Authors: Andrew Sabin¹, Steve Taddei¹, Abram Bailey¹
¹HearAdvisor LLC
Background: Predicting real-world hearing-aid benefit for speech understanding remains a central challenge in audiology. Most existing metrics focus on objective speech intelligibility in controlled lab conditions, which may not fully capture whether real users experience improvement in everyday listening scenarios.
Method: We developed a large-scale perceptual database by deploying a blind listening test on HearAdvisor.com. Over 120,000 ratings were collected from hearing aid consumers who evaluated 80 commercial devices across 72 ambisonic scenes recorded on an acoustic manikin. Participants rated the “ease of speech understanding” on a five-point scale. This database was then used to train a deep learning model utilizing a pretrained automatic speech recognition architecture to predict mean listener ratings from audio.
Results and Conclusion: The predictive model demonstrated high generalization on held-out devices (r > 0.92), substantially outperforming widely used objective intelligibility metrics. These findings introduce a scalable approach to hearing-aid evaluation that is grounded in actual user perception, prioritizing how easily a device allows a listener to understand speech in realistic environments.
