Second International Conference on Spoken Language Processing (ICSLP'92)
Banff, Alberta, Canada
HADIFIX is a speech synthesis system for German. It transforms a phonemic input string with accent markers into a speech signal. To obtain high-quality speech, time-domain elements with a structure similar to demisyllables are concatenated. A method to generate vowel-to-vowel transitions for time-domain units is used when necessary. Prosodic manipulations are carried out using the TD-PSOLA algorithm . An F0 contour is generated using Fujisaki's model [2,3]. The quality of the synthesized speech was compared with that of natural speech and of an LPC-based version using a rhyme test, a word test, a dictation, and a subjective ranking test. While the intelligibility is high, the difference in naturalness between synthesized and natural speech is still large. Another outcome of the test is the superiority of the time-domain units, compared with the LP-coded ones.
Bibliographic reference. Portele, Thomas / Steffan, Birgit / Preuß, Rainer / Sendlmeier, Walter F. / Hess, Wolfgang (1992): "HADIFIX - a speech synthesis system for German", In ICSLP-1992, 1227-1230.