Authors: Lars Bramsløw¹, Szymon Drgas², Peter Leer³
¹Demant A/S; ²Poznan University of Technology; ³Eriksholm Research Centre
Background: Recent applications of auditory models with hearing loss use ‘closed loop’ approaches: Deep learning is used to optimize the parameters of a deep neural network (DNN) by comparing a normal hearing and a hearing-impaired branch of the auditory model, thereby obtaining a new, unrestricted hearing-loss compensation (HLC). This approach builds upon the development of rationales like NAL-NL1 and NL2, but targets modern neural-based compensation strategies.
Method: Based on recent publications, we conducted a meta-analysis comparing available auditory models, including both physiological auditory nerve models and psychoacoustic loudness models. We evaluated both quantitative and qualitative outcomes to discuss the strengths and weaknesses of these candidate models in the context of HLC.
Results: Physiological models offer superior detail but are computationally demanding and difficult to personalize via audiogram, especially as supra-threshold measures are currently missing. Psychoacoustic models are simpler and easier to personalize but lack important properties like temporal fine structure. To date, no model has demonstrated clear superiority for hearing-loss compensation, showing only slight improvements over conventional NAL-RP linear amplification.
Conclusion: The application of advanced computational auditory models for hearing-loss compensation is in its early stages, with no single preferred model. Existing models still lack critical properties of individual hearing losses, highlighting the need for further research to enhance personalization and functional outcomes.
