The ESCA Workshop on Speech Synthesis

September 25-28, 1990
Autrans, France

F0 Generation with a Data Base of Natural F0 Patterns and with a Neural Network

Christof Traber

Group For Speech and Language Processing, Institute of Electronics, Swiss Federal Institute of Technology (ETH), Zurich, Switzerland

We present two approaches to generate F0 contours for German utterances based on a phonological transcription of the utterances (phonetic string, accents, and phrase boundaries). The first approach uses a data base of natural F0 patterns which are concatenated to new Fo contours. The second one uses a recurrent neural network to produce global F0 contours directly from the encoded phonological transcription. Our results show that both approaches are well-suited to produce high-quality F0 contours. So far, the resulting contours produced with the neural network are better than the ones produces with the patterns data base. Using a neural network for the generation of complete F0 contours with high quality is feasible and may require much less human effort than other approaches.


Full Paper

Bibliographic reference.  Traber, Christof (1990): "F0 generation with a data base of natural F0 patterns and with a neural network", In SSW1-1990, 141-144.