Symbolic segment dissimilarity measure and its applications in speech synthesis

This paper presents a novel method for determining a measure of dissimilarity of symbolic representations of speech segments. It is supposed that these symbolic representations can be completely derived from text, e.g. using a text analysis component of a text-to-speech synthesiser. An analysis is made to evaluate the correspondence of the particular symbolic features with the acoustic variability. A final dissimilarity measure is designed so as to maximise the correspondence with the selected acoustic measure. The designed measure is applied to the problem of sentence selection for the speech synthesis voice databases.