Second ESCA/IEEE Workshop on Speech Synthesis
September 12-15, 1994
In this paper, automatic derivation of F0 control parameters using Multiple Split Regression (MSR) is proposed. In this derivation, a superpositional F0-model was employed to reduce the number of parameters. Control parameters were determined automatically using MSR, and were optimized using a hill climbing algorithm to minimize the error between observed and predicted F0 contours. Major findings in a test using 200 read sentences of Japanese were as follows : (1) the dominant factor controlling amplitude of the accent command was accent type, (2) for the phrase command it was the number of morae in the previous phrase.
Bibliographic reference. Hirai, Toshio / Iwahashi, Naoto / Higuchi, Norio / Sagisaka, Yoshinori (1994): "Automatic extraction of FO control parameters using statistical analysis", In SSW2-1994, 57-60.