Electronic Thesis and Dissertation Repository


Master of Engineering Science


Electrical and Computer Engineering


Parsa Vijay


Speech enhancement in assistive hearing devices has been an area of research for many decades. Noise reduction is particularly challenging because of the wide variety of noise sources and the non-stationarity of speech and noise. Digital signal processing (DSP) algorithms deployed in modern hearing aids for noise reduction rely on certain assumptions on the statistical properties of undesired signals. This could be disadvantageous in accurate estimation of different noise types, which subsequently leads to suboptimal noise reduction. In this research, a relatively unexplored technique based on deep learning, i.e. Recurrent Neural Network (RNN), is used to perform noise reduction and dereverberation for assisting hearing-impaired listeners. For noise reduction, the performance of the deep learning model was evaluated objectively and compared with that of open Master Hearing Aid (openMHA), a conventional signal processing based framework, and a Deep Neural Network (DNN) based model. It was found that the RNN model can suppress noise and improve speech understanding better than the conventional hearing aid noise reduction algorithm and the DNN model. The same RNN model was shown to reduce reverberation components with proper training. A real-time implementation of the deep learning model is also discussed.