A systematic approach to the extraction of diphone elements from natural speech

Synthetic speech can be generated with an unrestricted vocabulary by concatenating stored units such as diphone elements. When joining speech segments that were not adjacent in the original context they were taken from, discontinuities in the spectral envelope may arise that impair intelligibility. The method proposed here attempts to find optimum diphone boundaries in order to minimize these discontinuities, Steady-state zones of all phones carrying a diphone boundary are specified by means of a centroid vector. Based on the centroids and on an objective distance measure, hypothetical boundary cost functions are defined. Their minimization together with the evaluation of a set of additional rules determines the boundary locations. A rhyme test carried out with speech generated by concatenating diphone elements extracted according to this method yielded an intelligibility score of 96.7 percent for isolated words.