checked, scene-pooled ratings intended to stabilize a perceptual target that is sparse at the single-recording level but much more reliable at the talker-pooled scene level (Sabin et al., 24 Jun 2026).

4. Learned prediction of ease of speech understanding

The learned HearAdvisor metric is intrusive: it requires both the aided audio and a matched clean-speech reference. Both are downmixed to mono and passed through the same frozen Whisper-small encoder, a 244M-parameter speech-to-text model, using only its encoder component. At layer $\pm 45^\circ$6, the hidden-state sequence $\pm 45^\circ$7 is mean-pooled over time to produce a 768-dimensional representation. The metric then constructs a difference embedding

$\pm 45^\circ$8

which is intended to capture the effect of hearing-aid processing and noise while factoring out shared speech content (Sabin et al., 24 Jun 2026).

The model uses different encoder layers for different acoustic regimes. Quiet scenes, defined as $\pm 45^\circ$9 dB SPL, are routed through layer $0^\circ + \pm 45^\circ$0, while loud scenes, defined as $0^\circ + \pm 45^\circ$1 dB SPL, are routed through layer $0^\circ + \pm 45^\circ$2. Each route has its own MLP head with architecture $0^\circ + \pm 45^\circ$3, with layer normalization, GELU, and dropout on the hidden layers. Training uses weighted mean-squared error on talker-pooled scene-level mean ratings, with each target weighted by $0^\circ + \pm 45^\circ$4, where $0^\circ + \pm 45^\circ$5 is the number of ratings for that scene-level target. Optimization uses AdamW with learning rate $0^\circ + \pm 45^\circ$6, weight decay $0^\circ + \pm 45^\circ$7, and 200 epochs. To reduce initialization sensitivity, five independently trained heads are used per route and their predictions are averaged, for a total of ten heads (Sabin et al., 24 Jun 2026).

On held-out devices at the talker-pooled scene level, the learned metric substantially outperforms HASPIv2. Reported correlations with mean listener ratings are: overall $0^\circ + \pm 45^\circ$8 versus $0^\circ + \pm 45^\circ$9 for HASPIv2, loud $>70$0 versus $>70$1, and quiet $>70$2 versus $>70$3. The split-half reliability ceiling is $>70$4 for loud scenes and $>70$5 for quiet scenes, so the model effectively reaches the ceiling in loud scenes and approaches it in quiet scenes. At the device level, the learned metric reaches $>70$6 in loud scenes and $>70$7 in quiet scenes. Sensitivity analyses further show that predicted ease of understanding decreases systematically as gain undershoot above $>70$8 kHz increases and increases with SNR, with a much larger SNR effect in loud scenes than in quiet scenes (Sabin et al., 24 Jun 2026).

A second misconception is that this model predicts only intelligibility. The paper instead frames it as a predictor of listener-rated ease of understanding, which is broader than sentence recognition and can respond to gain shaping, noise reduction, and other processing effects that users perceive as benefit or burden even when word recognition is not the only latent variable (Sabin et al., 24 Jun 2026).

5. Validity, interpretation, and limitations

HearAdvisor’s current learned metric is highly effective within its intended operating regime, but that regime is specific. First, the metric is intrusive, so it is best suited to lab testing, standardized recording pipelines, manufacturer QA, and controlled benchmarking rather than arbitrary in-the-wild recordings where clean reference speech is unavailable. Second, all fittings and HASPIv2 calculations in the reported study use the standard N3 audiogram, so generalization to other audiograms is untested. Third, the perceptual data come from self-selected website visitors using uncontrolled playback hardware and only coarse calibration; anchor gating and within-session normalization mitigate these issues, but they do not eliminate them (Sabin et al., 24 Jun 2026).

Scene and language coverage are also bounded. The corpus spans 12 backgrounds and 6 talker configurations, which is substantial but still limited relative to everyday acoustic variability. The material is English-language, and multilingual extension is not addressed. Quiet-scene prediction remains below the split-half ceiling, indicating residual unexplained perceptual variance. This suggests that scene realism alone is not sufficient: the mapping from commercial hearing-aid output to human-perceived ease still depends on factors not fully captured by the present representation and corpus (Sabin et al., 24 Jun 2026).

These limitations clarify what HearAdvisor should and should not be taken to represent. It is not a universal substitute for clinical fitting, nor a complete model of all hearing-aid percepts. It is instead a standardized, ecologically grounded comparison framework whose most developed axis is speech-related benefit as perceived by listeners with self-reported hearing loss under a specific recording and evaluation protocol (Sabin et al., 24 Jun 2026).

6. Relation to the broader hearable-evaluation ecosystem

The term “HearAdvisor-style” is used in adjacent research to denote a larger ecosystem of evaluation, fitting, and optimization tools for hearables and hearing aids. One foundational requirement in that ecosystem is individualized estimation of eardrum sound pressure for hear-through or acoustic-transparency systems. An influential approach predicts the receiver-to-eardrum transfer function from an inward-facing in-ear microphone using a PCA-based model and a low-frequency LS estimator, achieving approximately $>70$9 dB error up to $<70$0 kHz and thereby reducing dependence on invasive probe-tube measurements (Jin et al., 2021). This suggests that a mature HearAdvisor-like platform would not only rate speech-understanding benefit but also model how devices alter eardrum-level acoustics under transparency or mixed-reality operation.

A second component is reproducible algorithmic infrastructure. The open Master Hearing Aid, or openMHA, provides an open-source, AGPL-3.0, C++ hearing-aid research platform with calibration, filterbank, multiband compression, beamforming, noise reduction, and feedback-control plugins, along with cross-platform real-time deployment and reference configurations from published studies (Kayser et al., 2021). This makes openMHA a plausible back-end substrate for benchmarkable HearAdvisor-style pipelines in which device algorithms, fittings, and evaluation scenes must be replayed identically across laboratories and hardware targets.

A third component is acoustic-context understanding for device steering. The DEAR benchmark contains 1,158 audio tracks, each 30 seconds long, and evaluates foundation-model representations on eight hearable-relevant tasks spanning scene context, speech sources, and technical acoustic properties such as SNR, RT60, and DRR; among Wav2Vec2.0, HuBERT, WavLM, and BEATs, the BEATs model significantly surpasses the others (Gröger et al., 10 Feb 2025). A plausible implication is that future HearAdvisor systems may fuse recording-based perceptual metrics with learned scene representations to explain not only which device performs better, but also under which acoustic regimes and why.

Finally, the intrusive nature of the current HearAdvisor metric has already motivated adjacent work on non-intrusive prediction. Enhancer-guided intelligibility prediction for hearing-impaired listeners uses parallel enhanced-signal pathways, speech foundation models, and audiogram conditioning to outperform the non-intrusive CPC2 Champion baseline across CPC3 and Arehart datasets (Cao et al., 21 Sep 2025). If such approaches mature, HearAdvisor could evolve from a primarily laboratory-style platform toward a hybrid system that combines controlled recording benchmarks with reference-free field assessment.

In that broader sense, HearAdvisor denotes not just a website or a single metric, but an emerging methodological program: standardized device recording, perceptually grounded outcome modeling, individualized acoustic characterization, reproducible algorithmic evaluation, and progressively richer context awareness for hearing technologies.