Modeling Contextual Preferences of Hearing Aid Users

Maciej Jan Korzepa

Research output: Book/ReportPh.D. thesisResearch

7 Downloads (Pure)

Abstract

The current clinical approach for calibrating hearing aid (HA) settings to compensate for individual hearing loss employs amplification schemes which are established based on how an average hearing impaired listener perceives loudness or speech. This approach does not take into consideration the large variability across users nor the variety in auditory needs resulting from differences in listening environments and situations encountered throughout the day. While modern HAs provide a wide range of capabilities that can address these variations, the lack of resources in hearing health care prevents a successful personalization of the devices in clinical settings. A possible solution is to move the personalization process outside the clinics by employing algorithms to learn auditory preferences of HA users in real environments. This project explores how to model and learn the preferences of HA users dependent on the context. Firstly, an in-the-wild user study was run to explore setting preferences of HA users in real listening environments. The results showed that HA users may have preferences for HA settings very different from the ones provided in standard clinical settings and that these preferences vary not only between but also within individual users depending on the context. Secondly, a Gaussian process based approach is proposed to simulate contextual preferences of HA users in order to facilitate a systematic evaluation and agile development of HA personalization solutions. It incorporates aspects such as correlations of preferences across users, complexity of preferences, partial observability of context and user features inducing these preferences, and noisiness of perceptual user feedback. Simulation allows to circumvent the limitations of traditional user studies and enables a more scalable approach to run multiple
large-scale experiments validating different aspects of preference learning solutions, such as efficiency of learning in the presence of limited and noisy data,
leveraging similarities across users or accounting for latent factors. Thirdly, a novel probabilistic meta-learning method is proposed inspired by parallel research focusing on Bayesian neural networks and their connections with Gaussian processes. This meta-learning method is applied to the problem of contextual personalization of HA settings in a number of simulations with varying assumptions about the observability of user and context features. The results indicate potential both for using the probabilistic meta-learning framework to facilitate contextual personalization in the HA domain and for evaluating and comparing HA personalization solutions in simulated environments. This project has provided a foundation for taking a more holistic approach to
HA personalization by enabling development of scalable solutions that learn auditory preferences dependent on the context.
Original languageEnglish
PublisherTechnical University of Denmark
Number of pages157
Publication statusPublished - 2020

Fingerprint

Dive into the research topics of 'Modeling Contextual Preferences of Hearing Aid Users'. Together they form a unique fingerprint.

Cite this