Sixth European Conference on Speech Communication and Technology

Budapest, Hungary
September 5-9, 1999

Efficient Weight Training for Selection Based Synthesis

Yoram Meron, Keikichi Hirose

Department of Information and Communication Engineering, Faculty of Engineering, The University of Tokyo, Tokyo, Japan

Synthesis systems, based on unit selection from databases with a large number of unit examples from different phonetic and prosodic contexts, have been shown to allow a high quality speech synthesis [1]. One of the dificult problems associated with this kind of synthesis is determining the strategy for unit selection, and tuning its parameters. In this work we suggest a way to extend the usefulness of two existing training methods, by using phoneme pairs as the basic comparison unit. Using unit pairs is shown to significantly increase the efifciency of the exhaustive weight search training method on one hand, and refining the regression weight training method on the other.

Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Meron, Yoram / Hirose, Keikichi (1999): "Efficient weight training for selection based synthesis", In EUROSPEECH'99, 2319-2322.