Study On Objective Estimation Of Speech Quality Using Neural Networks

In traditional objective estimation of speech quality based on auditory model, it is necessary to introduce some mathematics models which is adapted to ear characteristics so as to describe the perception of auditory system, but it is a trouble in practice.In this thesis, the problem will be solved by using spectral distortion measures, that is, using simple characteristic parameters indicate the spectral character of speech signal, such as Mel Frequency Cepstral Coefficient (MFCC) and Bark Spectral Distance (BSD) measure, to replace those complex mathematic models, and defining and computing those complex processing function reflecting auditory characteristics in vector measure. In addition, compared with high non-linear auditory system, artificial neural network has the capability to learn from limit sample sets and map input to output in various dimensions. And under the condition that there are sufficient and representative samples, the network will come close to real auditory characteristics satisfactorily. Based on this way, this thesis aimed at the objective estimation of speech quality, multilayer perceptron and radial basis function network, instead of Euclidean distance corresponding to the parameters of MFCC and BSD, are trained by character parameters and the subjective assessment results named expected values. As a result, it really reflects the characteristics of ear effectively, and reduces the calculating time in spectral distortion measure as well.Radial basis function neural network has the competence of learning rapidly because of basis function. No matter what the non-linear function that the network wants to approximate is, it is not easy to influence the performance of the network. The key point is the choice of the center of basis function. The performance of network is not satisfied if the center is unsuitable or improper. So how to choose the center of basis function is outlined in detail soas to help network play in practice successfully.Moreover, the samples data in network training should be typical, well distributed but a few redundant data is permitted. Because of the high-dimension, high-redundant and full of noisy training samples, it is necessary to present the pre-procession of sample data.At last, compared with other training results, the results using what the thesis introduced is presented. And based on the comparisons and analysis, how the research work will go on is also indicated.