To generate a natural fundamental frequency pattern close to human vocalization by providing a phrase model changing means and a phrase component generating means for generating a phrase component by using the phrase model changed by the phrase model changing means.

A phrase component generating part 3 generates the phrase component of a fundamental frequency pattern. A phrase time length arithmetic part 4 calculates the time lengths Tp sec of the respective phrases from the starting time of each phrase obtained by adding the time lengths of the respective syllables. A time constant changing part 5 composing a phrase model changing means changes the value of a time constant (a) in a phrase component generating part 3 and decides the time constant (a) based on the following rule. The rule is that if the time length Tp of a phrase is less than one second, the time constant is decided as a=0.5 [1/sec]. When the time length Tp of the phrase is longer than one second, the time constant a=5.0/Tp [1/sec].