"Accent Morphing"

Supervised by Professor RK Moore

Abstract

This paper builds an analysis of the major issues involved when trying to tackle the problem of accent morphing and the techniques that may be used to accomplish this task. In recent years, many academic researchers have shown great interest in the field of voice morphing and made considerable efforts in exploring more insights. However, it might be argued that the accent properties of speaker individuality are often ignored. This project initiates by looking into the various properties of speaker individuality and continues by identifying the features that are most influential on the variability of accents. Furthermore, this project introduces three methods for converting a source speaker's accent towards a given target speaker's accent. In addition to implementing the widely used mean-variance conversion method, two further conversion methods (Cubic and GMM-Based) were investigated. The latter sections of this report, introduce the results of the various tests that were conducted in order to evaluate the performance of the algorithms looked at. It can be seen from the results of these tests that the GMM-based and Cubic conversion functions manage to achieve better results than the baseline conversion function. However, as a result of these tests, it was seen that the degree of improvement made by using these methods is related to the characteristics and the diversity of the training and test utterances used.