Personalization of Noise Reduction and Compression for Hearing Enhancement

Noise reduction and dynamic range compression constitute two main signal processing modules for hearing enhancement in modern digital hearing aid devices. This dissertation covers personalization solutions for both of these two modules. First, in contrast to the great majority of previous works that are designed based on pre-collected datasets, in this dissertation, a personalized noise reduction solution is developed for field deployment which is capable of dealing with unseen noisy audio environments. More specifically, a deep learning-based approach is devised to improve speech denoising in real-world audio environments by not requiring the availability of clean speech signals as reference. A fully convolutional neural network is designed based on two noisy realizations of the same speech signal, one used as the input and the other as the target of the network. The results of extensive experimentations are reported to show the superiority of the developed personalized deep learning-based speech denoising approach over existing approaches. In support of personalized noise reduction, a personalized noise classification is also developed in this dissertation by performing the noise classification in an unsupervised manner. Second, in contrast to the existing prescriptive compression strategies used in hearing aids which are devised based on gain averages from a group of users, a personalized compression solution is developed in this dissertation via a human-in-the-loop deep reinforcement learning approach. The developed approach is designed to learn a specific user’s hearing preferences in order to set compression ratios based on the user’s preference feedbacks. Both simulation and subject testing results are reported to show the superiority of the developed personalized deep learning-based compression over conventional prescriptive compression. In addition, four smartphone apps in support of the above two modules are developed in this dissertation which include the real-time implementation of noise classification, noise reduction, and dynamic range compression.