On the Detection of Discontinuities in Concatenative Speech Synthesis

Abstract

Last decade considerable work has been done in finding an objective distance measure which is able to predict audible discontinuities in concatenative speech synthesis. Speech segments in concatenative synthesis are extracted from disjoint phonetic contexts and discontinuities in spectral shape and phase mismatches tend to occur at unit boundaries. Many feature sets--most of them of spectral nature--and distances were tested. However there were significant discrepancies among the results. In this paper, we tested most of the distances that were proposed using the same listening experiment. Best score were given by AM&FM decomposition of the speech signal using Fisher's linear discriminant.