Date Approved


Embargo Period


Document Type


Degree Name

M.S. in Engineering


Electrical & Computer Engineering


Henry M. Rowan College of Engineering


Ramachandran, Ravi


Electrical engineering--Research; Signal theory (Telecommunication)


Electrical and Computer Engineering


Signal-to-noise ratio is an important concept in electrical communications, as it is a measurable ratio between a given transmitted signal and the inherent background noise of a transmission channel. Currently signal-to-noise ratio testing is primarily performed by using an intrusive method of comparing a corrupted signal to the original signal and giving it a score based on the comparison. However, this technique is inefficient and often impossible for practical use because it requires the original signal for comparison. A speech signal's characteristics and properties could be used to develop a non-intrusive method for determining SNR, or a method that does not require the presence of the original clean signal.

In this thesis, several extracted features were investigated to determine whether a neural network trained with data from corrupt speech signals could accurately estimate the SNR of a speech signal. A MultiLayer Perceptron (MLP) was trained on extracted features for each decibel level from 0dB to 30dB, in an attempt to create 'expert classifiers' for each SNR level. This type of architecture would then have 31 independent classifiers operating together to accurately estimate the signal-to-noise ratio of an unknown speech signal. Principal component analysis was also implemented to reduce dimensionality and increase class discrimination. The performance of several neural network classifier structures is examined, as well as analyzing the overall results to determine the optimal feature for estimating signal-to-noise ratio of an unknown speech signal. Decision-level fusion was the final procedure which combined the outputs of several classifier systems in an effort to reduce the estimation error.