PURPOSE: To increase a recognition rate and shorten processing time.
CONSTITUTION: A pitch extracting means 13 extracts the frequency of the pitch in the whole section of an input speech and the envelope of its frequency variation is obtained through a low-pass filter 14 of about 1-0.5Hz. The peak of enveloping (pitch peak) is obtaind by a detection means 21. The peak position of the enveloping of the pitch frequency variation statistically found as to the object speech to be recognized and respective time lengths form the position to the head and tail of the speech section to be recognizing object fare recorded in storage means 16-18 respectively, and speech section candidates for the recognition object are obtained from control parameters of the input speech from an LPC calculation part 3 on the basis of the previously found pitch peak according to the stored information and word registration and likelihood calculation are performed only for the section candidates to decides the candidate with high likelihood as a recognition result.
HONMA SHIGERU
KITAI MIKIO
ARAI KAZUHIRO
Next Patent: SPEAKER VERIFICATION METHOD USING GROUP NORMALIZATION SCORING