Committee: 

Dr. Anderson, Advisor 

Dr. Moore, Chair

Dr. Clements

Abstract:

The objective of the proposed research is to develop an efficient neural network-based speech enhancement processor.  Neural networks are powerful models that have in the last few years been applied to several speech processing problems including speech enhancement.  There, however, remain several questions such as the most suitable architectures, training features, and best practices for obtaining optimal results.  We investigate the use of the extreme learning machine, an algorithm that allows feed-forward networks to be quickly trained and provides good generalization, for speech enhancement.  We propose modifications to the extreme learning machine to increase its accuracy.  The proposed research also investigates the use of a time domain noise subtraction architecture for enhancing low-SNR signals; the use of artifact-free block processing methods; and the incorporation of perceptual criteria in the design of neural network-based speech enhancement systems.