摘要 |
PURPOSE:To obtain a voice in which an intonation is held by deriving a pitch frequency for satisfying a prescribed condition by a specific expression when the most probable greatest common resonance frequency is set as a pitch frequency from a filter gain. CONSTITUTION:A theoretical filter gain in fHz of a j-th filter in the expression I is denoted as Xj(f), and an observed filter gain is denoted as Gj. A voice passes through a filter bank 11 and a voice power detector 12, and by using an output of a low frequency side filter of the bank 11 and the voice power, an A(f) computer 13 executes a calculation of A(f) by the expression I, and stores it temporarily in an A(f) calculation result memory 14. Also, a recognition template generator 17 generates a recognition template by using an output of the bank 11 and stores it in a dictionary 20. Moreover, a voiced/silent discriminator 18 stores voiced/silent information in the dictionary. An Min SIGMAB(fk) computing element 16 calculates an (fk) train for satisfying Min SIGMAB(fk) by using the memory 14 and a value of a transition probability tape 15, stores it as a pitch frequency in the dictionary 20, and at this time point, the dictionary registration is completed. |