Neural networks for noise canceling applications

number: 
402
English
Degree: 
Author: 
Rasha R. A. Izzat
Supervisor: 
Dr. Thamer M. Jameil
year: 
2000
Abstract:

One of the important problems for many years is the reduction of noise in a specified environment. The problem of enhancing speech degrarded by additive noise may improve the overall quality, to increase the intelligibility and to reduce the listener fatigue. The approach to speech enhancement depends considerably on the environment in which the adaptive noise canceling (ANC) is needed. One approach that regards to reduce the additive noise that arises from different sources is the two microphones approach which employs both primary input (noise - corrupted speech) and reference input (noise alone).The adaptive noise canceller was achieved using many conventional techniques such as the adaptive filter .In the past few decades neural networks begun to find wide applicability in many real world applications because of the key features of neural networks.The main goal of this research is to implement a suitable neural network system with the following specific requirements to perform the task of noise cancellation .These requirements have been summarized as follows:
1-Choosing suitable learning rules to train that system.
2-Obtaining the simulation results with appropriate software programming.
3-Comparing these results to show the performance of the rules.
These rules were the L.M.S. algorithm in the adaptive system and Back propagation and Levenberg - Marquardt algorithms in the neural network system. These algorithms were tested with different types of noise with different cases in the present work . Results obtained have confirmed that (a) The neural network system has achieved the aim by canceling noise and (b) showing a better performance in performing the noise cancellation, with ten nodes in the hidden layer and white Gaussian noise at the reference microphone, the RMS in the back-propagation algorithm reaches convergence after 388 iterations and in the Levenberg-marquardt algorithm reaches convergence after 8 iterations. While with real speech noise at the reference microphone, the RMS in the back-propagation algorithm reaches convergence after 833 iterations and in the Levenberg-marquardt algorithm reaches convergence after 6 iterations.