Date of Award

December 2021

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Computer Engineering

Committee Member

Melissa Smith

Committee Member

Feng Luo

Committee Member

YingJie Lao

Abstract

Speech enhancement is a critical part in automatic speech recognition systems. Recently with the development of deep learning based techniques, those speech enhancement systems trained with neural networks can significantly improve performance. While many of the latest speech enhancement systems show advantages in maximizing the perceptual quality of the noisy signals, they expose drawbacks when the test noisy signals have noise types that never exist during the system training process. The systems have relatively poor performance when handling noisy signals with unseen noise in contrast to noisy signals with seen noise. The dissimilarity between the training and testing circumstances can cause a serious performance decline in a deep learning task.In this work, a new method is proposed to solve the noise types problem. The framework has three parts: the autoencoder, the gradient reverse layers and the recurrent neural networks. The proposed framework can weaken the noise types influences when handling random noisy signals. This work shows that the new method outperforms the baseline models in unseen noise situations.

Share

COinS
 
 

To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.