PURPOSE: To improve speech recognizing process performance and speech synthesizing process performance by correcting a misextracted local peak position.

CONSTITUTION: A fundamental frequency correction part 106 applies a linear regression model to a fundamental frequency pattern sent out of a fundamental frequency calculation part 105 to model a fundamental frequency in a proper analytic section. A fundamental frequency sample value in the analytic section is evaluated by using the modeled fundamental frequency pattern to evaluate whether or not local peak extraction is correct. Then the evaluation and extraction position correcting process is performed for the whole input speech and information on the local peak position after the correction and the fundamental frequency pattern are sent out to a fundamental frequency pattern output part 107. Further, the fundamental frequency pattern output part 107 displays the fundamental frequency pattern in synchronism with a speech waveform and also outputs information on a local peak position series to an output terminal 108.