VOICE RECOGNITION

PURPOSE: To improve the recognition rate by making a fundamental frequency logarithmic and correcting the time gradient, and normalizing the fundamental frequency and extracting a paragraph border, etc. CONSTITUTION: The fundamental frequency of a speech is found from an input speech signal and made logarithmic, and a soundless part and a voiceless part in the speech are interpolated to find time-series information on the fundamental frequency made logarithmic. Then the recursive straight line of this time-series information is found and subtracted from the time-series information to normalize the gradient. Then the maximum and minimum values are found from the time-series information of the fundamental frequency having the gradient normalized to extract the paragraph border and important words. Consequently, the speech recognition and the recognition rate of a speech understanding part are improved.