Online Personalization of Compression in Hearing Aids via Maximum Likelihood Inverse Reinforcement Learning

Abstract

A key function of modern hearing aids is compression or mapping of sound to the residual hearing range of those suffering from hearing loss. This paper presents a machine learning approach to personalize compression in hearing aids in an online manner. The online feature of this approach allows it to be deployed in the field. The significance of this personalized compression lies in enabling preferred hearing outcomes relative to the one-size-fits-all prescriptive compression rationales that are currently being used. This personalization approach utilizes maximum likelihood inverse reinforcement learning to establish a model of a hearing aid user's preference based on paired comparisons by the user. The results of the preference paired comparisons between the personalized and standard prescriptive settings from ten subjects indicated that personalized settings were preferred about 10 times more than the standard prescriptive settings. In addition, a word recognition comparison was conducted showing that the personalized settings had no adverse impact on speech understanding in either quiet or in competing noise conditions.