Speech Perception Benefit

Overview

We quantify the expected improvement in speech understanding for each device/fit combination via acoustic measurement. The published score is the average of two components computed on the same recordings:

The two components are deliberately different in spirit. HASPIv2 is a model of the impaired auditory periphery developed by hearing scientists and validated against objective intelligibility measures — it represents the professional research perspective. The second component is a machine-learning model trained on more than 100,000 ratings collected through our Blind Listening Challenge, in which website visitors with hearing loss rate recordings of unlabeled devices in terms of ease of understanding — it represents the perspective of real hearing aid consumers. Averaging the two yields a score informed equally by intelligibility research and by user preference.

Computation

Intelligibility Component (ΔHASPIv2)

We compute HASPIv2 using the N3 audiogram (Table 1) and RAU-transform[15] the output. We average across both ears and compute the difference between unaided and aided recordings:

Positive values indicate the hearing aid improved speech intelligibility; negative values indicate degradation.

Listener-Rated Component

The second component is produced by a model trained to predict listener-rated ease of speech understanding directly from audio. The aided recording and a matched clean-speech reference are each passed through a frozen speech foundation model (the Whisper encoder), and the difference between their internal representations is mapped to a predicted listener rating by a small trained network.

The training data come from our Blind Listening Challenge: over 100,000 quality-screened ratings in which website visitors with hearing loss rated recordings of unlabeled devices on a five-point ease-of-understanding scale. The dataset — 83 commercial products recorded across 72 realistic acoustic scenes — and the model are described in Sabin et al. (2026)[27].

The figure below shows the model's predictions against the mean listener rating for devices that were entirely held out of training. The model tracks consumer ratings closely in both loud and quiet scenes (r = 0.92 overall), substantially better than HASPIv2 on the same data (r = 0.83)[27]:

Each point is one device fit in one background scene. Adapted from Sabin et al. (2026)[27], Figure 2.

Quiet/Moderate vs. Loud Environments

We report both components separately for two environment categories:

For each category, we average each component across all qualifying acoustic scenes.

Mapping to the 5-Point Scale

The intelligibility component is mapped to our 5-point scale via linear scaling, as in previous versions of this methodology: a $\Delta$HASPIv2 of 0 (no benefit relative to the open ear) maps to 0 on our scale, with increasing benefit scaled linearly up to 5.

The listener-rated component is linearly rescaled so that its distribution of scores across all tested devices matches the distribution of the scaled intelligibility component. This ensures the two components contribute comparably to the final score and that published scores remain comparable to previous versions of this methodology.

The published Speech Perception Benefit score is the unweighted mean of the two components:

Both components are bounded below at 1.0 before averaging, so the published score does not distinguish among devices at the extreme low end of the scale. There is no upper bound: a device with strong benefit can score above 5.